2004-08-20 20:54:07 +02:00
|
|
|
/*
|
2011-02-12 01:29:35 +01:00
|
|
|
* Copyright (c) 2011 ARM Limited
|
|
|
|
* All rights reserved.
|
|
|
|
*
|
|
|
|
* The license below extends only to copyright in the software and shall
|
|
|
|
* not be construed as granting a license to any other intellectual
|
|
|
|
* property including but not limited to intellectual property relating
|
|
|
|
* to a hardware implementation of the functionality of the software
|
|
|
|
* licensed hereunder. You may use the software subject to the license
|
|
|
|
* terms below provided that you ensure that this notice is replicated
|
|
|
|
* unmodified and in its entirety in all distributions of the software,
|
|
|
|
* modified or unmodified, in source code or in binary form.
|
|
|
|
*
|
2006-06-13 01:04:42 +02:00
|
|
|
* Copyright (c) 2004-2006 The Regents of The University of Michigan
|
2010-02-12 20:53:19 +01:00
|
|
|
* Copyright (c) 2009 The University of Edinburgh
|
2004-08-20 20:54:07 +02:00
|
|
|
* All rights reserved.
|
|
|
|
*
|
|
|
|
* Redistribution and use in source and binary forms, with or without
|
|
|
|
* modification, are permitted provided that the following conditions are
|
|
|
|
* met: redistributions of source code must retain the above copyright
|
|
|
|
* notice, this list of conditions and the following disclaimer;
|
|
|
|
* redistributions in binary form must reproduce the above copyright
|
|
|
|
* notice, this list of conditions and the following disclaimer in the
|
|
|
|
* documentation and/or other materials provided with the distribution;
|
|
|
|
* neither the name of the copyright holders nor the names of its
|
|
|
|
* contributors may be used to endorse or promote products derived from
|
|
|
|
* this software without specific prior written permission.
|
|
|
|
*
|
|
|
|
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
|
|
|
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
|
|
|
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
|
|
|
|
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
|
|
|
|
* OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
|
|
|
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
|
|
|
|
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
|
|
|
|
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
|
|
|
|
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
|
|
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
|
|
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
2006-06-01 01:26:56 +02:00
|
|
|
*
|
|
|
|
* Authors: Kevin Lim
|
2010-02-12 20:53:19 +01:00
|
|
|
* Timothy M. Jones
|
2004-08-20 20:54:07 +02:00
|
|
|
*/
|
|
|
|
|
2005-02-26 00:00:49 +01:00
|
|
|
#ifndef __CPU_BASE_DYN_INST_HH__
|
|
|
|
#define __CPU_BASE_DYN_INST_HH__
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-06-14 19:12:41 +02:00
|
|
|
#include <bitset>
|
2006-04-23 00:26:48 +02:00
|
|
|
#include <list>
|
2004-08-20 20:54:07 +02:00
|
|
|
#include <string>
|
|
|
|
|
2006-06-03 00:15:20 +02:00
|
|
|
#include "arch/faults.hh"
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
#include "arch/utility.hh"
|
2004-08-20 20:54:07 +02:00
|
|
|
#include "base/fast_alloc.hh"
|
|
|
|
#include "base/trace.hh"
|
2005-08-30 19:18:54 +02:00
|
|
|
#include "config/full_system.hh"
|
2009-09-23 17:34:21 +02:00
|
|
|
#include "config/the_isa.hh"
|
2006-12-06 11:48:59 +01:00
|
|
|
#include "cpu/o3/comm.hh"
|
2005-02-26 00:00:49 +01:00
|
|
|
#include "cpu/exetrace.hh"
|
|
|
|
#include "cpu/inst_seq.hh"
|
2006-06-03 00:15:20 +02:00
|
|
|
#include "cpu/op_class.hh"
|
2005-02-26 00:00:49 +01:00
|
|
|
#include "cpu/static_inst.hh"
|
2010-02-12 20:53:19 +01:00
|
|
|
#include "cpu/translation.hh"
|
2006-06-03 00:15:20 +02:00
|
|
|
#include "mem/packet.hh"
|
2010-09-14 04:26:03 +02:00
|
|
|
#include "sim/byteswap.hh"
|
2006-04-23 00:26:48 +02:00
|
|
|
#include "sim/system.hh"
|
2009-04-09 07:21:27 +02:00
|
|
|
#include "sim/tlb.hh"
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/**
|
|
|
|
* @file
|
|
|
|
* Defines a dynamic instruction context.
|
|
|
|
*/
|
|
|
|
|
Update to make multiple instruction issue and different latencies work.
Also change to ref counted DynInst.
SConscript:
Add branch predictor, BTB, load store queue, and storesets.
arch/isa_parser.py:
Specify the template parameter for AlphaDynInst
base/traceflags.py:
Add load store queue, store set, and mem dependence unit to the
list of trace flags.
cpu/base_dyn_inst.cc:
Change formating, add in debug statement.
cpu/base_dyn_inst.hh:
Change DynInst to be RefCounted, add flag to clear whether or not this
instruction can commit. This is likely to be removed in the future.
cpu/beta_cpu/alpha_dyn_inst.cc:
AlphaDynInst has been changed to be templated, so now this CC file
is just used to force instantiations of AlphaDynInst.
cpu/beta_cpu/alpha_dyn_inst.hh:
Changed AlphaDynInst to be templated on Impl. Removed some unnecessary
functions.
cpu/beta_cpu/alpha_full_cpu.cc:
AlphaFullCPU has been changed to be templated, so this CC file is now
just used to force instantation of AlphaFullCPU.
cpu/beta_cpu/alpha_full_cpu.hh:
Change AlphaFullCPU to be templated on Impl.
cpu/beta_cpu/alpha_impl.hh:
Update it to reflect AlphaDynInst and AlphaFullCPU being templated
on Impl. Also removed time buffers from here, as they are really
a part of the CPU and are thus in the CPU policy now.
cpu/beta_cpu/alpha_params.hh:
Make AlphaSimpleParams inherit from the BaseFullCPU so that it doesn't
need to specifically declare any parameters that are already in the
BaseFullCPU.
cpu/beta_cpu/comm.hh:
Changed the structure of the time buffer communication structs. Now
they include the size of the packet of instructions it is sending.
Added some parameters to the backwards communication struct, mainly
for squashing.
cpu/beta_cpu/commit.hh:
Update typenames to reflect change in location of time buffer structs.
Update DynInst to DynInstPtr (it is refcounted now).
cpu/beta_cpu/commit_impl.hh:
Formatting changes mainly. Also sends back proper information
on branch mispredicts so that the bpred unit can update itself.
Updated behavior for non-speculative instructions (stores, any
other non-spec instructions): once they reach the head of the ROB,
the ROB signals back to the IQ that it can go ahead and issue the
non-speculative instruction. The instruction itself is updated so that
commit won't try to commit it again until it is done executing.
cpu/beta_cpu/cpu_policy.hh:
Added branch prediction unit, mem dependence prediction unit, load
store queue. Moved time buffer structs from AlphaSimpleImpl to here.
cpu/beta_cpu/decode.hh:
Changed typedefs to reflect change in location of time buffer structs
and also the change from DynInst to ref counted DynInstPtr.
cpu/beta_cpu/decode_impl.hh:
Continues to buffer instructions even while unblocking now. Changed
how it loops through groups of instructions so it can properly block
during the middle of a group of instructions.
cpu/beta_cpu/fetch.hh:
Changed typedefs to reflect change in location of time buffer structs
and the change to ref counted DynInsts. Also added in branch
brediction unit.
cpu/beta_cpu/fetch_impl.hh:
Add in branch prediction. Changed how fetch checks inputs and its
current state to make for easier logic.
cpu/beta_cpu/free_list.cc:
Changed int regs and float regs to logically use one flat namespace.
Future change will be moving them to a single scoreboard to conserve
space.
cpu/beta_cpu/free_list.hh:
Mostly debugging statements. Might be removed for performance in future.
cpu/beta_cpu/full_cpu.cc:
Added in some debugging statements. Updated BaseFullCPU to take
a params object.
cpu/beta_cpu/full_cpu.hh:
Added params class within BaseCPU that other param classes will be
able to inherit from. Updated typedefs to reflect change in location
of time buffer structs and ref counted DynInst.
cpu/beta_cpu/iew.hh:
Updated typedefs to reflect change in location of time buffer structs
and use of ref counted DynInsts.
cpu/beta_cpu/iew_impl.hh:
Added in load store queue, updated iew to be able to execute non-
speculative instructions, instead of having them execute in commit.
cpu/beta_cpu/inst_queue.hh:
Updated change to ref counted DynInsts. Changed inst queue to hold
non-speculative instructions as well, which are issued only when
commit signals backwards that a nonspeculative instruction is at
the head of the ROB.
cpu/beta_cpu/inst_queue_impl.hh:
Updated to allow for non-speculative instructions to be in the inst
queue. Also added some debug functions.
cpu/beta_cpu/regfile.hh:
Added debugging statements, changed formatting.
cpu/beta_cpu/rename.hh:
Updated typedefs, added some functions to clean up code.
cpu/beta_cpu/rename_impl.hh:
Moved some code into functions to make it easier to read.
cpu/beta_cpu/rename_map.cc:
Changed int and float reg behavior to use a single flat namespace. In
the future, the rename maps can be combined to a single rename map to
save space.
cpu/beta_cpu/rename_map.hh:
Added destructor.
cpu/beta_cpu/rob.hh:
Updated it with change from DynInst to ref counted DynInst.
cpu/beta_cpu/rob_impl.hh:
Formatting, updated to use ref counted DynInst.
cpu/static_inst.hh:
Updated forward declaration for AlphaDynInst now that it is templated.
--HG--
extra : convert_revision : 1045f240ee9b6a4bd368e1806aca029ebbdc6dd3
2004-09-23 20:06:03 +02:00
|
|
|
// Forward declaration.
|
|
|
|
class StaticInstPtr;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
template <class Impl>
|
Update to make multiple instruction issue and different latencies work.
Also change to ref counted DynInst.
SConscript:
Add branch predictor, BTB, load store queue, and storesets.
arch/isa_parser.py:
Specify the template parameter for AlphaDynInst
base/traceflags.py:
Add load store queue, store set, and mem dependence unit to the
list of trace flags.
cpu/base_dyn_inst.cc:
Change formating, add in debug statement.
cpu/base_dyn_inst.hh:
Change DynInst to be RefCounted, add flag to clear whether or not this
instruction can commit. This is likely to be removed in the future.
cpu/beta_cpu/alpha_dyn_inst.cc:
AlphaDynInst has been changed to be templated, so now this CC file
is just used to force instantiations of AlphaDynInst.
cpu/beta_cpu/alpha_dyn_inst.hh:
Changed AlphaDynInst to be templated on Impl. Removed some unnecessary
functions.
cpu/beta_cpu/alpha_full_cpu.cc:
AlphaFullCPU has been changed to be templated, so this CC file is now
just used to force instantation of AlphaFullCPU.
cpu/beta_cpu/alpha_full_cpu.hh:
Change AlphaFullCPU to be templated on Impl.
cpu/beta_cpu/alpha_impl.hh:
Update it to reflect AlphaDynInst and AlphaFullCPU being templated
on Impl. Also removed time buffers from here, as they are really
a part of the CPU and are thus in the CPU policy now.
cpu/beta_cpu/alpha_params.hh:
Make AlphaSimpleParams inherit from the BaseFullCPU so that it doesn't
need to specifically declare any parameters that are already in the
BaseFullCPU.
cpu/beta_cpu/comm.hh:
Changed the structure of the time buffer communication structs. Now
they include the size of the packet of instructions it is sending.
Added some parameters to the backwards communication struct, mainly
for squashing.
cpu/beta_cpu/commit.hh:
Update typenames to reflect change in location of time buffer structs.
Update DynInst to DynInstPtr (it is refcounted now).
cpu/beta_cpu/commit_impl.hh:
Formatting changes mainly. Also sends back proper information
on branch mispredicts so that the bpred unit can update itself.
Updated behavior for non-speculative instructions (stores, any
other non-spec instructions): once they reach the head of the ROB,
the ROB signals back to the IQ that it can go ahead and issue the
non-speculative instruction. The instruction itself is updated so that
commit won't try to commit it again until it is done executing.
cpu/beta_cpu/cpu_policy.hh:
Added branch prediction unit, mem dependence prediction unit, load
store queue. Moved time buffer structs from AlphaSimpleImpl to here.
cpu/beta_cpu/decode.hh:
Changed typedefs to reflect change in location of time buffer structs
and also the change from DynInst to ref counted DynInstPtr.
cpu/beta_cpu/decode_impl.hh:
Continues to buffer instructions even while unblocking now. Changed
how it loops through groups of instructions so it can properly block
during the middle of a group of instructions.
cpu/beta_cpu/fetch.hh:
Changed typedefs to reflect change in location of time buffer structs
and the change to ref counted DynInsts. Also added in branch
brediction unit.
cpu/beta_cpu/fetch_impl.hh:
Add in branch prediction. Changed how fetch checks inputs and its
current state to make for easier logic.
cpu/beta_cpu/free_list.cc:
Changed int regs and float regs to logically use one flat namespace.
Future change will be moving them to a single scoreboard to conserve
space.
cpu/beta_cpu/free_list.hh:
Mostly debugging statements. Might be removed for performance in future.
cpu/beta_cpu/full_cpu.cc:
Added in some debugging statements. Updated BaseFullCPU to take
a params object.
cpu/beta_cpu/full_cpu.hh:
Added params class within BaseCPU that other param classes will be
able to inherit from. Updated typedefs to reflect change in location
of time buffer structs and ref counted DynInst.
cpu/beta_cpu/iew.hh:
Updated typedefs to reflect change in location of time buffer structs
and use of ref counted DynInsts.
cpu/beta_cpu/iew_impl.hh:
Added in load store queue, updated iew to be able to execute non-
speculative instructions, instead of having them execute in commit.
cpu/beta_cpu/inst_queue.hh:
Updated change to ref counted DynInsts. Changed inst queue to hold
non-speculative instructions as well, which are issued only when
commit signals backwards that a nonspeculative instruction is at
the head of the ROB.
cpu/beta_cpu/inst_queue_impl.hh:
Updated to allow for non-speculative instructions to be in the inst
queue. Also added some debug functions.
cpu/beta_cpu/regfile.hh:
Added debugging statements, changed formatting.
cpu/beta_cpu/rename.hh:
Updated typedefs, added some functions to clean up code.
cpu/beta_cpu/rename_impl.hh:
Moved some code into functions to make it easier to read.
cpu/beta_cpu/rename_map.cc:
Changed int and float reg behavior to use a single flat namespace. In
the future, the rename maps can be combined to a single rename map to
save space.
cpu/beta_cpu/rename_map.hh:
Added destructor.
cpu/beta_cpu/rob.hh:
Updated it with change from DynInst to ref counted DynInst.
cpu/beta_cpu/rob_impl.hh:
Formatting, updated to use ref counted DynInst.
cpu/static_inst.hh:
Updated forward declaration for AlphaDynInst now that it is templated.
--HG--
extra : convert_revision : 1045f240ee9b6a4bd368e1806aca029ebbdc6dd3
2004-09-23 20:06:03 +02:00
|
|
|
class BaseDynInst : public FastAlloc, public RefCounted
|
2004-08-20 20:54:07 +02:00
|
|
|
{
|
|
|
|
public:
|
|
|
|
// Typedef for the CPU.
|
2006-06-16 23:08:47 +02:00
|
|
|
typedef typename Impl::CPUType ImplCPU;
|
|
|
|
typedef typename ImplCPU::ImplState ImplState;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
// Logical register index type.
|
Changes to untemplate StaticInst and StaticInstPtr, change the isa to a namespace instead of a class, an improvement to the architecture specific header file selection system, and fixed up a few include paths.
arch/alpha/alpha_linux_process.cc:
Added using directive for AlphaISA namespace
arch/alpha/alpha_memory.hh:
arch/alpha/isa/branch.isa:
cpu/pc_event.hh:
Added typedefs for Addr
arch/alpha/alpha_tru64_process.cc:
arch/alpha/arguments.cc:
Added using directive for AlphaISA
arch/alpha/ev5.hh:
Added an include of arch/alpha/isa_traits.hh, and a using directive for the AlphaISA namespace.
arch/alpha/faults.hh:
Added a typedef for the Addr type, and changed the formatting of the faults slightly.
arch/alpha/isa/main.isa:
Untemplatized StaticInst, added a using for namespace AlphaISA to show up in decoder.cc and the exec.ccs, relocated makeNop to decoder.hh
arch/alpha/isa/mem.isa:
Untemplatized StaticInst and StaticInstPtr
arch/alpha/isa/pal.isa:
cpu/base_dyn_inst.cc:
Untemplatized StaticInstPtr
arch/alpha/isa_traits.hh:
Changed variables to be externs instead of static since they are part of a namespace and not a class.
arch/alpha/stacktrace.cc:
Untemplatized StaticInstPtr, and added a using directive for AlphaISA.
arch/alpha/stacktrace.hh:
Added some typedefs for Addr and MachInst, and untemplatized StaticInstPtr
arch/alpha/vtophys.cc:
Added a using directive for AlphaISA
arch/alpha/vtophys.hh:
Added the AlphaISA namespace specifier where needed
arch/isa_parser.py:
Changed the placement of the definition of the decodeInst function to be outside the namespaceInst namespace.
base/loader/object_file.hh:
cpu/o3/bpred_unit.hh:
Added a typedef for Addr
base/loader/symtab.hh:
Added a typedef for Addr, and added a TheISA to Addr in another typedef
base/remote_gdb.cc:
Added a using namespace TheISA, and untemplatized StaticInstPtr
base/remote_gdb.hh:
Added typedefs for Addr and MachInst
cpu/base.cc:
Added TheISA specifier to some variables exported from the isa.
cpu/base.hh:
Added a typedef for Addr, and TheISA to some variables from the ISA
cpu/base_dyn_inst.hh:
Untemplatized StaticInstPtr, and added TheISA specifier to some variables from the ISA.
cpu/exec_context.hh:
Added some typedefs for types from the isa, and added TheISA specifier to some variables from the isa
cpu/exetrace.hh:
Added typedefs for some types from the ISA, and untemplatized StaticInstPtr
cpu/memtest/memtest.cc:
cpu/o3/btb.cc:
dev/baddev.cc:
dev/ide_ctrl.cc:
dev/ide_disk.cc:
dev/isa_fake.cc:
dev/ns_gige.cc:
dev/pciconfigall.cc:
dev/platform.cc:
dev/sinic.cc:
dev/uart8250.cc:
kern/freebsd/freebsd_system.cc:
kern/linux/linux_system.cc:
kern/system_events.cc:
kern/tru64/dump_mbuf.cc:
kern/tru64/tru64_events.cc:
sim/process.cc:
sim/pseudo_inst.cc:
sim/system.cc:
Added using namespace TheISA
cpu/memtest/memtest.hh:
cpu/trace/opt_cpu.hh:
cpu/trace/reader/itx_reader.hh:
dev/ide_disk.hh:
dev/pcidev.hh:
dev/platform.hh:
dev/tsunami.hh:
sim/system.hh:
sim/vptr.hh:
Added typedef for Addr
cpu/o3/2bit_local_pred.hh:
Changed the include to use arch/isa_traits.hh instead of arch/alpha/isa_traits.hh. Added typedef for Addr
cpu/o3/alpha_cpu.hh:
Added typedefs for Addr and IntReg
cpu/o3/alpha_cpu_impl.hh:
Added this-> to setNextPC to fix a problem since it didn't depend on template parameters any more. Removed "typename" where it was no longer needed.
cpu/o3/alpha_dyn_inst.hh:
Cleaned up some typedefs, and untemplatized StaticInst
cpu/o3/alpha_dyn_inst_impl.hh:
untemplatized StaticInstPtr
cpu/o3/alpha_impl.hh:
Fixed up a typedef of MachInst
cpu/o3/bpred_unit_impl.hh:
Added a using TheISA::MachInst to a function
cpu/o3/btb.hh:
Changed an include from arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef for Addr
cpu/o3/commit.hh:
Removed a typedef of Impl::ISA as ISA, since TheISA takes care of this now.
cpu/o3/cpu.cc:
Cleaned up namespace issues
cpu/o3/cpu.hh:
Cleaned up namespace usage
cpu/o3/decode.hh:
Removed typedef of ISA, and changed it to TheISA
cpu/o3/fetch.hh:
Fized up typedefs, and changed ISA to TheISA
cpu/o3/free_list.hh:
Changed include of arch/alpha/isa_traits.hh to arch/isa_traits.hh
cpu/o3/iew.hh:
Removed typedef of ISA
cpu/o3/iew_impl.hh:
Added TheISA namespace specifier to MachInst
cpu/o3/ras.hh:
Changed include from arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef for Addr.
cpu/o3/regfile.hh:
Changed ISA to TheISA, and added some typedefs for Addr, IntReg, FloatReg, and MiscRegFile
cpu/o3/rename.hh:
Changed ISA to TheISA, and added a typedef for RegIndex
cpu/o3/rename_map.hh:
Added an include for arch/isa_traits.hh, and a typedef for RegIndex
cpu/o3/rob.hh:
Added a typedef for RegIndex
cpu/o3/store_set.hh:
cpu/o3/tournament_pred.hh:
Changed an include of arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef of Addr
cpu/ozone/cpu.hh:
Changed ISA into TheISA, and untemplatized StaticInst
cpu/pc_event.cc:
Added namespace specifier TheISA to Addr types
cpu/profile.hh:
kern/kernel_stats.hh:
Added typedef for Addr, and untemplatized StaticInstPtr
cpu/simple/cpu.cc:
Changed using directive from LittleEndianGuest to AlphaISA, which will contain both namespaces. Added TheISA where needed, and untemplatized StaticInst
cpu/simple/cpu.hh:
Added a typedef for MachInst, and untemplatized StaticInst
cpu/static_inst.cc:
Untemplatized StaticInst
cpu/static_inst.hh:
Untemplatized StaticInst by using the TheISA namespace
dev/alpha_console.cc:
Added using namespace AlphaISA
dev/simple_disk.hh:
Added typedef for Addr and fixed up some formatting
dev/sinicreg.hh:
Added TheISA namespace specifier where needed
dev/tsunami.cc:
dev/tsunami_io.cc:
dev/tsunami_pchip.cc:
Added using namespace TheISA. It might be better for it to be AlphaISA
dev/tsunami_cchip.cc:
Added typedef for TheISA. It might be better for it to be AlphaISA
kern/linux/aligned.hh:
sim/pseudo_inst.hh:
Added TheISA namespace specifier to Addr
kern/linux/linux_threadinfo.hh:
Added typedef for Addr, and TheISA namespace specifier to StackPointerReg
kern/tru64/mbuf.hh:
Added TheISA to Addr type in structs
sim/process.hh:
Added typedefs of Addr, RegFile, and MachInst
sim/syscall_emul.cc:
Added using namespace TheISA, and a cast of VMPageSize to the int type
sim/syscall_emul.hh:
Added typecast for Addr, and TheISA namespace specifier for where needed
--HG--
extra : convert_revision : 91d4f6ca33a73b21c1f1771d74bfdea3b80eff45
2006-02-19 08:34:37 +01:00
|
|
|
typedef TheISA::RegIndex RegIndex;
|
2006-06-08 22:58:50 +02:00
|
|
|
// Integer register type.
|
Changes to untemplate StaticInst and StaticInstPtr, change the isa to a namespace instead of a class, an improvement to the architecture specific header file selection system, and fixed up a few include paths.
arch/alpha/alpha_linux_process.cc:
Added using directive for AlphaISA namespace
arch/alpha/alpha_memory.hh:
arch/alpha/isa/branch.isa:
cpu/pc_event.hh:
Added typedefs for Addr
arch/alpha/alpha_tru64_process.cc:
arch/alpha/arguments.cc:
Added using directive for AlphaISA
arch/alpha/ev5.hh:
Added an include of arch/alpha/isa_traits.hh, and a using directive for the AlphaISA namespace.
arch/alpha/faults.hh:
Added a typedef for the Addr type, and changed the formatting of the faults slightly.
arch/alpha/isa/main.isa:
Untemplatized StaticInst, added a using for namespace AlphaISA to show up in decoder.cc and the exec.ccs, relocated makeNop to decoder.hh
arch/alpha/isa/mem.isa:
Untemplatized StaticInst and StaticInstPtr
arch/alpha/isa/pal.isa:
cpu/base_dyn_inst.cc:
Untemplatized StaticInstPtr
arch/alpha/isa_traits.hh:
Changed variables to be externs instead of static since they are part of a namespace and not a class.
arch/alpha/stacktrace.cc:
Untemplatized StaticInstPtr, and added a using directive for AlphaISA.
arch/alpha/stacktrace.hh:
Added some typedefs for Addr and MachInst, and untemplatized StaticInstPtr
arch/alpha/vtophys.cc:
Added a using directive for AlphaISA
arch/alpha/vtophys.hh:
Added the AlphaISA namespace specifier where needed
arch/isa_parser.py:
Changed the placement of the definition of the decodeInst function to be outside the namespaceInst namespace.
base/loader/object_file.hh:
cpu/o3/bpred_unit.hh:
Added a typedef for Addr
base/loader/symtab.hh:
Added a typedef for Addr, and added a TheISA to Addr in another typedef
base/remote_gdb.cc:
Added a using namespace TheISA, and untemplatized StaticInstPtr
base/remote_gdb.hh:
Added typedefs for Addr and MachInst
cpu/base.cc:
Added TheISA specifier to some variables exported from the isa.
cpu/base.hh:
Added a typedef for Addr, and TheISA to some variables from the ISA
cpu/base_dyn_inst.hh:
Untemplatized StaticInstPtr, and added TheISA specifier to some variables from the ISA.
cpu/exec_context.hh:
Added some typedefs for types from the isa, and added TheISA specifier to some variables from the isa
cpu/exetrace.hh:
Added typedefs for some types from the ISA, and untemplatized StaticInstPtr
cpu/memtest/memtest.cc:
cpu/o3/btb.cc:
dev/baddev.cc:
dev/ide_ctrl.cc:
dev/ide_disk.cc:
dev/isa_fake.cc:
dev/ns_gige.cc:
dev/pciconfigall.cc:
dev/platform.cc:
dev/sinic.cc:
dev/uart8250.cc:
kern/freebsd/freebsd_system.cc:
kern/linux/linux_system.cc:
kern/system_events.cc:
kern/tru64/dump_mbuf.cc:
kern/tru64/tru64_events.cc:
sim/process.cc:
sim/pseudo_inst.cc:
sim/system.cc:
Added using namespace TheISA
cpu/memtest/memtest.hh:
cpu/trace/opt_cpu.hh:
cpu/trace/reader/itx_reader.hh:
dev/ide_disk.hh:
dev/pcidev.hh:
dev/platform.hh:
dev/tsunami.hh:
sim/system.hh:
sim/vptr.hh:
Added typedef for Addr
cpu/o3/2bit_local_pred.hh:
Changed the include to use arch/isa_traits.hh instead of arch/alpha/isa_traits.hh. Added typedef for Addr
cpu/o3/alpha_cpu.hh:
Added typedefs for Addr and IntReg
cpu/o3/alpha_cpu_impl.hh:
Added this-> to setNextPC to fix a problem since it didn't depend on template parameters any more. Removed "typename" where it was no longer needed.
cpu/o3/alpha_dyn_inst.hh:
Cleaned up some typedefs, and untemplatized StaticInst
cpu/o3/alpha_dyn_inst_impl.hh:
untemplatized StaticInstPtr
cpu/o3/alpha_impl.hh:
Fixed up a typedef of MachInst
cpu/o3/bpred_unit_impl.hh:
Added a using TheISA::MachInst to a function
cpu/o3/btb.hh:
Changed an include from arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef for Addr
cpu/o3/commit.hh:
Removed a typedef of Impl::ISA as ISA, since TheISA takes care of this now.
cpu/o3/cpu.cc:
Cleaned up namespace issues
cpu/o3/cpu.hh:
Cleaned up namespace usage
cpu/o3/decode.hh:
Removed typedef of ISA, and changed it to TheISA
cpu/o3/fetch.hh:
Fized up typedefs, and changed ISA to TheISA
cpu/o3/free_list.hh:
Changed include of arch/alpha/isa_traits.hh to arch/isa_traits.hh
cpu/o3/iew.hh:
Removed typedef of ISA
cpu/o3/iew_impl.hh:
Added TheISA namespace specifier to MachInst
cpu/o3/ras.hh:
Changed include from arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef for Addr.
cpu/o3/regfile.hh:
Changed ISA to TheISA, and added some typedefs for Addr, IntReg, FloatReg, and MiscRegFile
cpu/o3/rename.hh:
Changed ISA to TheISA, and added a typedef for RegIndex
cpu/o3/rename_map.hh:
Added an include for arch/isa_traits.hh, and a typedef for RegIndex
cpu/o3/rob.hh:
Added a typedef for RegIndex
cpu/o3/store_set.hh:
cpu/o3/tournament_pred.hh:
Changed an include of arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef of Addr
cpu/ozone/cpu.hh:
Changed ISA into TheISA, and untemplatized StaticInst
cpu/pc_event.cc:
Added namespace specifier TheISA to Addr types
cpu/profile.hh:
kern/kernel_stats.hh:
Added typedef for Addr, and untemplatized StaticInstPtr
cpu/simple/cpu.cc:
Changed using directive from LittleEndianGuest to AlphaISA, which will contain both namespaces. Added TheISA where needed, and untemplatized StaticInst
cpu/simple/cpu.hh:
Added a typedef for MachInst, and untemplatized StaticInst
cpu/static_inst.cc:
Untemplatized StaticInst
cpu/static_inst.hh:
Untemplatized StaticInst by using the TheISA namespace
dev/alpha_console.cc:
Added using namespace AlphaISA
dev/simple_disk.hh:
Added typedef for Addr and fixed up some formatting
dev/sinicreg.hh:
Added TheISA namespace specifier where needed
dev/tsunami.cc:
dev/tsunami_io.cc:
dev/tsunami_pchip.cc:
Added using namespace TheISA. It might be better for it to be AlphaISA
dev/tsunami_cchip.cc:
Added typedef for TheISA. It might be better for it to be AlphaISA
kern/linux/aligned.hh:
sim/pseudo_inst.hh:
Added TheISA namespace specifier to Addr
kern/linux/linux_threadinfo.hh:
Added typedef for Addr, and TheISA namespace specifier to StackPointerReg
kern/tru64/mbuf.hh:
Added TheISA to Addr type in structs
sim/process.hh:
Added typedefs of Addr, RegFile, and MachInst
sim/syscall_emul.cc:
Added using namespace TheISA, and a cast of VMPageSize to the int type
sim/syscall_emul.hh:
Added typecast for Addr, and TheISA namespace specifier for where needed
--HG--
extra : convert_revision : 91d4f6ca33a73b21c1f1771d74bfdea3b80eff45
2006-02-19 08:34:37 +01:00
|
|
|
typedef TheISA::IntReg IntReg;
|
2006-06-08 22:58:50 +02:00
|
|
|
// Floating point register type.
|
|
|
|
typedef TheISA::FloatReg FloatReg;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
// The DynInstPtr type.
|
|
|
|
typedef typename Impl::DynInstPtr DynInstPtr;
|
|
|
|
|
|
|
|
// The list of instructions iterator type.
|
|
|
|
typedef typename std::list<DynInstPtr>::iterator ListIt;
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
enum {
|
2008-09-10 20:26:15 +02:00
|
|
|
MaxInstSrcRegs = TheISA::MaxInstSrcRegs, /// Max source regs
|
|
|
|
MaxInstDestRegs = TheISA::MaxInstDestRegs, /// Max dest regs
|
2004-08-20 20:54:07 +02:00
|
|
|
};
|
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/** The StaticInst used by this BaseDynInst. */
|
Changes to untemplate StaticInst and StaticInstPtr, change the isa to a namespace instead of a class, an improvement to the architecture specific header file selection system, and fixed up a few include paths.
arch/alpha/alpha_linux_process.cc:
Added using directive for AlphaISA namespace
arch/alpha/alpha_memory.hh:
arch/alpha/isa/branch.isa:
cpu/pc_event.hh:
Added typedefs for Addr
arch/alpha/alpha_tru64_process.cc:
arch/alpha/arguments.cc:
Added using directive for AlphaISA
arch/alpha/ev5.hh:
Added an include of arch/alpha/isa_traits.hh, and a using directive for the AlphaISA namespace.
arch/alpha/faults.hh:
Added a typedef for the Addr type, and changed the formatting of the faults slightly.
arch/alpha/isa/main.isa:
Untemplatized StaticInst, added a using for namespace AlphaISA to show up in decoder.cc and the exec.ccs, relocated makeNop to decoder.hh
arch/alpha/isa/mem.isa:
Untemplatized StaticInst and StaticInstPtr
arch/alpha/isa/pal.isa:
cpu/base_dyn_inst.cc:
Untemplatized StaticInstPtr
arch/alpha/isa_traits.hh:
Changed variables to be externs instead of static since they are part of a namespace and not a class.
arch/alpha/stacktrace.cc:
Untemplatized StaticInstPtr, and added a using directive for AlphaISA.
arch/alpha/stacktrace.hh:
Added some typedefs for Addr and MachInst, and untemplatized StaticInstPtr
arch/alpha/vtophys.cc:
Added a using directive for AlphaISA
arch/alpha/vtophys.hh:
Added the AlphaISA namespace specifier where needed
arch/isa_parser.py:
Changed the placement of the definition of the decodeInst function to be outside the namespaceInst namespace.
base/loader/object_file.hh:
cpu/o3/bpred_unit.hh:
Added a typedef for Addr
base/loader/symtab.hh:
Added a typedef for Addr, and added a TheISA to Addr in another typedef
base/remote_gdb.cc:
Added a using namespace TheISA, and untemplatized StaticInstPtr
base/remote_gdb.hh:
Added typedefs for Addr and MachInst
cpu/base.cc:
Added TheISA specifier to some variables exported from the isa.
cpu/base.hh:
Added a typedef for Addr, and TheISA to some variables from the ISA
cpu/base_dyn_inst.hh:
Untemplatized StaticInstPtr, and added TheISA specifier to some variables from the ISA.
cpu/exec_context.hh:
Added some typedefs for types from the isa, and added TheISA specifier to some variables from the isa
cpu/exetrace.hh:
Added typedefs for some types from the ISA, and untemplatized StaticInstPtr
cpu/memtest/memtest.cc:
cpu/o3/btb.cc:
dev/baddev.cc:
dev/ide_ctrl.cc:
dev/ide_disk.cc:
dev/isa_fake.cc:
dev/ns_gige.cc:
dev/pciconfigall.cc:
dev/platform.cc:
dev/sinic.cc:
dev/uart8250.cc:
kern/freebsd/freebsd_system.cc:
kern/linux/linux_system.cc:
kern/system_events.cc:
kern/tru64/dump_mbuf.cc:
kern/tru64/tru64_events.cc:
sim/process.cc:
sim/pseudo_inst.cc:
sim/system.cc:
Added using namespace TheISA
cpu/memtest/memtest.hh:
cpu/trace/opt_cpu.hh:
cpu/trace/reader/itx_reader.hh:
dev/ide_disk.hh:
dev/pcidev.hh:
dev/platform.hh:
dev/tsunami.hh:
sim/system.hh:
sim/vptr.hh:
Added typedef for Addr
cpu/o3/2bit_local_pred.hh:
Changed the include to use arch/isa_traits.hh instead of arch/alpha/isa_traits.hh. Added typedef for Addr
cpu/o3/alpha_cpu.hh:
Added typedefs for Addr and IntReg
cpu/o3/alpha_cpu_impl.hh:
Added this-> to setNextPC to fix a problem since it didn't depend on template parameters any more. Removed "typename" where it was no longer needed.
cpu/o3/alpha_dyn_inst.hh:
Cleaned up some typedefs, and untemplatized StaticInst
cpu/o3/alpha_dyn_inst_impl.hh:
untemplatized StaticInstPtr
cpu/o3/alpha_impl.hh:
Fixed up a typedef of MachInst
cpu/o3/bpred_unit_impl.hh:
Added a using TheISA::MachInst to a function
cpu/o3/btb.hh:
Changed an include from arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef for Addr
cpu/o3/commit.hh:
Removed a typedef of Impl::ISA as ISA, since TheISA takes care of this now.
cpu/o3/cpu.cc:
Cleaned up namespace issues
cpu/o3/cpu.hh:
Cleaned up namespace usage
cpu/o3/decode.hh:
Removed typedef of ISA, and changed it to TheISA
cpu/o3/fetch.hh:
Fized up typedefs, and changed ISA to TheISA
cpu/o3/free_list.hh:
Changed include of arch/alpha/isa_traits.hh to arch/isa_traits.hh
cpu/o3/iew.hh:
Removed typedef of ISA
cpu/o3/iew_impl.hh:
Added TheISA namespace specifier to MachInst
cpu/o3/ras.hh:
Changed include from arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef for Addr.
cpu/o3/regfile.hh:
Changed ISA to TheISA, and added some typedefs for Addr, IntReg, FloatReg, and MiscRegFile
cpu/o3/rename.hh:
Changed ISA to TheISA, and added a typedef for RegIndex
cpu/o3/rename_map.hh:
Added an include for arch/isa_traits.hh, and a typedef for RegIndex
cpu/o3/rob.hh:
Added a typedef for RegIndex
cpu/o3/store_set.hh:
cpu/o3/tournament_pred.hh:
Changed an include of arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef of Addr
cpu/ozone/cpu.hh:
Changed ISA into TheISA, and untemplatized StaticInst
cpu/pc_event.cc:
Added namespace specifier TheISA to Addr types
cpu/profile.hh:
kern/kernel_stats.hh:
Added typedef for Addr, and untemplatized StaticInstPtr
cpu/simple/cpu.cc:
Changed using directive from LittleEndianGuest to AlphaISA, which will contain both namespaces. Added TheISA where needed, and untemplatized StaticInst
cpu/simple/cpu.hh:
Added a typedef for MachInst, and untemplatized StaticInst
cpu/static_inst.cc:
Untemplatized StaticInst
cpu/static_inst.hh:
Untemplatized StaticInst by using the TheISA namespace
dev/alpha_console.cc:
Added using namespace AlphaISA
dev/simple_disk.hh:
Added typedef for Addr and fixed up some formatting
dev/sinicreg.hh:
Added TheISA namespace specifier where needed
dev/tsunami.cc:
dev/tsunami_io.cc:
dev/tsunami_pchip.cc:
Added using namespace TheISA. It might be better for it to be AlphaISA
dev/tsunami_cchip.cc:
Added typedef for TheISA. It might be better for it to be AlphaISA
kern/linux/aligned.hh:
sim/pseudo_inst.hh:
Added TheISA namespace specifier to Addr
kern/linux/linux_threadinfo.hh:
Added typedef for Addr, and TheISA namespace specifier to StackPointerReg
kern/tru64/mbuf.hh:
Added TheISA to Addr type in structs
sim/process.hh:
Added typedefs of Addr, RegFile, and MachInst
sim/syscall_emul.cc:
Added using namespace TheISA, and a cast of VMPageSize to the int type
sim/syscall_emul.hh:
Added typecast for Addr, and TheISA namespace specifier for where needed
--HG--
extra : convert_revision : 91d4f6ca33a73b21c1f1771d74bfdea3b80eff45
2006-02-19 08:34:37 +01:00
|
|
|
StaticInstPtr staticInst;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
////////////////////////////////////////////
|
|
|
|
//
|
|
|
|
// INSTRUCTION EXECUTION
|
|
|
|
//
|
|
|
|
////////////////////////////////////////////
|
2006-04-23 00:26:48 +02:00
|
|
|
/** InstRecord that tracks this instructions. */
|
2004-08-20 20:54:07 +02:00
|
|
|
Trace::InstRecord *traceData;
|
|
|
|
|
2008-02-27 05:38:51 +01:00
|
|
|
void demapPage(Addr vaddr, uint64_t asn)
|
|
|
|
{
|
|
|
|
cpu->demapPage(vaddr, asn);
|
|
|
|
}
|
|
|
|
void demapInstPage(Addr vaddr, uint64_t asn)
|
|
|
|
{
|
|
|
|
cpu->demapPage(vaddr, asn);
|
|
|
|
}
|
|
|
|
void demapDataPage(Addr vaddr, uint64_t asn)
|
|
|
|
{
|
|
|
|
cpu->demapPage(vaddr, asn);
|
|
|
|
}
|
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/**
|
|
|
|
* Does a read to a given address.
|
|
|
|
* @param addr The address to read.
|
|
|
|
* @param data The read's data is written into this parameter.
|
|
|
|
* @param flags The request's flags.
|
|
|
|
* @return Returns any fault due to the read.
|
|
|
|
*/
|
2004-08-20 20:54:07 +02:00
|
|
|
template <class T>
|
2006-02-22 02:10:40 +01:00
|
|
|
Fault read(Addr addr, T &data, unsigned flags);
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2010-08-13 15:16:02 +02:00
|
|
|
Fault readBytes(Addr addr, uint8_t *data, unsigned size, unsigned flags);
|
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/**
|
|
|
|
* Does a write to a given address.
|
|
|
|
* @param data The data to be written.
|
|
|
|
* @param addr The address to write to.
|
|
|
|
* @param flags The request's flags.
|
|
|
|
* @param res The result of the write (for load locked/store conditionals).
|
|
|
|
* @return Returns any fault due to the write.
|
|
|
|
*/
|
2004-08-20 20:54:07 +02:00
|
|
|
template <class T>
|
2010-02-12 20:53:19 +01:00
|
|
|
Fault write(T data, Addr addr, unsigned flags, uint64_t *res);
|
|
|
|
|
2010-08-13 15:16:02 +02:00
|
|
|
Fault writeBytes(uint8_t *data, unsigned size,
|
|
|
|
Addr addr, unsigned flags, uint64_t *res);
|
|
|
|
|
2010-02-12 20:53:20 +01:00
|
|
|
/** Splits a request in two if it crosses a dcache block. */
|
|
|
|
void splitRequest(RequestPtr req, RequestPtr &sreqLow,
|
|
|
|
RequestPtr &sreqHigh);
|
|
|
|
|
2010-02-12 20:53:19 +01:00
|
|
|
/** Initiate a DTB address translation. */
|
2010-02-12 20:53:20 +01:00
|
|
|
void initiateTranslation(RequestPtr req, RequestPtr sreqLow,
|
|
|
|
RequestPtr sreqHigh, uint64_t *res,
|
2010-02-12 20:53:19 +01:00
|
|
|
BaseTLB::Mode mode);
|
|
|
|
|
|
|
|
/** Finish a DTB address translation. */
|
|
|
|
void finishTranslation(WholeTranslationState *state);
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2011-02-12 01:29:35 +01:00
|
|
|
/** True if the DTB address translation has started. */
|
|
|
|
bool translationStarted;
|
|
|
|
|
|
|
|
/** True if the DTB address translation has completed. */
|
|
|
|
bool translationCompleted;
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Returns true if the DTB address translation is being delayed due to a hw
|
|
|
|
* page table walk.
|
|
|
|
*/
|
|
|
|
bool isTranslationDelayed() const
|
|
|
|
{
|
|
|
|
return (translationStarted && !translationCompleted);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Saved memory requests (needed when the DTB address translation is
|
|
|
|
* delayed due to a hw page table walk).
|
|
|
|
*/
|
|
|
|
RequestPtr savedReq;
|
|
|
|
RequestPtr savedSreqLow;
|
|
|
|
RequestPtr savedSreqHigh;
|
|
|
|
|
2005-05-19 07:28:25 +02:00
|
|
|
/** @todo: Consider making this private. */
|
2004-08-20 20:54:07 +02:00
|
|
|
public:
|
|
|
|
/** The sequence number of the instruction. */
|
|
|
|
InstSeqNum seqNum;
|
|
|
|
|
2006-06-14 19:12:41 +02:00
|
|
|
enum Status {
|
|
|
|
IqEntry, /// Instruction is in the IQ
|
|
|
|
RobEntry, /// Instruction is in the ROB
|
|
|
|
LsqEntry, /// Instruction is in the LSQ
|
|
|
|
Completed, /// Instruction has completed
|
|
|
|
ResultReady, /// Instruction has its result
|
|
|
|
CanIssue, /// Instruction can issue and execute
|
|
|
|
Issued, /// Instruction has issued
|
|
|
|
Executed, /// Instruction has executed
|
|
|
|
CanCommit, /// Instruction can commit
|
|
|
|
AtCommit, /// Instruction has reached commit
|
|
|
|
Committed, /// Instruction has committed
|
|
|
|
Squashed, /// Instruction is squashed
|
|
|
|
SquashedInIQ, /// Instruction is squashed in the IQ
|
|
|
|
SquashedInLSQ, /// Instruction is squashed in the LSQ
|
|
|
|
SquashedInROB, /// Instruction is squashed in the ROB
|
|
|
|
RecoverInst, /// Is a recover instruction
|
|
|
|
BlockingInst, /// Is a blocking instruction
|
|
|
|
ThreadsyncWait, /// Is a thread synchronization instruction
|
|
|
|
SerializeBefore, /// Needs to serialize on
|
|
|
|
/// instructions ahead of it
|
|
|
|
SerializeAfter, /// Needs to serialize instructions behind it
|
|
|
|
SerializeHandled, /// Serialization has been handled
|
|
|
|
NumStatus
|
|
|
|
};
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-06-14 19:12:41 +02:00
|
|
|
/** The status of this BaseDynInst. Several bits can be set. */
|
|
|
|
std::bitset<NumStatus> status;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** The thread this instruction is from. */
|
2009-05-26 18:23:13 +02:00
|
|
|
ThreadID threadNumber;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** data address space ID, for loads & stores. */
|
|
|
|
short asid;
|
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/** How many source registers are ready. */
|
|
|
|
unsigned readyRegs;
|
|
|
|
|
2006-06-16 23:08:47 +02:00
|
|
|
/** Pointer to the Impl's CPU object. */
|
|
|
|
ImplCPU *cpu;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-06-06 23:32:21 +02:00
|
|
|
/** Pointer to the thread state. */
|
2006-04-23 00:26:48 +02:00
|
|
|
ImplState *thread;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** The kind of fault this instruction has generated. */
|
2006-02-22 02:10:40 +01:00
|
|
|
Fault fault;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-06-13 01:04:42 +02:00
|
|
|
/** Pointer to the data for the memory access. */
|
2006-06-03 00:15:20 +02:00
|
|
|
uint8_t *memData;
|
2006-04-23 00:26:48 +02:00
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** The effective virtual address (lds & stores only). */
|
|
|
|
Addr effAddr;
|
|
|
|
|
2007-03-23 16:33:08 +01:00
|
|
|
/** Is the effective virtual address valid. */
|
|
|
|
bool effAddrValid;
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** The effective physical address. */
|
|
|
|
Addr physEffAddr;
|
|
|
|
|
|
|
|
/** Effective virtual address for a copy source. */
|
|
|
|
Addr copySrcEffAddr;
|
|
|
|
|
|
|
|
/** Effective physical address for a copy source. */
|
|
|
|
Addr copySrcPhysEffAddr;
|
|
|
|
|
|
|
|
/** The memory request flags (from translation). */
|
|
|
|
unsigned memReqFlags;
|
|
|
|
|
2005-02-26 00:00:49 +01:00
|
|
|
union Result {
|
|
|
|
uint64_t integer;
|
2006-08-11 23:42:59 +02:00
|
|
|
// float fp;
|
2005-02-26 00:00:49 +01:00
|
|
|
double dbl;
|
|
|
|
};
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2005-02-26 00:00:49 +01:00
|
|
|
/** The result of the instruction; assumes for now that there's only one
|
|
|
|
* destination register.
|
|
|
|
*/
|
|
|
|
Result instResult;
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-10-23 20:00:07 +02:00
|
|
|
/** Records changes to result? */
|
|
|
|
bool recordResult;
|
|
|
|
|
2010-08-23 18:18:40 +02:00
|
|
|
/** Did this instruction execute, or is it predicated false */
|
|
|
|
bool predicate;
|
|
|
|
|
2006-12-28 20:27:45 +01:00
|
|
|
protected:
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
/** PC state for this instruction. */
|
|
|
|
TheISA::PCState pc;
|
2006-07-23 19:39:42 +02:00
|
|
|
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
/** Predicted PC state after this instruction. */
|
|
|
|
TheISA::PCState predPC;
|
2007-04-14 19:13:18 +02:00
|
|
|
|
2006-12-16 13:22:19 +01:00
|
|
|
/** If this is a branch that was predicted taken */
|
|
|
|
bool predTaken;
|
|
|
|
|
2006-12-28 20:27:45 +01:00
|
|
|
public:
|
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
#ifdef DEBUG
|
|
|
|
void dumpSNList();
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/** Whether or not the source register is ready.
|
|
|
|
* @todo: Not sure this should be here vs the derived class.
|
2004-08-20 20:54:07 +02:00
|
|
|
*/
|
|
|
|
bool _readySrcRegIdx[MaxInstSrcRegs];
|
|
|
|
|
2006-12-06 11:48:59 +01:00
|
|
|
protected:
|
|
|
|
/** Flattened register index of the destination registers of this
|
|
|
|
* instruction.
|
|
|
|
*/
|
|
|
|
TheISA::RegIndex _flatDestRegIdx[TheISA::MaxInstDestRegs];
|
|
|
|
|
|
|
|
/** Flattened register index of the source registers of this
|
|
|
|
* instruction.
|
|
|
|
*/
|
|
|
|
TheISA::RegIndex _flatSrcRegIdx[TheISA::MaxInstSrcRegs];
|
|
|
|
|
|
|
|
/** Physical register index of the destination registers of this
|
|
|
|
* instruction.
|
|
|
|
*/
|
|
|
|
PhysRegIndex _destRegIdx[TheISA::MaxInstDestRegs];
|
|
|
|
|
|
|
|
/** Physical register index of the source registers of this
|
|
|
|
* instruction.
|
|
|
|
*/
|
|
|
|
PhysRegIndex _srcRegIdx[TheISA::MaxInstSrcRegs];
|
|
|
|
|
|
|
|
/** Physical register index of the previous producers of the
|
|
|
|
* architected destinations.
|
|
|
|
*/
|
|
|
|
PhysRegIndex _prevDestRegIdx[TheISA::MaxInstDestRegs];
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
public:
|
2006-12-06 11:48:59 +01:00
|
|
|
|
|
|
|
/** Returns the physical register index of the i'th destination
|
|
|
|
* register.
|
|
|
|
*/
|
|
|
|
PhysRegIndex renamedDestRegIdx(int idx) const
|
|
|
|
{
|
|
|
|
return _destRegIdx[idx];
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Returns the physical register index of the i'th source register. */
|
|
|
|
PhysRegIndex renamedSrcRegIdx(int idx) const
|
|
|
|
{
|
|
|
|
return _srcRegIdx[idx];
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Returns the flattened register index of the i'th destination
|
|
|
|
* register.
|
|
|
|
*/
|
|
|
|
TheISA::RegIndex flattenedDestRegIdx(int idx) const
|
|
|
|
{
|
|
|
|
return _flatDestRegIdx[idx];
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Returns the flattened register index of the i'th source register */
|
|
|
|
TheISA::RegIndex flattenedSrcRegIdx(int idx) const
|
|
|
|
{
|
|
|
|
return _flatSrcRegIdx[idx];
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Returns the physical register index of the previous physical register
|
|
|
|
* that remapped to the same logical register index.
|
|
|
|
*/
|
|
|
|
PhysRegIndex prevDestRegIdx(int idx) const
|
|
|
|
{
|
|
|
|
return _prevDestRegIdx[idx];
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Renames a destination register to a physical register. Also records
|
|
|
|
* the previous physical register that the logical register mapped to.
|
|
|
|
*/
|
|
|
|
void renameDestReg(int idx,
|
|
|
|
PhysRegIndex renamed_dest,
|
|
|
|
PhysRegIndex previous_rename)
|
|
|
|
{
|
|
|
|
_destRegIdx[idx] = renamed_dest;
|
|
|
|
_prevDestRegIdx[idx] = previous_rename;
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Renames a source logical register to the physical register which
|
|
|
|
* has/will produce that logical register's result.
|
|
|
|
* @todo: add in whether or not the source register is ready.
|
|
|
|
*/
|
|
|
|
void renameSrcReg(int idx, PhysRegIndex renamed_src)
|
|
|
|
{
|
|
|
|
_srcRegIdx[idx] = renamed_src;
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Flattens a source architectural register index into a logical index.
|
|
|
|
*/
|
|
|
|
void flattenSrcReg(int idx, TheISA::RegIndex flattened_src)
|
|
|
|
{
|
|
|
|
_flatSrcRegIdx[idx] = flattened_src;
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Flattens a destination architectural register index into a logical
|
|
|
|
* index.
|
|
|
|
*/
|
|
|
|
void flattenDestReg(int idx, TheISA::RegIndex flattened_dest)
|
|
|
|
{
|
|
|
|
_flatDestRegIdx[idx] = flattened_dest;
|
|
|
|
}
|
2007-04-14 19:13:18 +02:00
|
|
|
/** BaseDynInst constructor given a binary instruction.
|
|
|
|
* @param staticInst A StaticInstPtr to the underlying instruction.
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
* @param pc The PC state for the instruction.
|
|
|
|
* @param predPC The predicted next PC state for the instruction.
|
2007-04-14 19:13:18 +02:00
|
|
|
* @param seq_num The sequence number of the instruction.
|
|
|
|
* @param cpu Pointer to the instruction's CPU.
|
|
|
|
*/
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
BaseDynInst(StaticInstPtr staticInst, TheISA::PCState pc,
|
|
|
|
TheISA::PCState predPC, InstSeqNum seq_num, ImplCPU *cpu);
|
2006-12-06 11:48:59 +01:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/** BaseDynInst constructor given a binary instruction.
|
|
|
|
* @param inst The binary instruction.
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
* @param _pc The PC state for the instruction.
|
|
|
|
* @param _predPC The predicted next PC state for the instruction.
|
2006-04-23 00:26:48 +02:00
|
|
|
* @param seq_num The sequence number of the instruction.
|
|
|
|
* @param cpu Pointer to the instruction's CPU.
|
|
|
|
*/
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
BaseDynInst(TheISA::ExtMachInst inst, TheISA::PCState pc,
|
|
|
|
TheISA::PCState predPC, InstSeqNum seq_num, ImplCPU *cpu);
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/** BaseDynInst constructor given a StaticInst pointer.
|
|
|
|
* @param _staticInst The StaticInst for this BaseDynInst.
|
|
|
|
*/
|
Changes to untemplate StaticInst and StaticInstPtr, change the isa to a namespace instead of a class, an improvement to the architecture specific header file selection system, and fixed up a few include paths.
arch/alpha/alpha_linux_process.cc:
Added using directive for AlphaISA namespace
arch/alpha/alpha_memory.hh:
arch/alpha/isa/branch.isa:
cpu/pc_event.hh:
Added typedefs for Addr
arch/alpha/alpha_tru64_process.cc:
arch/alpha/arguments.cc:
Added using directive for AlphaISA
arch/alpha/ev5.hh:
Added an include of arch/alpha/isa_traits.hh, and a using directive for the AlphaISA namespace.
arch/alpha/faults.hh:
Added a typedef for the Addr type, and changed the formatting of the faults slightly.
arch/alpha/isa/main.isa:
Untemplatized StaticInst, added a using for namespace AlphaISA to show up in decoder.cc and the exec.ccs, relocated makeNop to decoder.hh
arch/alpha/isa/mem.isa:
Untemplatized StaticInst and StaticInstPtr
arch/alpha/isa/pal.isa:
cpu/base_dyn_inst.cc:
Untemplatized StaticInstPtr
arch/alpha/isa_traits.hh:
Changed variables to be externs instead of static since they are part of a namespace and not a class.
arch/alpha/stacktrace.cc:
Untemplatized StaticInstPtr, and added a using directive for AlphaISA.
arch/alpha/stacktrace.hh:
Added some typedefs for Addr and MachInst, and untemplatized StaticInstPtr
arch/alpha/vtophys.cc:
Added a using directive for AlphaISA
arch/alpha/vtophys.hh:
Added the AlphaISA namespace specifier where needed
arch/isa_parser.py:
Changed the placement of the definition of the decodeInst function to be outside the namespaceInst namespace.
base/loader/object_file.hh:
cpu/o3/bpred_unit.hh:
Added a typedef for Addr
base/loader/symtab.hh:
Added a typedef for Addr, and added a TheISA to Addr in another typedef
base/remote_gdb.cc:
Added a using namespace TheISA, and untemplatized StaticInstPtr
base/remote_gdb.hh:
Added typedefs for Addr and MachInst
cpu/base.cc:
Added TheISA specifier to some variables exported from the isa.
cpu/base.hh:
Added a typedef for Addr, and TheISA to some variables from the ISA
cpu/base_dyn_inst.hh:
Untemplatized StaticInstPtr, and added TheISA specifier to some variables from the ISA.
cpu/exec_context.hh:
Added some typedefs for types from the isa, and added TheISA specifier to some variables from the isa
cpu/exetrace.hh:
Added typedefs for some types from the ISA, and untemplatized StaticInstPtr
cpu/memtest/memtest.cc:
cpu/o3/btb.cc:
dev/baddev.cc:
dev/ide_ctrl.cc:
dev/ide_disk.cc:
dev/isa_fake.cc:
dev/ns_gige.cc:
dev/pciconfigall.cc:
dev/platform.cc:
dev/sinic.cc:
dev/uart8250.cc:
kern/freebsd/freebsd_system.cc:
kern/linux/linux_system.cc:
kern/system_events.cc:
kern/tru64/dump_mbuf.cc:
kern/tru64/tru64_events.cc:
sim/process.cc:
sim/pseudo_inst.cc:
sim/system.cc:
Added using namespace TheISA
cpu/memtest/memtest.hh:
cpu/trace/opt_cpu.hh:
cpu/trace/reader/itx_reader.hh:
dev/ide_disk.hh:
dev/pcidev.hh:
dev/platform.hh:
dev/tsunami.hh:
sim/system.hh:
sim/vptr.hh:
Added typedef for Addr
cpu/o3/2bit_local_pred.hh:
Changed the include to use arch/isa_traits.hh instead of arch/alpha/isa_traits.hh. Added typedef for Addr
cpu/o3/alpha_cpu.hh:
Added typedefs for Addr and IntReg
cpu/o3/alpha_cpu_impl.hh:
Added this-> to setNextPC to fix a problem since it didn't depend on template parameters any more. Removed "typename" where it was no longer needed.
cpu/o3/alpha_dyn_inst.hh:
Cleaned up some typedefs, and untemplatized StaticInst
cpu/o3/alpha_dyn_inst_impl.hh:
untemplatized StaticInstPtr
cpu/o3/alpha_impl.hh:
Fixed up a typedef of MachInst
cpu/o3/bpred_unit_impl.hh:
Added a using TheISA::MachInst to a function
cpu/o3/btb.hh:
Changed an include from arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef for Addr
cpu/o3/commit.hh:
Removed a typedef of Impl::ISA as ISA, since TheISA takes care of this now.
cpu/o3/cpu.cc:
Cleaned up namespace issues
cpu/o3/cpu.hh:
Cleaned up namespace usage
cpu/o3/decode.hh:
Removed typedef of ISA, and changed it to TheISA
cpu/o3/fetch.hh:
Fized up typedefs, and changed ISA to TheISA
cpu/o3/free_list.hh:
Changed include of arch/alpha/isa_traits.hh to arch/isa_traits.hh
cpu/o3/iew.hh:
Removed typedef of ISA
cpu/o3/iew_impl.hh:
Added TheISA namespace specifier to MachInst
cpu/o3/ras.hh:
Changed include from arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef for Addr.
cpu/o3/regfile.hh:
Changed ISA to TheISA, and added some typedefs for Addr, IntReg, FloatReg, and MiscRegFile
cpu/o3/rename.hh:
Changed ISA to TheISA, and added a typedef for RegIndex
cpu/o3/rename_map.hh:
Added an include for arch/isa_traits.hh, and a typedef for RegIndex
cpu/o3/rob.hh:
Added a typedef for RegIndex
cpu/o3/store_set.hh:
cpu/o3/tournament_pred.hh:
Changed an include of arch/alpha/isa_traits.hh to arch/isa_traits.hh, and added a typedef of Addr
cpu/ozone/cpu.hh:
Changed ISA into TheISA, and untemplatized StaticInst
cpu/pc_event.cc:
Added namespace specifier TheISA to Addr types
cpu/profile.hh:
kern/kernel_stats.hh:
Added typedef for Addr, and untemplatized StaticInstPtr
cpu/simple/cpu.cc:
Changed using directive from LittleEndianGuest to AlphaISA, which will contain both namespaces. Added TheISA where needed, and untemplatized StaticInst
cpu/simple/cpu.hh:
Added a typedef for MachInst, and untemplatized StaticInst
cpu/static_inst.cc:
Untemplatized StaticInst
cpu/static_inst.hh:
Untemplatized StaticInst by using the TheISA namespace
dev/alpha_console.cc:
Added using namespace AlphaISA
dev/simple_disk.hh:
Added typedef for Addr and fixed up some formatting
dev/sinicreg.hh:
Added TheISA namespace specifier where needed
dev/tsunami.cc:
dev/tsunami_io.cc:
dev/tsunami_pchip.cc:
Added using namespace TheISA. It might be better for it to be AlphaISA
dev/tsunami_cchip.cc:
Added typedef for TheISA. It might be better for it to be AlphaISA
kern/linux/aligned.hh:
sim/pseudo_inst.hh:
Added TheISA namespace specifier to Addr
kern/linux/linux_threadinfo.hh:
Added typedef for Addr, and TheISA namespace specifier to StackPointerReg
kern/tru64/mbuf.hh:
Added TheISA to Addr type in structs
sim/process.hh:
Added typedefs of Addr, RegFile, and MachInst
sim/syscall_emul.cc:
Added using namespace TheISA, and a cast of VMPageSize to the int type
sim/syscall_emul.hh:
Added typecast for Addr, and TheISA namespace specifier for where needed
--HG--
extra : convert_revision : 91d4f6ca33a73b21c1f1771d74bfdea3b80eff45
2006-02-19 08:34:37 +01:00
|
|
|
BaseDynInst(StaticInstPtr &_staticInst);
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** BaseDynInst destructor. */
|
|
|
|
~BaseDynInst();
|
|
|
|
|
2005-02-26 00:00:49 +01:00
|
|
|
private:
|
2005-05-19 07:28:25 +02:00
|
|
|
/** Function to initialize variables in the constructors. */
|
2005-02-26 00:00:49 +01:00
|
|
|
void initVars();
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2005-02-26 00:00:49 +01:00
|
|
|
public:
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Dumps out contents of this BaseDynInst. */
|
|
|
|
void dump();
|
|
|
|
|
|
|
|
/** Dumps out contents of this BaseDynInst into given string. */
|
|
|
|
void dump(std::string &outstring);
|
|
|
|
|
2006-10-23 20:00:07 +02:00
|
|
|
/** Read this CPU's ID. */
|
2008-11-03 03:56:57 +01:00
|
|
|
int cpuId() { return cpu->cpuId(); }
|
2006-10-23 20:00:07 +02:00
|
|
|
|
2008-11-03 03:57:07 +01:00
|
|
|
/** Read this context's system-wide ID **/
|
|
|
|
int contextId() { return thread->contextId(); }
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Returns the fault type. */
|
2006-02-22 02:10:40 +01:00
|
|
|
Fault getFault() { return fault; }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** Checks whether or not this instruction has had its branch target
|
|
|
|
* calculated yet. For now it is not utilized and is hacked to be
|
|
|
|
* always false.
|
2006-04-23 00:26:48 +02:00
|
|
|
* @todo: Actually use this instruction.
|
2004-08-20 20:54:07 +02:00
|
|
|
*/
|
|
|
|
bool doneTargCalc() { return false; }
|
|
|
|
|
|
|
|
/** Set the predicted target of this current instruction. */
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
void setPredTarg(const TheISA::PCState &_predPC)
|
2006-12-16 13:22:19 +01:00
|
|
|
{
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
predPC = _predPC;
|
2006-12-16 13:22:19 +01:00
|
|
|
}
|
2004-08-20 20:54:07 +02:00
|
|
|
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
const TheISA::PCState &readPredTarg() { return predPC; }
|
|
|
|
|
2006-12-16 13:22:19 +01:00
|
|
|
/** Returns the predicted PC immediately after the branch. */
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
Addr predInstAddr() { return predPC.instAddr(); }
|
2006-12-16 13:22:19 +01:00
|
|
|
|
|
|
|
/** Returns the predicted PC two instructions after the branch */
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
Addr predNextInstAddr() { return predPC.nextInstAddr(); }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2007-04-14 19:13:18 +02:00
|
|
|
/** Returns the predicted micro PC after the branch */
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
Addr predMicroPC() { return predPC.microPC(); }
|
2007-04-14 19:13:18 +02:00
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Returns whether the instruction was predicted taken or not. */
|
2006-12-16 13:22:19 +01:00
|
|
|
bool readPredTaken()
|
|
|
|
{
|
|
|
|
return predTaken;
|
|
|
|
}
|
|
|
|
|
|
|
|
void setPredTaken(bool predicted_taken)
|
|
|
|
{
|
|
|
|
predTaken = predicted_taken;
|
|
|
|
}
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** Returns whether the instruction mispredicted. */
|
2006-07-23 19:39:42 +02:00
|
|
|
bool mispredicted()
|
2006-12-16 13:22:19 +01:00
|
|
|
{
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
TheISA::PCState tempPC = pc;
|
|
|
|
TheISA::advancePC(tempPC, staticInst);
|
|
|
|
return !(tempPC == predPC);
|
2006-12-16 13:22:19 +01:00
|
|
|
}
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
//
|
|
|
|
// Instruction types. Forward checks to StaticInst object.
|
|
|
|
//
|
2008-09-10 20:26:15 +02:00
|
|
|
bool isNop() const { return staticInst->isNop(); }
|
|
|
|
bool isMemRef() const { return staticInst->isMemRef(); }
|
|
|
|
bool isLoad() const { return staticInst->isLoad(); }
|
|
|
|
bool isStore() const { return staticInst->isStore(); }
|
2006-05-23 20:38:16 +02:00
|
|
|
bool isStoreConditional() const
|
|
|
|
{ return staticInst->isStoreConditional(); }
|
2004-08-20 20:54:07 +02:00
|
|
|
bool isInstPrefetch() const { return staticInst->isInstPrefetch(); }
|
|
|
|
bool isDataPrefetch() const { return staticInst->isDataPrefetch(); }
|
|
|
|
bool isCopy() const { return staticInst->isCopy(); }
|
2008-09-10 20:26:15 +02:00
|
|
|
bool isInteger() const { return staticInst->isInteger(); }
|
|
|
|
bool isFloating() const { return staticInst->isFloating(); }
|
|
|
|
bool isControl() const { return staticInst->isControl(); }
|
|
|
|
bool isCall() const { return staticInst->isCall(); }
|
|
|
|
bool isReturn() const { return staticInst->isReturn(); }
|
|
|
|
bool isDirectCtrl() const { return staticInst->isDirectCtrl(); }
|
2004-08-20 20:54:07 +02:00
|
|
|
bool isIndirectCtrl() const { return staticInst->isIndirectCtrl(); }
|
2008-09-10 20:26:15 +02:00
|
|
|
bool isCondCtrl() const { return staticInst->isCondCtrl(); }
|
|
|
|
bool isUncondCtrl() const { return staticInst->isUncondCtrl(); }
|
2006-07-23 19:39:42 +02:00
|
|
|
bool isCondDelaySlot() const { return staticInst->isCondDelaySlot(); }
|
2004-08-20 20:54:07 +02:00
|
|
|
bool isThreadSync() const { return staticInst->isThreadSync(); }
|
|
|
|
bool isSerializing() const { return staticInst->isSerializing(); }
|
2006-04-23 00:26:48 +02:00
|
|
|
bool isSerializeBefore() const
|
2006-06-14 19:12:41 +02:00
|
|
|
{ return staticInst->isSerializeBefore() || status[SerializeBefore]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
bool isSerializeAfter() const
|
2006-06-14 19:12:41 +02:00
|
|
|
{ return staticInst->isSerializeAfter() || status[SerializeAfter]; }
|
2010-12-08 01:19:57 +01:00
|
|
|
bool isSquashAfter() const { return staticInst->isSquashAfter(); }
|
2004-08-20 20:54:07 +02:00
|
|
|
bool isMemBarrier() const { return staticInst->isMemBarrier(); }
|
|
|
|
bool isWriteBarrier() const { return staticInst->isWriteBarrier(); }
|
|
|
|
bool isNonSpeculative() const { return staticInst->isNonSpeculative(); }
|
2006-04-23 00:26:48 +02:00
|
|
|
bool isQuiesce() const { return staticInst->isQuiesce(); }
|
2006-05-23 20:38:16 +02:00
|
|
|
bool isIprAccess() const { return staticInst->isIprAccess(); }
|
2006-05-11 20:12:34 +02:00
|
|
|
bool isUnverifiable() const { return staticInst->isUnverifiable(); }
|
2007-08-01 02:34:08 +02:00
|
|
|
bool isSyscall() const { return staticInst->isSyscall(); }
|
2007-06-21 04:46:45 +02:00
|
|
|
bool isMacroop() const { return staticInst->isMacroop(); }
|
|
|
|
bool isMicroop() const { return staticInst->isMicroop(); }
|
2007-04-14 19:13:18 +02:00
|
|
|
bool isDelayedCommit() const { return staticInst->isDelayedCommit(); }
|
2007-06-21 04:46:45 +02:00
|
|
|
bool isLastMicroop() const { return staticInst->isLastMicroop(); }
|
|
|
|
bool isFirstMicroop() const { return staticInst->isFirstMicroop(); }
|
2007-04-14 19:13:18 +02:00
|
|
|
bool isMicroBranch() const { return staticInst->isMicroBranch(); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Temporarily sets this instruction as a serialize before instruction. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setSerializeBefore() { status.set(SerializeBefore); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Clears the serializeBefore part of this instruction. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void clearSerializeBefore() { status.reset(SerializeBefore); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Checks if this serializeBefore is only temporarily set. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isTempSerializeBefore() { return status[SerializeBefore]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Temporarily sets this instruction as a serialize after instruction. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setSerializeAfter() { status.set(SerializeAfter); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Clears the serializeAfter part of this instruction.*/
|
2006-06-14 19:12:41 +02:00
|
|
|
void clearSerializeAfter() { status.reset(SerializeAfter); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Checks if this serializeAfter is only temporarily set. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isTempSerializeAfter() { return status[SerializeAfter]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
2006-06-14 19:12:41 +02:00
|
|
|
/** Sets the serialization part of this instruction as handled. */
|
|
|
|
void setSerializeHandled() { status.set(SerializeHandled); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Checks if the serialization part of this instruction has been
|
|
|
|
* handled. This does not apply to the temporary serializing
|
|
|
|
* state; it only applies to this instruction's own permanent
|
|
|
|
* serializing state.
|
|
|
|
*/
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isSerializeHandled() { return status[SerializeHandled]; }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2005-02-26 00:00:49 +01:00
|
|
|
/** Returns the opclass of this instruction. */
|
|
|
|
OpClass opClass() const { return staticInst->opClass(); }
|
|
|
|
|
|
|
|
/** Returns the branch target address. */
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
TheISA::PCState branchTarget() const
|
|
|
|
{ return staticInst->branchTarget(pc); }
|
2005-02-26 00:00:49 +01:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/** Returns the number of source registers. */
|
2008-09-10 20:26:15 +02:00
|
|
|
int8_t numSrcRegs() const { return staticInst->numSrcRegs(); }
|
2005-05-19 07:28:25 +02:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/** Returns the number of destination registers. */
|
2004-08-20 20:54:07 +02:00
|
|
|
int8_t numDestRegs() const { return staticInst->numDestRegs(); }
|
|
|
|
|
|
|
|
// the following are used to track physical register usage
|
|
|
|
// for machines with separate int & FP reg files
|
|
|
|
int8_t numFPDestRegs() const { return staticInst->numFPDestRegs(); }
|
|
|
|
int8_t numIntDestRegs() const { return staticInst->numIntDestRegs(); }
|
|
|
|
|
|
|
|
/** Returns the logical register index of the i'th destination register. */
|
2006-04-23 00:26:48 +02:00
|
|
|
RegIndex destRegIdx(int i) const { return staticInst->destRegIdx(i); }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** Returns the logical register index of the i'th source register. */
|
2006-04-23 00:26:48 +02:00
|
|
|
RegIndex srcRegIdx(int i) const { return staticInst->srcRegIdx(i); }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2005-05-19 07:28:25 +02:00
|
|
|
/** Returns the result of an integer instruction. */
|
2005-02-26 00:00:49 +01:00
|
|
|
uint64_t readIntResult() { return instResult.integer; }
|
2005-05-19 07:28:25 +02:00
|
|
|
|
|
|
|
/** Returns the result of a floating point instruction. */
|
2006-08-11 23:42:59 +02:00
|
|
|
float readFloatResult() { return (float)instResult.dbl; }
|
2005-05-19 07:28:25 +02:00
|
|
|
|
|
|
|
/** Returns the result of a floating point (double) instruction. */
|
2005-02-26 00:00:49 +01:00
|
|
|
double readDoubleResult() { return instResult.dbl; }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-06-13 01:04:42 +02:00
|
|
|
/** Records an integer register being set to a value. */
|
2006-12-12 18:58:40 +01:00
|
|
|
void setIntRegOperand(const StaticInst *si, int idx, uint64_t val)
|
2006-05-11 20:12:34 +02:00
|
|
|
{
|
2006-10-23 20:00:07 +02:00
|
|
|
if (recordResult)
|
|
|
|
instResult.integer = val;
|
2006-05-11 20:12:34 +02:00
|
|
|
}
|
|
|
|
|
2006-06-13 01:04:42 +02:00
|
|
|
/** Records an fp register being set to a value. */
|
2006-12-12 18:58:40 +01:00
|
|
|
void setFloatRegOperand(const StaticInst *si, int idx, FloatReg val,
|
|
|
|
int width)
|
2006-06-08 22:58:50 +02:00
|
|
|
{
|
2006-10-23 20:00:07 +02:00
|
|
|
if (recordResult) {
|
|
|
|
if (width == 32)
|
|
|
|
instResult.dbl = (double)val;
|
|
|
|
else if (width == 64)
|
|
|
|
instResult.dbl = val;
|
|
|
|
else
|
|
|
|
panic("Unsupported width!");
|
|
|
|
}
|
2006-06-08 22:58:50 +02:00
|
|
|
}
|
|
|
|
|
2006-06-13 01:04:42 +02:00
|
|
|
/** Records an fp register being set to a value. */
|
2006-12-12 18:58:40 +01:00
|
|
|
void setFloatRegOperand(const StaticInst *si, int idx, FloatReg val)
|
2006-05-11 20:12:34 +02:00
|
|
|
{
|
2006-10-23 20:00:07 +02:00
|
|
|
if (recordResult)
|
|
|
|
instResult.dbl = (double)val;
|
2006-05-11 20:12:34 +02:00
|
|
|
}
|
|
|
|
|
2006-06-13 01:04:42 +02:00
|
|
|
/** Records an fp register being set to an integer value. */
|
2006-12-12 18:58:40 +01:00
|
|
|
void setFloatRegOperandBits(const StaticInst *si, int idx, uint64_t val,
|
|
|
|
int width)
|
2006-05-11 20:12:34 +02:00
|
|
|
{
|
2006-10-23 20:00:07 +02:00
|
|
|
if (recordResult)
|
|
|
|
instResult.integer = val;
|
2006-05-11 20:12:34 +02:00
|
|
|
}
|
|
|
|
|
2006-06-13 01:04:42 +02:00
|
|
|
/** Records an fp register being set to an integer value. */
|
2006-12-12 18:58:40 +01:00
|
|
|
void setFloatRegOperandBits(const StaticInst *si, int idx, uint64_t val)
|
2006-05-11 20:12:34 +02:00
|
|
|
{
|
2006-10-23 20:00:07 +02:00
|
|
|
if (recordResult)
|
|
|
|
instResult.integer = val;
|
2006-05-11 20:12:34 +02:00
|
|
|
}
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Records that one of the source registers is ready. */
|
2006-04-23 00:26:48 +02:00
|
|
|
void markSrcRegReady();
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-05-23 22:57:14 +02:00
|
|
|
/** Marks a specific register as ready. */
|
2006-04-23 00:26:48 +02:00
|
|
|
void markSrcRegReady(RegIndex src_idx);
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2005-05-19 07:28:25 +02:00
|
|
|
/** Returns if a source register is ready. */
|
2005-02-26 00:00:49 +01:00
|
|
|
bool isReadySrcRegIdx(int idx) const
|
|
|
|
{
|
|
|
|
return this->_readySrcRegIdx[idx];
|
|
|
|
}
|
|
|
|
|
2005-05-19 07:28:25 +02:00
|
|
|
/** Sets this instruction as completed. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setCompleted() { status.set(Completed); }
|
2005-02-26 00:00:49 +01:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/** Returns whether or not this instruction is completed. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isCompleted() const { return status[Completed]; }
|
2005-02-26 00:00:49 +01:00
|
|
|
|
2006-06-14 19:12:41 +02:00
|
|
|
/** Marks the result as ready. */
|
|
|
|
void setResultReady() { status.set(ResultReady); }
|
2006-05-11 20:12:34 +02:00
|
|
|
|
2006-06-14 19:12:41 +02:00
|
|
|
/** Returns whether or not the result is ready. */
|
|
|
|
bool isResultReady() const { return status[ResultReady]; }
|
2006-05-11 20:12:34 +02:00
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Sets this instruction as ready to issue. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setCanIssue() { status.set(CanIssue); }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** Returns whether or not this instruction is ready to issue. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool readyToIssue() const { return status[CanIssue]; }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2007-03-23 16:33:08 +01:00
|
|
|
/** Clears this instruction being able to issue. */
|
|
|
|
void clearCanIssue() { status.reset(CanIssue); }
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Sets this instruction as issued from the IQ. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setIssued() { status.set(Issued); }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** Returns whether or not this instruction has issued. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isIssued() const { return status[Issued]; }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2007-03-23 16:33:08 +01:00
|
|
|
/** Clears this instruction as being issued. */
|
|
|
|
void clearIssued() { status.reset(Issued); }
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Sets this instruction as executed. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setExecuted() { status.set(Executed); }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** Returns whether or not this instruction has executed. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isExecuted() const { return status[Executed]; }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** Sets this instruction as ready to commit. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setCanCommit() { status.set(CanCommit); }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
Update to make multiple instruction issue and different latencies work.
Also change to ref counted DynInst.
SConscript:
Add branch predictor, BTB, load store queue, and storesets.
arch/isa_parser.py:
Specify the template parameter for AlphaDynInst
base/traceflags.py:
Add load store queue, store set, and mem dependence unit to the
list of trace flags.
cpu/base_dyn_inst.cc:
Change formating, add in debug statement.
cpu/base_dyn_inst.hh:
Change DynInst to be RefCounted, add flag to clear whether or not this
instruction can commit. This is likely to be removed in the future.
cpu/beta_cpu/alpha_dyn_inst.cc:
AlphaDynInst has been changed to be templated, so now this CC file
is just used to force instantiations of AlphaDynInst.
cpu/beta_cpu/alpha_dyn_inst.hh:
Changed AlphaDynInst to be templated on Impl. Removed some unnecessary
functions.
cpu/beta_cpu/alpha_full_cpu.cc:
AlphaFullCPU has been changed to be templated, so this CC file is now
just used to force instantation of AlphaFullCPU.
cpu/beta_cpu/alpha_full_cpu.hh:
Change AlphaFullCPU to be templated on Impl.
cpu/beta_cpu/alpha_impl.hh:
Update it to reflect AlphaDynInst and AlphaFullCPU being templated
on Impl. Also removed time buffers from here, as they are really
a part of the CPU and are thus in the CPU policy now.
cpu/beta_cpu/alpha_params.hh:
Make AlphaSimpleParams inherit from the BaseFullCPU so that it doesn't
need to specifically declare any parameters that are already in the
BaseFullCPU.
cpu/beta_cpu/comm.hh:
Changed the structure of the time buffer communication structs. Now
they include the size of the packet of instructions it is sending.
Added some parameters to the backwards communication struct, mainly
for squashing.
cpu/beta_cpu/commit.hh:
Update typenames to reflect change in location of time buffer structs.
Update DynInst to DynInstPtr (it is refcounted now).
cpu/beta_cpu/commit_impl.hh:
Formatting changes mainly. Also sends back proper information
on branch mispredicts so that the bpred unit can update itself.
Updated behavior for non-speculative instructions (stores, any
other non-spec instructions): once they reach the head of the ROB,
the ROB signals back to the IQ that it can go ahead and issue the
non-speculative instruction. The instruction itself is updated so that
commit won't try to commit it again until it is done executing.
cpu/beta_cpu/cpu_policy.hh:
Added branch prediction unit, mem dependence prediction unit, load
store queue. Moved time buffer structs from AlphaSimpleImpl to here.
cpu/beta_cpu/decode.hh:
Changed typedefs to reflect change in location of time buffer structs
and also the change from DynInst to ref counted DynInstPtr.
cpu/beta_cpu/decode_impl.hh:
Continues to buffer instructions even while unblocking now. Changed
how it loops through groups of instructions so it can properly block
during the middle of a group of instructions.
cpu/beta_cpu/fetch.hh:
Changed typedefs to reflect change in location of time buffer structs
and the change to ref counted DynInsts. Also added in branch
brediction unit.
cpu/beta_cpu/fetch_impl.hh:
Add in branch prediction. Changed how fetch checks inputs and its
current state to make for easier logic.
cpu/beta_cpu/free_list.cc:
Changed int regs and float regs to logically use one flat namespace.
Future change will be moving them to a single scoreboard to conserve
space.
cpu/beta_cpu/free_list.hh:
Mostly debugging statements. Might be removed for performance in future.
cpu/beta_cpu/full_cpu.cc:
Added in some debugging statements. Updated BaseFullCPU to take
a params object.
cpu/beta_cpu/full_cpu.hh:
Added params class within BaseCPU that other param classes will be
able to inherit from. Updated typedefs to reflect change in location
of time buffer structs and ref counted DynInst.
cpu/beta_cpu/iew.hh:
Updated typedefs to reflect change in location of time buffer structs
and use of ref counted DynInsts.
cpu/beta_cpu/iew_impl.hh:
Added in load store queue, updated iew to be able to execute non-
speculative instructions, instead of having them execute in commit.
cpu/beta_cpu/inst_queue.hh:
Updated change to ref counted DynInsts. Changed inst queue to hold
non-speculative instructions as well, which are issued only when
commit signals backwards that a nonspeculative instruction is at
the head of the ROB.
cpu/beta_cpu/inst_queue_impl.hh:
Updated to allow for non-speculative instructions to be in the inst
queue. Also added some debug functions.
cpu/beta_cpu/regfile.hh:
Added debugging statements, changed formatting.
cpu/beta_cpu/rename.hh:
Updated typedefs, added some functions to clean up code.
cpu/beta_cpu/rename_impl.hh:
Moved some code into functions to make it easier to read.
cpu/beta_cpu/rename_map.cc:
Changed int and float reg behavior to use a single flat namespace. In
the future, the rename maps can be combined to a single rename map to
save space.
cpu/beta_cpu/rename_map.hh:
Added destructor.
cpu/beta_cpu/rob.hh:
Updated it with change from DynInst to ref counted DynInst.
cpu/beta_cpu/rob_impl.hh:
Formatting, updated to use ref counted DynInst.
cpu/static_inst.hh:
Updated forward declaration for AlphaDynInst now that it is templated.
--HG--
extra : convert_revision : 1045f240ee9b6a4bd368e1806aca029ebbdc6dd3
2004-09-23 20:06:03 +02:00
|
|
|
/** Clears this instruction as being ready to commit. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void clearCanCommit() { status.reset(CanCommit); }
|
Update to make multiple instruction issue and different latencies work.
Also change to ref counted DynInst.
SConscript:
Add branch predictor, BTB, load store queue, and storesets.
arch/isa_parser.py:
Specify the template parameter for AlphaDynInst
base/traceflags.py:
Add load store queue, store set, and mem dependence unit to the
list of trace flags.
cpu/base_dyn_inst.cc:
Change formating, add in debug statement.
cpu/base_dyn_inst.hh:
Change DynInst to be RefCounted, add flag to clear whether or not this
instruction can commit. This is likely to be removed in the future.
cpu/beta_cpu/alpha_dyn_inst.cc:
AlphaDynInst has been changed to be templated, so now this CC file
is just used to force instantiations of AlphaDynInst.
cpu/beta_cpu/alpha_dyn_inst.hh:
Changed AlphaDynInst to be templated on Impl. Removed some unnecessary
functions.
cpu/beta_cpu/alpha_full_cpu.cc:
AlphaFullCPU has been changed to be templated, so this CC file is now
just used to force instantation of AlphaFullCPU.
cpu/beta_cpu/alpha_full_cpu.hh:
Change AlphaFullCPU to be templated on Impl.
cpu/beta_cpu/alpha_impl.hh:
Update it to reflect AlphaDynInst and AlphaFullCPU being templated
on Impl. Also removed time buffers from here, as they are really
a part of the CPU and are thus in the CPU policy now.
cpu/beta_cpu/alpha_params.hh:
Make AlphaSimpleParams inherit from the BaseFullCPU so that it doesn't
need to specifically declare any parameters that are already in the
BaseFullCPU.
cpu/beta_cpu/comm.hh:
Changed the structure of the time buffer communication structs. Now
they include the size of the packet of instructions it is sending.
Added some parameters to the backwards communication struct, mainly
for squashing.
cpu/beta_cpu/commit.hh:
Update typenames to reflect change in location of time buffer structs.
Update DynInst to DynInstPtr (it is refcounted now).
cpu/beta_cpu/commit_impl.hh:
Formatting changes mainly. Also sends back proper information
on branch mispredicts so that the bpred unit can update itself.
Updated behavior for non-speculative instructions (stores, any
other non-spec instructions): once they reach the head of the ROB,
the ROB signals back to the IQ that it can go ahead and issue the
non-speculative instruction. The instruction itself is updated so that
commit won't try to commit it again until it is done executing.
cpu/beta_cpu/cpu_policy.hh:
Added branch prediction unit, mem dependence prediction unit, load
store queue. Moved time buffer structs from AlphaSimpleImpl to here.
cpu/beta_cpu/decode.hh:
Changed typedefs to reflect change in location of time buffer structs
and also the change from DynInst to ref counted DynInstPtr.
cpu/beta_cpu/decode_impl.hh:
Continues to buffer instructions even while unblocking now. Changed
how it loops through groups of instructions so it can properly block
during the middle of a group of instructions.
cpu/beta_cpu/fetch.hh:
Changed typedefs to reflect change in location of time buffer structs
and the change to ref counted DynInsts. Also added in branch
brediction unit.
cpu/beta_cpu/fetch_impl.hh:
Add in branch prediction. Changed how fetch checks inputs and its
current state to make for easier logic.
cpu/beta_cpu/free_list.cc:
Changed int regs and float regs to logically use one flat namespace.
Future change will be moving them to a single scoreboard to conserve
space.
cpu/beta_cpu/free_list.hh:
Mostly debugging statements. Might be removed for performance in future.
cpu/beta_cpu/full_cpu.cc:
Added in some debugging statements. Updated BaseFullCPU to take
a params object.
cpu/beta_cpu/full_cpu.hh:
Added params class within BaseCPU that other param classes will be
able to inherit from. Updated typedefs to reflect change in location
of time buffer structs and ref counted DynInst.
cpu/beta_cpu/iew.hh:
Updated typedefs to reflect change in location of time buffer structs
and use of ref counted DynInsts.
cpu/beta_cpu/iew_impl.hh:
Added in load store queue, updated iew to be able to execute non-
speculative instructions, instead of having them execute in commit.
cpu/beta_cpu/inst_queue.hh:
Updated change to ref counted DynInsts. Changed inst queue to hold
non-speculative instructions as well, which are issued only when
commit signals backwards that a nonspeculative instruction is at
the head of the ROB.
cpu/beta_cpu/inst_queue_impl.hh:
Updated to allow for non-speculative instructions to be in the inst
queue. Also added some debug functions.
cpu/beta_cpu/regfile.hh:
Added debugging statements, changed formatting.
cpu/beta_cpu/rename.hh:
Updated typedefs, added some functions to clean up code.
cpu/beta_cpu/rename_impl.hh:
Moved some code into functions to make it easier to read.
cpu/beta_cpu/rename_map.cc:
Changed int and float reg behavior to use a single flat namespace. In
the future, the rename maps can be combined to a single rename map to
save space.
cpu/beta_cpu/rename_map.hh:
Added destructor.
cpu/beta_cpu/rob.hh:
Updated it with change from DynInst to ref counted DynInst.
cpu/beta_cpu/rob_impl.hh:
Formatting, updated to use ref counted DynInst.
cpu/static_inst.hh:
Updated forward declaration for AlphaDynInst now that it is templated.
--HG--
extra : convert_revision : 1045f240ee9b6a4bd368e1806aca029ebbdc6dd3
2004-09-23 20:06:03 +02:00
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Returns whether or not this instruction is ready to commit. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool readyToCommit() const { return status[CanCommit]; }
|
|
|
|
|
|
|
|
void setAtCommit() { status.set(AtCommit); }
|
|
|
|
|
|
|
|
bool isAtCommit() { return status[AtCommit]; }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/** Sets this instruction as committed. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setCommitted() { status.set(Committed); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Returns whether or not this instruction is committed. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isCommitted() const { return status[Committed]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Sets this instruction as squashed. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setSquashed() { status.set(Squashed); }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** Returns whether or not this instruction is squashed. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isSquashed() const { return status[Squashed]; }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
//Instruction Queue Entry
|
|
|
|
//-----------------------
|
|
|
|
/** Sets this instruction as a entry the IQ. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setInIQ() { status.set(IqEntry); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Sets this instruction as a entry the IQ. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void clearInIQ() { status.reset(IqEntry); }
|
|
|
|
|
|
|
|
/** Returns whether or not this instruction has issued. */
|
|
|
|
bool isInIQ() const { return status[IqEntry]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
/** Sets this instruction as squashed in the IQ. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setSquashedInIQ() { status.set(SquashedInIQ); status.set(Squashed);}
|
2004-08-20 20:54:07 +02:00
|
|
|
|
|
|
|
/** Returns whether or not this instruction is squashed in the IQ. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isSquashedInIQ() const { return status[SquashedInIQ]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
//Load / Store Queue Functions
|
|
|
|
//-----------------------
|
|
|
|
/** Sets this instruction as a entry the LSQ. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setInLSQ() { status.set(LsqEntry); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Sets this instruction as a entry the LSQ. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void removeInLSQ() { status.reset(LsqEntry); }
|
|
|
|
|
|
|
|
/** Returns whether or not this instruction is in the LSQ. */
|
|
|
|
bool isInLSQ() const { return status[LsqEntry]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Sets this instruction as squashed in the LSQ. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setSquashedInLSQ() { status.set(SquashedInLSQ);}
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Returns whether or not this instruction is squashed in the LSQ. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isSquashedInLSQ() const { return status[SquashedInLSQ]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
//Reorder Buffer Functions
|
|
|
|
//-----------------------
|
|
|
|
/** Sets this instruction as a entry the ROB. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setInROB() { status.set(RobEntry); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Sets this instruction as a entry the ROB. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void clearInROB() { status.reset(RobEntry); }
|
|
|
|
|
|
|
|
/** Returns whether or not this instruction is in the ROB. */
|
|
|
|
bool isInROB() const { return status[RobEntry]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Sets this instruction as squashed in the ROB. */
|
2006-06-14 19:12:41 +02:00
|
|
|
void setSquashedInROB() { status.set(SquashedInROB); }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Returns whether or not this instruction is squashed in the ROB. */
|
2006-06-14 19:12:41 +02:00
|
|
|
bool isSquashedInROB() const { return status[SquashedInROB]; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
/** Read the PC state of this instruction. */
|
|
|
|
const TheISA::PCState pcState() const { return pc; }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
/** Set the PC state of this instruction. */
|
|
|
|
const void pcState(const TheISA::PCState &val) { pc = val; }
|
2007-04-14 19:13:18 +02:00
|
|
|
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
/** Read the PC of this instruction. */
|
|
|
|
const Addr instAddr() const { return pc.instAddr(); }
|
2004-08-20 20:54:07 +02:00
|
|
|
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
/** Read the PC of the next instruction. */
|
|
|
|
const Addr nextInstAddr() const { return pc.nextInstAddr(); }
|
2006-07-23 19:39:42 +02:00
|
|
|
|
ISA,CPU,etc: Create an ISA defined PC type that abstracts out ISA behaviors.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
2010-10-31 08:07:20 +01:00
|
|
|
/**Read the micro PC of this instruction. */
|
|
|
|
const Addr microPC() const { return pc.microPC(); }
|
2007-04-14 19:13:18 +02:00
|
|
|
|
2010-08-23 18:18:40 +02:00
|
|
|
bool readPredicate()
|
|
|
|
{
|
|
|
|
return predicate;
|
|
|
|
}
|
|
|
|
|
|
|
|
void setPredicate(bool val)
|
|
|
|
{
|
|
|
|
predicate = val;
|
2010-08-23 18:18:41 +02:00
|
|
|
|
|
|
|
if (traceData) {
|
|
|
|
traceData->setPredicate(val);
|
|
|
|
}
|
2010-08-23 18:18:40 +02:00
|
|
|
}
|
|
|
|
|
2006-06-13 01:04:42 +02:00
|
|
|
/** Sets the ASID. */
|
2006-04-23 00:26:48 +02:00
|
|
|
void setASID(short addr_space_id) { asid = addr_space_id; }
|
|
|
|
|
2006-06-13 01:04:42 +02:00
|
|
|
/** Sets the thread id. */
|
2009-05-26 18:23:13 +02:00
|
|
|
void setTid(ThreadID tid) { threadNumber = tid; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
2006-06-14 19:12:41 +02:00
|
|
|
/** Sets the pointer to the thread state. */
|
2006-06-13 01:04:42 +02:00
|
|
|
void setThreadState(ImplState *state) { thread = state; }
|
2006-04-23 00:26:48 +02:00
|
|
|
|
2006-06-14 19:12:41 +02:00
|
|
|
/** Returns the thread context. */
|
2006-06-06 23:32:21 +02:00
|
|
|
ThreadContext *tcBase() { return thread->getTC(); }
|
2005-02-26 00:00:49 +01:00
|
|
|
|
|
|
|
private:
|
2005-05-19 07:28:25 +02:00
|
|
|
/** Instruction effective address.
|
|
|
|
* @todo: Consider if this is necessary or not.
|
|
|
|
*/
|
2005-02-26 00:00:49 +01:00
|
|
|
Addr instEffAddr;
|
2006-04-23 00:26:48 +02:00
|
|
|
|
2005-05-19 07:28:25 +02:00
|
|
|
/** Whether or not the effective address calculation is completed.
|
|
|
|
* @todo: Consider if this is necessary or not.
|
|
|
|
*/
|
2005-02-26 00:00:49 +01:00
|
|
|
bool eaCalcDone;
|
|
|
|
|
2007-03-23 16:33:08 +01:00
|
|
|
/** Is this instruction's memory access uncacheable. */
|
|
|
|
bool isUncacheable;
|
|
|
|
|
|
|
|
/** Has this instruction generated a memory request. */
|
|
|
|
bool reqMade;
|
|
|
|
|
2005-02-26 00:00:49 +01:00
|
|
|
public:
|
2005-05-19 07:28:25 +02:00
|
|
|
/** Sets the effective address. */
|
2005-02-26 00:00:49 +01:00
|
|
|
void setEA(Addr &ea) { instEffAddr = ea; eaCalcDone = true; }
|
2005-05-19 07:28:25 +02:00
|
|
|
|
|
|
|
/** Returns the effective address. */
|
2006-06-03 00:15:20 +02:00
|
|
|
const Addr &getEA() const { return instEffAddr; }
|
2005-05-19 07:28:25 +02:00
|
|
|
|
|
|
|
/** Returns whether or not the eff. addr. calculation has been completed. */
|
2005-02-26 00:00:49 +01:00
|
|
|
bool doneEACalc() { return eaCalcDone; }
|
2005-05-19 07:28:25 +02:00
|
|
|
|
|
|
|
/** Returns whether or not the eff. addr. source registers are ready. */
|
2005-02-26 00:00:49 +01:00
|
|
|
bool eaSrcsReady();
|
2005-05-03 16:56:47 +02:00
|
|
|
|
2006-04-23 00:26:48 +02:00
|
|
|
/** Whether or not the memory operation is done. */
|
|
|
|
bool memOpDone;
|
|
|
|
|
2007-03-23 16:33:08 +01:00
|
|
|
/** Is this instruction's memory access uncacheable. */
|
|
|
|
bool uncacheable() { return isUncacheable; }
|
|
|
|
|
|
|
|
/** Has this instruction generated a memory request. */
|
|
|
|
bool hasRequest() { return reqMade; }
|
|
|
|
|
2005-05-03 16:56:47 +02:00
|
|
|
public:
|
2005-05-19 07:28:25 +02:00
|
|
|
/** Load queue index. */
|
2005-05-03 16:56:47 +02:00
|
|
|
int16_t lqIdx;
|
2005-05-19 07:28:25 +02:00
|
|
|
|
|
|
|
/** Store queue index. */
|
2005-05-03 16:56:47 +02:00
|
|
|
int16_t sqIdx;
|
2006-04-23 00:26:48 +02:00
|
|
|
|
|
|
|
/** Iterator pointing to this BaseDynInst in the list of all insts. */
|
|
|
|
ListIt instListIt;
|
|
|
|
|
|
|
|
/** Returns iterator to this instruction in the list of all insts. */
|
|
|
|
ListIt &getInstListIt() { return instListIt; }
|
|
|
|
|
|
|
|
/** Sets iterator for this instruction in the list of all insts. */
|
|
|
|
void setInstListIt(ListIt _instListIt) { instListIt = _instListIt; }
|
2006-10-23 20:00:07 +02:00
|
|
|
|
|
|
|
public:
|
|
|
|
/** Returns the number of consecutive store conditional failures. */
|
|
|
|
unsigned readStCondFailures()
|
|
|
|
{ return thread->storeCondFailures; }
|
|
|
|
|
|
|
|
/** Sets the number of consecutive store conditional failures. */
|
|
|
|
void setStCondFailures(unsigned sc_failures)
|
|
|
|
{ thread->storeCondFailures = sc_failures; }
|
2004-08-20 20:54:07 +02:00
|
|
|
};
|
|
|
|
|
|
|
|
template<class Impl>
|
2010-08-13 15:16:02 +02:00
|
|
|
Fault
|
|
|
|
BaseDynInst<Impl>::readBytes(Addr addr, uint8_t *data,
|
|
|
|
unsigned size, unsigned flags)
|
2004-08-20 20:54:07 +02:00
|
|
|
{
|
2007-03-23 16:33:08 +01:00
|
|
|
reqMade = true;
|
2011-02-12 01:29:35 +01:00
|
|
|
Request *req = NULL;
|
2010-02-12 20:53:20 +01:00
|
|
|
Request *sreqLow = NULL;
|
|
|
|
Request *sreqHigh = NULL;
|
|
|
|
|
2011-02-12 01:29:35 +01:00
|
|
|
if (reqMade && translationStarted) {
|
|
|
|
req = savedReq;
|
|
|
|
sreqLow = savedSreqLow;
|
|
|
|
sreqHigh = savedSreqHigh;
|
2005-05-19 07:28:25 +02:00
|
|
|
} else {
|
2011-02-12 01:29:35 +01:00
|
|
|
req = new Request(asid, addr, size, flags, this->pc.instAddr(),
|
|
|
|
thread->contextId(), threadNumber);
|
|
|
|
|
|
|
|
// Only split the request if the ISA supports unaligned accesses.
|
|
|
|
if (TheISA::HasUnalignedMemAcc) {
|
|
|
|
splitRequest(req, sreqLow, sreqHigh);
|
|
|
|
}
|
|
|
|
initiateTranslation(req, sreqLow, sreqHigh, NULL, BaseTLB::Read);
|
2004-08-20 20:54:07 +02:00
|
|
|
}
|
|
|
|
|
2011-02-12 01:29:35 +01:00
|
|
|
if (translationCompleted) {
|
|
|
|
if (fault == NoFault) {
|
|
|
|
effAddr = req->getVaddr();
|
|
|
|
effAddrValid = true;
|
|
|
|
fault = cpu->read(req, sreqLow, sreqHigh, data, lqIdx);
|
|
|
|
} else {
|
|
|
|
// Commit will have to clean up whatever happened. Set this
|
|
|
|
// instruction as executed.
|
|
|
|
this->setExecuted();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (fault != NoFault) {
|
|
|
|
// Return a fixed value to keep simulation deterministic even
|
|
|
|
// along misspeculated paths.
|
|
|
|
if (data)
|
|
|
|
bzero(data, size);
|
|
|
|
}
|
2010-08-23 18:18:39 +02:00
|
|
|
}
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
if (traceData) {
|
|
|
|
traceData->setAddr(addr);
|
|
|
|
}
|
|
|
|
|
|
|
|
return fault;
|
|
|
|
}
|
|
|
|
|
|
|
|
template<class Impl>
|
|
|
|
template<class T>
|
2006-02-22 02:10:40 +01:00
|
|
|
inline Fault
|
2010-08-13 15:16:02 +02:00
|
|
|
BaseDynInst<Impl>::read(Addr addr, T &data, unsigned flags)
|
2004-08-20 20:54:07 +02:00
|
|
|
{
|
2010-08-13 15:16:02 +02:00
|
|
|
Fault fault = readBytes(addr, (uint8_t *)&data, sizeof(T), flags);
|
|
|
|
|
|
|
|
data = TheISA::gtoh(data);
|
|
|
|
|
2004-08-20 20:54:07 +02:00
|
|
|
if (traceData) {
|
|
|
|
traceData->setData(data);
|
|
|
|
}
|
|
|
|
|
2010-08-13 15:16:02 +02:00
|
|
|
return fault;
|
|
|
|
}
|
|
|
|
|
|
|
|
template<class Impl>
|
|
|
|
Fault
|
|
|
|
BaseDynInst<Impl>::writeBytes(uint8_t *data, unsigned size,
|
|
|
|
Addr addr, unsigned flags, uint64_t *res)
|
|
|
|
{
|
|
|
|
if (traceData) {
|
|
|
|
traceData->setAddr(addr);
|
|
|
|
}
|
|
|
|
|
2007-03-23 16:33:08 +01:00
|
|
|
reqMade = true;
|
2011-02-12 01:29:35 +01:00
|
|
|
Request *req = NULL;
|
2010-02-12 20:53:20 +01:00
|
|
|
Request *sreqLow = NULL;
|
|
|
|
Request *sreqHigh = NULL;
|
|
|
|
|
2011-02-12 01:29:35 +01:00
|
|
|
if (reqMade && translationStarted) {
|
|
|
|
req = savedReq;
|
|
|
|
sreqLow = savedSreqLow;
|
|
|
|
sreqHigh = savedSreqHigh;
|
|
|
|
} else {
|
|
|
|
req = new Request(asid, addr, size, flags, this->pc.instAddr(),
|
|
|
|
thread->contextId(), threadNumber);
|
|
|
|
|
|
|
|
// Only split the request if the ISA supports unaligned accesses.
|
|
|
|
if (TheISA::HasUnalignedMemAcc) {
|
|
|
|
splitRequest(req, sreqLow, sreqHigh);
|
|
|
|
}
|
|
|
|
initiateTranslation(req, sreqLow, sreqHigh, res, BaseTLB::Write);
|
2010-02-12 20:53:20 +01:00
|
|
|
}
|
2007-03-23 16:33:08 +01:00
|
|
|
|
2011-02-12 01:29:35 +01:00
|
|
|
if (fault == NoFault && translationCompleted) {
|
2006-06-06 00:14:39 +02:00
|
|
|
effAddr = req->getVaddr();
|
2007-03-23 16:33:08 +01:00
|
|
|
effAddrValid = true;
|
2010-02-20 21:11:58 +01:00
|
|
|
fault = cpu->write(req, sreqLow, sreqHigh, data, sqIdx);
|
2010-02-12 20:53:19 +01:00
|
|
|
}
|
2007-04-08 03:42:42 +02:00
|
|
|
|
2010-02-12 20:53:19 +01:00
|
|
|
return fault;
|
|
|
|
}
|
|
|
|
|
2010-08-13 15:16:02 +02:00
|
|
|
template<class Impl>
|
|
|
|
template<class T>
|
|
|
|
inline Fault
|
|
|
|
BaseDynInst<Impl>::write(T data, Addr addr, unsigned flags, uint64_t *res)
|
|
|
|
{
|
|
|
|
if (traceData) {
|
|
|
|
traceData->setData(data);
|
|
|
|
}
|
|
|
|
data = TheISA::htog(data);
|
|
|
|
return writeBytes((uint8_t *)&data, sizeof(T), addr, flags, res);
|
|
|
|
}
|
|
|
|
|
2010-02-12 20:53:19 +01:00
|
|
|
template<class Impl>
|
|
|
|
inline void
|
2010-02-12 20:53:20 +01:00
|
|
|
BaseDynInst<Impl>::splitRequest(RequestPtr req, RequestPtr &sreqLow,
|
|
|
|
RequestPtr &sreqHigh)
|
|
|
|
{
|
|
|
|
// Check to see if the request crosses the next level block boundary.
|
|
|
|
unsigned block_size = cpu->getDcachePort()->peerBlockSize();
|
|
|
|
Addr addr = req->getVaddr();
|
|
|
|
Addr split_addr = roundDown(addr + req->getSize() - 1, block_size);
|
|
|
|
assert(split_addr <= addr || split_addr - addr < block_size);
|
|
|
|
|
|
|
|
// Spans two blocks.
|
|
|
|
if (split_addr > addr) {
|
|
|
|
req->splitOnVaddr(split_addr, sreqLow, sreqHigh);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template<class Impl>
|
|
|
|
inline void
|
|
|
|
BaseDynInst<Impl>::initiateTranslation(RequestPtr req, RequestPtr sreqLow,
|
|
|
|
RequestPtr sreqHigh, uint64_t *res,
|
2010-02-12 20:53:19 +01:00
|
|
|
BaseTLB::Mode mode)
|
|
|
|
{
|
2011-02-12 01:29:35 +01:00
|
|
|
translationStarted = true;
|
|
|
|
|
2010-02-12 20:53:20 +01:00
|
|
|
if (!TheISA::HasUnalignedMemAcc || sreqLow == NULL) {
|
|
|
|
WholeTranslationState *state =
|
|
|
|
new WholeTranslationState(req, NULL, res, mode);
|
|
|
|
|
|
|
|
// One translation if the request isn't split.
|
|
|
|
DataTranslation<BaseDynInst<Impl> > *trans =
|
|
|
|
new DataTranslation<BaseDynInst<Impl> >(this, state);
|
|
|
|
cpu->dtb->translateTiming(req, thread->getTC(), trans, mode);
|
2011-02-12 01:29:35 +01:00
|
|
|
if (!translationCompleted) {
|
|
|
|
// Save memory requests.
|
|
|
|
savedReq = state->mainReq;
|
|
|
|
savedSreqLow = state->sreqLow;
|
|
|
|
savedSreqHigh = state->sreqHigh;
|
|
|
|
}
|
2010-02-12 20:53:20 +01:00
|
|
|
} else {
|
|
|
|
WholeTranslationState *state =
|
|
|
|
new WholeTranslationState(req, sreqLow, sreqHigh, NULL, res, mode);
|
|
|
|
|
|
|
|
// Two translations when the request is split.
|
|
|
|
DataTranslation<BaseDynInst<Impl> > *stransLow =
|
|
|
|
new DataTranslation<BaseDynInst<Impl> >(this, state, 0);
|
|
|
|
DataTranslation<BaseDynInst<Impl> > *stransHigh =
|
|
|
|
new DataTranslation<BaseDynInst<Impl> >(this, state, 1);
|
|
|
|
|
|
|
|
cpu->dtb->translateTiming(sreqLow, thread->getTC(), stransLow, mode);
|
|
|
|
cpu->dtb->translateTiming(sreqHigh, thread->getTC(), stransHigh, mode);
|
2011-02-12 01:29:35 +01:00
|
|
|
if (!translationCompleted) {
|
|
|
|
// Save memory requests.
|
|
|
|
savedReq = state->mainReq;
|
|
|
|
savedSreqLow = state->sreqLow;
|
|
|
|
savedSreqHigh = state->sreqHigh;
|
|
|
|
}
|
2010-02-12 20:53:20 +01:00
|
|
|
}
|
2010-02-12 20:53:19 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
template<class Impl>
|
|
|
|
inline void
|
|
|
|
BaseDynInst<Impl>::finishTranslation(WholeTranslationState *state)
|
|
|
|
{
|
|
|
|
fault = state->getFault();
|
|
|
|
|
|
|
|
if (state->isUncacheable())
|
|
|
|
isUncacheable = true;
|
|
|
|
|
|
|
|
if (fault == NoFault) {
|
|
|
|
physEffAddr = state->getPaddr();
|
|
|
|
memReqFlags = state->getFlags();
|
|
|
|
|
|
|
|
if (state->mainReq->isCondSwap()) {
|
|
|
|
assert(state->res);
|
|
|
|
state->mainReq->setExtraData(*state->res);
|
2006-04-23 00:26:48 +02:00
|
|
|
}
|
2010-02-12 20:53:19 +01:00
|
|
|
|
2007-03-23 16:33:08 +01:00
|
|
|
} else {
|
2010-02-12 20:53:19 +01:00
|
|
|
state->deleteReqs();
|
2004-08-20 20:54:07 +02:00
|
|
|
}
|
2010-02-12 20:53:19 +01:00
|
|
|
delete state;
|
2011-02-12 01:29:35 +01:00
|
|
|
|
|
|
|
translationCompleted = true;
|
2004-08-20 20:54:07 +02:00
|
|
|
}
|
|
|
|
|
2005-02-26 00:00:49 +01:00
|
|
|
#endif // __CPU_BASE_DYN_INST_HH__
|