2006-06-09 09:57:25 +02:00
|
|
|
// -*- mode:c++ -*-
|
2006-02-16 08:51:04 +01:00
|
|
|
|
2007-11-15 20:21:01 +01:00
|
|
|
// Copyright (c) 2007 MIPS Technologies, Inc.
|
|
|
|
// 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.
|
|
|
|
//
|
|
|
|
// Authors: Korey Sewell
|
|
|
|
// Brett Miller
|
|
|
|
// Jaidev Patwardhan
|
2006-06-10 00:19:08 +02:00
|
|
|
|
2006-02-08 00:36:08 +01:00
|
|
|
////////////////////////////////////////////////////////////////////
|
|
|
|
//
|
|
|
|
// The actual MIPS32 ISA decoder
|
|
|
|
// -----------------------------
|
|
|
|
// The following instructions are specified in the MIPS32 ISA
|
|
|
|
// Specification. Decoding closely follows the style specified
|
2006-06-09 09:57:25 +02:00
|
|
|
// in the MIPS32 ISA specification document starting with Table
|
2007-06-23 01:03:42 +02:00
|
|
|
// A-2 (document available @ http://www.mips.com)
|
2006-02-08 00:36:08 +01:00
|
|
|
//
|
2006-02-15 03:26:01 +01:00
|
|
|
decode OPCODE_HI default Unknown::unknown() {
|
2006-06-09 09:57:25 +02:00
|
|
|
//Table A-2
|
2006-02-18 09:12:04 +01:00
|
|
|
0x0: decode OPCODE_LO {
|
2006-02-08 00:36:08 +01:00
|
|
|
0x0: decode FUNCTION_HI {
|
|
|
|
0x0: decode FUNCTION_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x1: decode MOVCI {
|
|
|
|
format BasicOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0: movf({{
|
|
|
|
Rd = (getCondCode(FCSR, CC) == 0) ? Rd : Rs;
|
|
|
|
}});
|
|
|
|
1: movt({{
|
|
|
|
Rd = (getCondCode(FCSR, CC) == 1) ? Rd : Rs;
|
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
format BasicOp {
|
2006-06-09 09:57:25 +02:00
|
|
|
//Table A-3 Note: "Specific encodings of the rd, rs, and
|
|
|
|
//rt fields are used to distinguish SLL, SSNOP, and EHB
|
|
|
|
//functions
|
2006-03-15 00:28:51 +01:00
|
|
|
0x0: decode RS {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x0: decode RT_RD {
|
2007-06-23 01:03:42 +02:00
|
|
|
0x0: decode SA default Nop::nop() {
|
2007-11-13 22:58:16 +01:00
|
|
|
0x1: ssnop({{;}});
|
|
|
|
0x3: ehb({{;}});
|
2006-06-09 09:57:25 +02:00
|
|
|
}
|
|
|
|
default: sll({{ Rd = Rt.uw << SA; }});
|
2006-03-15 00:28:51 +01:00
|
|
|
}
|
2006-03-08 22:53:44 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-04-10 18:23:17 +02:00
|
|
|
0x2: decode RS_SRL {
|
|
|
|
0x0:decode SRL {
|
|
|
|
0: srl({{ Rd = Rt.uw >> SA; }});
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Hardcoded assuming 32-bit ISA,
|
|
|
|
//probably need parameter here
|
|
|
|
1: rotr({{
|
|
|
|
Rd = (Rt.uw << (32 - SA)) | (Rt.uw >> SA);
|
|
|
|
}});
|
2006-04-10 18:23:17 +02:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-04-10 18:23:17 +02:00
|
|
|
0x3: decode RS {
|
|
|
|
0x0: sra({{
|
|
|
|
uint32_t temp = Rt >> SA;
|
|
|
|
if ( (Rt & 0x80000000) > 0 ) {
|
|
|
|
uint32_t mask = 0x80000000;
|
|
|
|
for(int i=0; i < SA; i++) {
|
|
|
|
temp |= mask;
|
|
|
|
mask = mask >> 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
Rd = temp;
|
|
|
|
}});
|
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
0x4: sllv({{ Rd = Rt.uw << Rs<4:0>; }});
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
0x6: decode SRLV {
|
|
|
|
0: srlv({{ Rd = Rt.uw >> Rs<4:0>; }});
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Hardcoded assuming 32-bit ISA,
|
|
|
|
//probably need parameter here
|
|
|
|
1: rotrv({{
|
|
|
|
Rd = (Rt.uw << (32 - Rs<4:0>)) |
|
|
|
|
(Rt.uw >> Rs<4:0>);
|
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-04-10 18:23:17 +02:00
|
|
|
0x7: srav({{
|
|
|
|
int shift_amt = Rs<4:0>;
|
|
|
|
|
|
|
|
uint32_t temp = Rt >> shift_amt;
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
if ((Rt & 0x80000000) > 0) {
|
|
|
|
uint32_t mask = 0x80000000;
|
|
|
|
for (int i = 0; i < shift_amt; i++) {
|
|
|
|
temp |= mask;
|
|
|
|
mask = mask >> 1;
|
2006-04-10 18:23:17 +02:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
}
|
2006-04-10 18:23:17 +02:00
|
|
|
Rd = temp;
|
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
0x1: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
//Table A-3 Note: "Specific encodings of the hint field are
|
|
|
|
//used to distinguish JR from JR.HB and JALR from JALR.HB"
|
2006-02-20 07:49:16 +01:00
|
|
|
format Jump {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x0: decode HINT {
|
2009-07-21 05:14:15 +02:00
|
|
|
0x1: jr_hb({{
|
|
|
|
Config1Reg config1 = Config1;
|
|
|
|
if (config1.ca == 0) {
|
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.nnpc(Rs);
|
2009-07-21 05:14:15 +02:00
|
|
|
} else {
|
|
|
|
panic("MIPS16e not supported\n");
|
|
|
|
}
|
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
|
|
|
PCS = pc;
|
2009-07-21 05:14:15 +02:00
|
|
|
}}, IsReturn, ClearHazards);
|
|
|
|
default: jr({{
|
|
|
|
Config1Reg config1 = Config1;
|
|
|
|
if (config1.ca == 0) {
|
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.nnpc(Rs);
|
2009-07-21 05:14:15 +02:00
|
|
|
} else {
|
|
|
|
panic("MIPS16e not supported\n");
|
|
|
|
}
|
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
|
|
|
PCS = pc;
|
2009-07-21 05:14:15 +02:00
|
|
|
}}, IsReturn);
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
0x1: decode HINT {
|
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
|
|
|
0x1: jalr_hb({{
|
|
|
|
Rd = pc.nnpc();
|
|
|
|
pc.nnpc(Rs);
|
|
|
|
PCS = pc;
|
|
|
|
}}, IsCall, ClearHazards);
|
|
|
|
default: jalr({{
|
|
|
|
Rd = pc.nnpc();
|
|
|
|
pc.nnpc(Rs);
|
|
|
|
PCS = pc;
|
|
|
|
}}, IsCall);
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
format BasicOp {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x2: movz({{ Rd = (Rt == 0) ? Rs : Rd; }});
|
|
|
|
0x3: movn({{ Rd = (Rt != 0) ? Rs : Rd; }});
|
2007-11-13 22:58:16 +01:00
|
|
|
#if FULL_SYSTEM
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: syscall({{ fault = new SystemCallFault(); }});
|
2007-11-13 22:58:16 +01:00
|
|
|
#else
|
2006-07-27 00:47:06 +02:00
|
|
|
0x4: syscall({{ xc->syscall(R2); }},
|
2009-04-18 16:42:29 +02:00
|
|
|
IsSerializeAfter, IsNonSpeculative);
|
2007-11-13 22:58:16 +01:00
|
|
|
#endif
|
2006-06-09 09:57:25 +02:00
|
|
|
0x7: sync({{ ; }}, IsMemBarrier);
|
2007-11-13 22:58:16 +01:00
|
|
|
0x5: break({{fault = new BreakpointFault();}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
0x2: decode FUNCTION_LO {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: HiLoRsSelOp::mfhi({{ Rd = HI_RS_SEL; }},
|
|
|
|
IntMultOp, IsIprAccess);
|
2007-06-23 01:03:42 +02:00
|
|
|
0x1: HiLoRdSelOp::mthi({{ HI_RD_SEL = Rs; }});
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: HiLoRsSelOp::mflo({{ Rd = LO_RS_SEL; }},
|
|
|
|
IntMultOp, IsIprAccess);
|
2007-06-23 01:03:42 +02:00
|
|
|
0x3: HiLoRdSelOp::mtlo({{ LO_RD_SEL = Rs; }});
|
2006-02-15 03:26:01 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
|
|
|
0x3: decode FUNCTION_LO {
|
2007-06-23 01:03:42 +02:00
|
|
|
format HiLoRdSelValOp {
|
2007-11-13 22:58:16 +01:00
|
|
|
0x0: mult({{ val = Rs.sd * Rt.sd; }}, IntMultOp);
|
|
|
|
0x1: multu({{ val = Rs.ud * Rt.ud; }}, IntMultOp);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
format HiLoOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: div({{
|
|
|
|
if (Rt.sd != 0) {
|
|
|
|
HI0 = Rs.sd % Rt.sd;
|
|
|
|
LO0 = Rs.sd / Rt.sd;
|
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
}}, IntDivOp);
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
0x3: divu({{
|
|
|
|
if (Rt.ud != 0) {
|
|
|
|
HI0 = Rs.ud % Rt.ud;
|
|
|
|
LO0 = Rs.ud / Rt.ud;
|
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
}}, IntDivOp);
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-04-10 18:23:17 +02:00
|
|
|
0x4: decode HINT {
|
|
|
|
0x0: decode FUNCTION_LO {
|
|
|
|
format IntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: add({{
|
|
|
|
/* More complicated since an ADD can cause
|
|
|
|
an arithmetic overflow exception */
|
|
|
|
int64_t Src1 = Rs.sw;
|
|
|
|
int64_t Src2 = Rt.sw;
|
|
|
|
int64_t temp_result;
|
|
|
|
#if FULL_SYSTEM
|
|
|
|
if (((Src1 >> 31) & 1) == 1)
|
|
|
|
Src1 |= 0x100000000LL;
|
2007-11-13 22:58:16 +01:00
|
|
|
#endif
|
2009-07-22 10:51:10 +02:00
|
|
|
temp_result = Src1 + Src2;
|
|
|
|
#if FULL_SYSTEM
|
|
|
|
if (bits(temp_result, 31) ==
|
|
|
|
bits(temp_result, 32)) {
|
2007-11-13 22:58:16 +01:00
|
|
|
#endif
|
2009-07-22 10:51:10 +02:00
|
|
|
Rd.sw = temp_result;
|
|
|
|
#if FULL_SYSTEM
|
|
|
|
} else {
|
|
|
|
fault = new ArithmeticFault();
|
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
#endif
|
2009-07-22 10:51:10 +02:00
|
|
|
}});
|
2006-04-10 18:23:17 +02:00
|
|
|
0x1: addu({{ Rd.sw = Rs.sw + Rt.sw;}});
|
2007-11-13 22:58:16 +01:00
|
|
|
0x2: sub({{
|
2009-07-22 10:51:10 +02:00
|
|
|
/* More complicated since an SUB can cause
|
|
|
|
an arithmetic overflow exception */
|
|
|
|
int64_t Src1 = Rs.sw;
|
|
|
|
int64_t Src2 = Rt.sw;
|
|
|
|
int64_t temp_result = Src1 - Src2;
|
2007-11-13 22:58:16 +01:00
|
|
|
#if FULL_SYSTEM
|
2009-07-22 10:51:10 +02:00
|
|
|
if (bits(temp_result, 31) ==
|
|
|
|
bits(temp_result, 32)) {
|
2007-11-13 22:58:16 +01:00
|
|
|
#endif
|
2009-07-22 10:51:10 +02:00
|
|
|
Rd.sw = temp_result;
|
2007-11-13 22:58:16 +01:00
|
|
|
#if FULL_SYSTEM
|
2009-07-22 10:51:10 +02:00
|
|
|
} else {
|
|
|
|
fault = new ArithmeticFault();
|
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
#endif
|
2009-07-22 10:51:10 +02:00
|
|
|
}});
|
|
|
|
0x3: subu({{ Rd.sw = Rs.sw - Rt.sw; }});
|
|
|
|
0x4: and({{ Rd = Rs & Rt; }});
|
|
|
|
0x5: or({{ Rd = Rs | Rt; }});
|
|
|
|
0x6: xor({{ Rd = Rs ^ Rt; }});
|
|
|
|
0x7: nor({{ Rd = ~(Rs | Rt); }});
|
2006-04-10 18:23:17 +02:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
|
2006-04-10 18:23:17 +02:00
|
|
|
0x5: decode HINT {
|
|
|
|
0x0: decode FUNCTION_LO {
|
|
|
|
format IntOp{
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: slt({{ Rd.sw = (Rs.sw < Rt.sw) ? 1 : 0 }});
|
|
|
|
0x3: sltu({{ Rd.uw = (Rs.uw < Rt.uw) ? 1 : 0 }});
|
2006-04-10 18:23:17 +02:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
|
|
|
0x6: decode FUNCTION_LO {
|
2006-03-08 08:05:38 +01:00
|
|
|
format Trap {
|
|
|
|
0x0: tge({{ cond = (Rs.sw >= Rt.sw); }});
|
2006-02-18 09:12:04 +01:00
|
|
|
0x1: tgeu({{ cond = (Rs.uw >= Rt.uw); }});
|
|
|
|
0x2: tlt({{ cond = (Rs.sw < Rt.sw); }});
|
2007-11-13 22:58:16 +01:00
|
|
|
0x3: tltu({{ cond = (Rs.uw < Rt.uw); }});
|
2006-02-18 09:12:04 +01:00
|
|
|
0x4: teq({{ cond = (Rs.sw == Rt.sw); }});
|
|
|
|
0x6: tne({{ cond = (Rs.sw != Rt.sw); }});
|
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
0x1: decode REGIMM_HI {
|
|
|
|
0x0: decode REGIMM_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
format Branch {
|
|
|
|
0x0: bltz({{ cond = (Rs.sw < 0); }});
|
|
|
|
0x1: bgez({{ cond = (Rs.sw >= 0); }});
|
2006-06-09 09:57:25 +02:00
|
|
|
0x2: bltzl({{ cond = (Rs.sw < 0); }}, Likely);
|
|
|
|
0x3: bgezl({{ cond = (Rs.sw >= 0); }}, Likely);
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
0x1: decode REGIMM_LO {
|
2007-11-13 22:58:16 +01:00
|
|
|
format TrapImm {
|
|
|
|
0x0: tgei( {{ cond = (Rs.sw >= (int16_t)INTIMM); }});
|
2009-07-22 10:51:10 +02:00
|
|
|
0x1: tgeiu({{
|
|
|
|
cond = (Rs.uw >= (uint32_t)(int32_t)(int16_t)INTIMM);
|
|
|
|
}});
|
2007-11-13 22:58:16 +01:00
|
|
|
0x2: tlti( {{ cond = (Rs.sw < (int16_t)INTIMM); }});
|
2009-07-22 10:51:10 +02:00
|
|
|
0x3: tltiu({{
|
|
|
|
cond = (Rs.uw < (uint32_t)(int32_t)(int16_t)INTIMM);
|
|
|
|
}});
|
|
|
|
0x4: teqi( {{ cond = (Rs.sw == (int16_t)INTIMM); }});
|
|
|
|
0x6: tnei( {{ cond = (Rs.sw != (int16_t)INTIMM); }});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
0x2: decode REGIMM_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
format Branch {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x0: bltzal({{ cond = (Rs.sw < 0); }}, Link);
|
|
|
|
0x1: decode RS {
|
|
|
|
0x0: bal ({{ cond = 1; }}, IsCall, Link);
|
|
|
|
default: bgezal({{ cond = (Rs.sw >= 0); }}, Link);
|
|
|
|
}
|
|
|
|
0x2: bltzall({{ cond = (Rs.sw < 0); }}, Link, Likely);
|
|
|
|
0x3: bgezall({{ cond = (Rs.sw >= 0); }}, Link, Likely);
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
0x3: decode REGIMM_LO {
|
2009-07-22 10:51:10 +02:00
|
|
|
// from Table 5-4 MIPS32 REGIMM Encoding of rt Field
|
|
|
|
// (DSP ASE MANUAL)
|
2007-06-23 01:03:42 +02:00
|
|
|
0x4: DspBranch::bposge32({{ cond = (dspctl<5:0> >= 32); }});
|
2006-02-18 09:12:04 +01:00
|
|
|
format WarnUnimpl {
|
|
|
|
0x7: synci();
|
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2006-02-20 07:49:16 +01:00
|
|
|
format Jump {
|
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
|
|
|
0x2: j({{
|
|
|
|
pc.nnpc((pc.npc() & 0xF0000000) | (JMPTARG << 2));
|
|
|
|
PCS = pc;
|
|
|
|
}});
|
|
|
|
0x3: jal({{
|
|
|
|
pc.nnpc((pc.npc() & 0xF0000000) | (JMPTARG << 2));
|
|
|
|
PCS = pc;
|
|
|
|
}}, IsCall, Link);
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
|
2006-02-16 08:51:04 +01:00
|
|
|
format Branch {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x4: decode RS_RT {
|
|
|
|
0x0: b({{ cond = 1; }});
|
|
|
|
default: beq({{ cond = (Rs.sw == Rt.sw); }});
|
2006-04-10 18:23:17 +02:00
|
|
|
}
|
2006-06-09 09:57:25 +02:00
|
|
|
0x5: bne({{ cond = (Rs.sw != Rt.sw); }});
|
|
|
|
0x6: blez({{ cond = (Rs.sw <= 0); }});
|
|
|
|
0x7: bgtz({{ cond = (Rs.sw > 0); }});
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
0x1: decode OPCODE_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format IntImmOp {
|
2007-11-13 22:58:16 +01:00
|
|
|
0x0: addi({{
|
2009-07-22 10:51:10 +02:00
|
|
|
int64_t Src1 = Rs.sw;
|
|
|
|
int64_t Src2 = imm;
|
|
|
|
int64_t temp_result;
|
2007-11-13 22:58:16 +01:00
|
|
|
#if FULL_SYSTEM
|
2009-07-22 10:51:10 +02:00
|
|
|
if (((Src1 >> 31) & 1) == 1)
|
|
|
|
Src1 |= 0x100000000LL;
|
2007-11-13 22:58:16 +01:00
|
|
|
#endif
|
2009-07-22 10:51:10 +02:00
|
|
|
temp_result = Src1 + Src2;
|
2007-11-13 22:58:16 +01:00
|
|
|
#if FULL_SYSTEM
|
2009-07-22 10:51:10 +02:00
|
|
|
if (bits(temp_result, 31) == bits(temp_result, 32)) {
|
2007-11-13 22:58:16 +01:00
|
|
|
#endif
|
2009-07-22 10:51:10 +02:00
|
|
|
Rt.sw = temp_result;
|
2007-11-13 22:58:16 +01:00
|
|
|
#if FULL_SYSTEM
|
2009-07-22 10:51:10 +02:00
|
|
|
} else {
|
|
|
|
fault = new ArithmeticFault();
|
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
#endif
|
2009-07-22 10:51:10 +02:00
|
|
|
}});
|
|
|
|
0x1: addiu({{ Rt.sw = Rs.sw + imm; }});
|
|
|
|
0x2: slti({{ Rt.sw = (Rs.sw < imm) ? 1 : 0 }});
|
2007-06-23 01:03:42 +02:00
|
|
|
|
|
|
|
//Edited to include MIPS AVP Pass/Fail instructions and
|
|
|
|
//default to the sltiu instruction
|
|
|
|
0x3: decode RS_RT_INTIMM {
|
2009-07-22 10:51:10 +02:00
|
|
|
0xabc1: BasicOp::fail({{
|
|
|
|
exitSimLoop("AVP/SRVP Test Failed");
|
|
|
|
}});
|
|
|
|
0xabc2: BasicOp::pass({{
|
|
|
|
exitSimLoop("AVP/SRVP Test Passed");
|
|
|
|
}});
|
|
|
|
default: sltiu({{
|
|
|
|
Rt.uw = (Rs.uw < (uint32_t)sextImm) ? 1 : 0;
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: andi({{ Rt.sw = Rs.sw & zextImm; }});
|
|
|
|
0x5: ori({{ Rt.sw = Rs.sw | zextImm; }});
|
|
|
|
0x6: xori({{ Rt.sw = Rs.sw ^ zextImm; }});
|
2006-04-10 18:23:17 +02:00
|
|
|
|
|
|
|
0x7: decode RS {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: lui({{ Rt = imm << 16; }});
|
2006-04-10 18:23:17 +02:00
|
|
|
}
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
0x2: decode OPCODE_LO {
|
|
|
|
//Table A-11 MIPS32 COP0 Encoding of rs Field
|
|
|
|
0x0: decode RS_MSB {
|
|
|
|
0x0: decode RS {
|
2009-07-22 10:51:10 +02:00
|
|
|
format CP0Control {
|
|
|
|
0x0: mfc0({{
|
|
|
|
Config3Reg config3 = Config3;
|
|
|
|
PageGrainReg pageGrain = PageGrain;
|
|
|
|
Rt = CP0_RD_SEL;
|
|
|
|
/* Hack for PageMask */
|
|
|
|
if (RD == 5) {
|
|
|
|
// PageMask
|
|
|
|
if (config3.sp == 0 || pageGrain.esp == 0)
|
|
|
|
Rt &= 0xFFFFE7FF;
|
|
|
|
}
|
|
|
|
}});
|
|
|
|
0x4: mtc0({{
|
|
|
|
CP0_RD_SEL = Rt;
|
|
|
|
CauseReg cause = Cause;
|
|
|
|
IntCtlReg intCtl = IntCtl;
|
|
|
|
if (RD == 11) {
|
|
|
|
// Compare
|
|
|
|
if (cause.ti == 1) {
|
|
|
|
cause.ti = 0;
|
|
|
|
int offset = 10; // corresponding to cause.ip0
|
|
|
|
offset += intCtl.ipti - 2;
|
|
|
|
replaceBits(cause, offset, offset, 0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
Cause = cause;
|
|
|
|
}});
|
|
|
|
}
|
|
|
|
format CP0Unimpl {
|
|
|
|
0x1: dmfc0();
|
|
|
|
0x5: dmtc0();
|
|
|
|
default: unknown();
|
|
|
|
}
|
|
|
|
format MT_MFTR {
|
|
|
|
// Decode MIPS MT MFTR instruction into sub-instructions
|
2007-06-23 01:03:42 +02:00
|
|
|
0x8: decode MT_U {
|
2009-07-21 05:14:15 +02:00
|
|
|
0x0: mftc0({{
|
|
|
|
data = xc->readRegOtherThread((RT << 3 | SEL) +
|
|
|
|
Ctrl_Base_DepTag);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
0x1: decode SEL {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: mftgpr({{
|
|
|
|
data = xc->readRegOtherThread(RT);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
0x1: decode RT {
|
2009-07-22 08:38:26 +02:00
|
|
|
0x0: mftlo_dsp0({{ data = xc->readRegOtherThread(INTREG_DSP_LO0); }});
|
|
|
|
0x1: mfthi_dsp0({{ data = xc->readRegOtherThread(INTREG_DSP_HI0); }});
|
|
|
|
0x2: mftacx_dsp0({{ data = xc->readRegOtherThread(INTREG_DSP_ACX0); }});
|
|
|
|
0x4: mftlo_dsp1({{ data = xc->readRegOtherThread(INTREG_DSP_LO1); }});
|
|
|
|
0x5: mfthi_dsp1({{ data = xc->readRegOtherThread(INTREG_DSP_HI1); }});
|
|
|
|
0x6: mftacx_dsp1({{ data = xc->readRegOtherThread(INTREG_DSP_ACX1); }});
|
|
|
|
0x8: mftlo_dsp2({{ data = xc->readRegOtherThread(INTREG_DSP_LO2); }});
|
|
|
|
0x9: mfthi_dsp2({{ data = xc->readRegOtherThread(INTREG_DSP_HI2); }});
|
|
|
|
0x10: mftacx_dsp2({{ data = xc->readRegOtherThread(INTREG_DSP_ACX2); }});
|
|
|
|
0x12: mftlo_dsp3({{ data = xc->readRegOtherThread(INTREG_DSP_LO3); }});
|
|
|
|
0x13: mfthi_dsp3({{ data = xc->readRegOtherThread(INTREG_DSP_HI3); }});
|
|
|
|
0x14: mftacx_dsp3({{ data = xc->readRegOtherThread(INTREG_DSP_ACX3); }});
|
|
|
|
0x16: mftdsp({{ data = xc->readRegOtherThread(INTREG_DSP_CONTROL); }});
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP0Unimpl::unknown();
|
2006-06-09 09:57:25 +02:00
|
|
|
}
|
2007-06-23 01:03:42 +02:00
|
|
|
0x2: decode MT_H {
|
|
|
|
0x0: mftc1({{ data = xc->readRegOtherThread(RT +
|
|
|
|
FP_Base_DepTag);
|
2009-07-22 10:51:10 +02:00
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
0x1: mfthc1({{ data = xc->readRegOtherThread(RT +
|
|
|
|
FP_Base_DepTag);
|
2009-07-22 10:51:10 +02:00
|
|
|
}});
|
|
|
|
}
|
|
|
|
0x3: cftc1({{
|
|
|
|
uint32_t fcsr_val = xc->readRegOtherThread(FLOATREG_FCSR +
|
2007-06-23 01:03:42 +02:00
|
|
|
FP_Base_DepTag);
|
2009-07-22 10:51:10 +02:00
|
|
|
switch (RT) {
|
|
|
|
case 0:
|
|
|
|
data = xc->readRegOtherThread(FLOATREG_FIR +
|
|
|
|
Ctrl_Base_DepTag);
|
|
|
|
break;
|
|
|
|
case 25:
|
|
|
|
data = (fcsr_val & 0xFE000000 >> 24) |
|
|
|
|
(fcsr_val & 0x00800000 >> 23);
|
|
|
|
break;
|
|
|
|
case 26:
|
|
|
|
data = fcsr_val & 0x0003F07C;
|
|
|
|
break;
|
|
|
|
case 28:
|
|
|
|
data = (fcsr_val & 0x00000F80) |
|
|
|
|
(fcsr_val & 0x01000000 >> 21) |
|
|
|
|
(fcsr_val & 0x00000003);
|
|
|
|
break;
|
|
|
|
case 31:
|
|
|
|
data = fcsr_val;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
fatal("FP Control Value (%d) Not Valid");
|
|
|
|
}
|
|
|
|
}});
|
|
|
|
default: CP0Unimpl::unknown();
|
|
|
|
}
|
|
|
|
}
|
2006-06-09 09:57:25 +02:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
format MT_MTTR {
|
|
|
|
// Decode MIPS MT MTTR instruction into sub-instructions
|
2007-06-23 01:03:42 +02:00
|
|
|
0xC: decode MT_U {
|
|
|
|
0x0: mttc0({{ xc->setRegOtherThread((RD << 3 | SEL) + Ctrl_Base_DepTag,
|
|
|
|
Rt);
|
|
|
|
}});
|
|
|
|
0x1: decode SEL {
|
|
|
|
0x0: mttgpr({{ xc->setRegOtherThread(RD, Rt); }});
|
|
|
|
0x1: decode RT {
|
2009-07-22 08:38:26 +02:00
|
|
|
0x0: mttlo_dsp0({{ xc->setRegOtherThread(INTREG_DSP_LO0, Rt);
|
2007-06-23 01:03:42 +02:00
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x1: mtthi_dsp0({{ xc->setRegOtherThread(INTREG_DSP_HI0,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x2: mttacx_dsp0({{ xc->setRegOtherThread(INTREG_DSP_ACX0,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x4: mttlo_dsp1({{ xc->setRegOtherThread(INTREG_DSP_LO1,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x5: mtthi_dsp1({{ xc->setRegOtherThread(INTREG_DSP_HI1,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x6: mttacx_dsp1({{ xc->setRegOtherThread(INTREG_DSP_ACX1,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x8: mttlo_dsp2({{ xc->setRegOtherThread(INTREG_DSP_LO2,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x9: mtthi_dsp2({{ xc->setRegOtherThread(INTREG_DSP_HI2,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x10: mttacx_dsp2({{ xc->setRegOtherThread(INTREG_DSP_ACX2,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x12: mttlo_dsp3({{ xc->setRegOtherThread(INTREG_DSP_LO3,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x13: mtthi_dsp3({{ xc->setRegOtherThread(INTREG_DSP_HI3,
|
2007-06-23 01:03:42 +02:00
|
|
|
Rt);
|
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x14: mttacx_dsp3({{ xc->setRegOtherThread(INTREG_DSP_ACX3, Rt);
|
2007-06-23 01:03:42 +02:00
|
|
|
}});
|
2009-07-22 08:38:26 +02:00
|
|
|
0x16: mttdsp({{ xc->setRegOtherThread(INTREG_DSP_CONTROL, Rt); }});
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP0Unimpl::unknown();
|
2007-11-13 22:58:16 +01:00
|
|
|
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: mttc1({{
|
|
|
|
uint64_t data = xc->readRegOtherThread(RD +
|
|
|
|
FP_Base_DepTag);
|
|
|
|
data = insertBits(data, top_bit,
|
|
|
|
bottom_bit, Rt);
|
|
|
|
xc->setRegOtherThread(RD + FP_Base_DepTag,
|
|
|
|
data);
|
|
|
|
}});
|
|
|
|
0x3: cttc1({{
|
|
|
|
uint32_t data;
|
|
|
|
switch (RD) {
|
|
|
|
case 25:
|
|
|
|
data = (Rt.uw<7:1> << 25) | // move 31-25
|
|
|
|
(FCSR & 0x01000000) | // bit 24
|
|
|
|
(FCSR & 0x004FFFFF); // bit 22-0
|
|
|
|
break;
|
|
|
|
case 26:
|
|
|
|
data = (FCSR & 0xFFFC0000) | // move 31-18
|
|
|
|
Rt.uw<17:12> << 12 | // bit 17-12
|
|
|
|
(FCSR & 0x00000F80) << 7 | // bit 11-7
|
|
|
|
Rt.uw<6:2> << 2 | // bit 6-2
|
|
|
|
(FCSR & 0x00000002); // bit 1...0
|
|
|
|
break;
|
|
|
|
case 28:
|
|
|
|
data = (FCSR & 0xFE000000) | // move 31-25
|
|
|
|
Rt.uw<2:2> << 24 | // bit 24
|
|
|
|
(FCSR & 0x00FFF000) << 23 | // bit 23-12
|
|
|
|
Rt.uw<11:7> << 7 | // bit 24
|
|
|
|
(FCSR & 0x000007E) |
|
|
|
|
Rt.uw<1:0>; // bit 22-0
|
|
|
|
break;
|
|
|
|
case 31:
|
|
|
|
data = Rt.uw;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
panic("FP Control Value (%d) "
|
|
|
|
"Not Available. Ignoring "
|
|
|
|
"Access to Floating Control "
|
|
|
|
"Status Register", FS);
|
|
|
|
}
|
|
|
|
xc->setRegOtherThread(FLOATREG_FCSR + FP_Base_DepTag, data);
|
|
|
|
}});
|
|
|
|
default: CP0Unimpl::unknown();
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
0xB: decode RD {
|
|
|
|
format MT_Control {
|
|
|
|
0x0: decode POS {
|
|
|
|
0x0: decode SEL {
|
|
|
|
0x1: decode SC {
|
2009-07-21 05:14:15 +02:00
|
|
|
0x0: dvpe({{
|
|
|
|
MVPControlReg mvpControl = MVPControl;
|
|
|
|
VPEConf0Reg vpeConf0 = VPEConf0;
|
|
|
|
Rt = MVPControl;
|
|
|
|
if (vpeConf0.mvp == 1)
|
|
|
|
mvpControl.evp = 0;
|
|
|
|
MVPControl = mvpControl;
|
|
|
|
}});
|
|
|
|
0x1: evpe({{
|
|
|
|
MVPControlReg mvpControl = MVPControl;
|
|
|
|
VPEConf0Reg vpeConf0 = VPEConf0;
|
|
|
|
Rt = MVPControl;
|
|
|
|
if (vpeConf0.mvp == 1)
|
|
|
|
mvpControl.evp = 1;
|
|
|
|
MVPControl = mvpControl;
|
|
|
|
}});
|
2007-11-13 22:58:16 +01:00
|
|
|
default:CP0Unimpl::unknown();
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default:CP0Unimpl::unknown();
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default:CP0Unimpl::unknown();
|
|
|
|
}
|
2007-06-23 01:03:42 +02:00
|
|
|
0x1: decode POS {
|
|
|
|
0xF: decode SEL {
|
|
|
|
0x1: decode SC {
|
2009-07-21 05:14:15 +02:00
|
|
|
0x0: dmt({{
|
|
|
|
VPEControlReg vpeControl = VPEControl;
|
|
|
|
Rt = vpeControl;
|
|
|
|
vpeControl.te = 0;
|
|
|
|
VPEControl = vpeControl;
|
|
|
|
}});
|
|
|
|
0x1: emt({{
|
|
|
|
VPEControlReg vpeControl = VPEControl;
|
|
|
|
Rt = vpeControl;
|
|
|
|
vpeControl.te = 1;
|
|
|
|
VPEControl = vpeControl;
|
|
|
|
}});
|
2007-11-13 22:58:16 +01:00
|
|
|
default:CP0Unimpl::unknown();
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default:CP0Unimpl::unknown();
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
default:CP0Unimpl::unknown();
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0xC: decode POS {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: decode SC {
|
|
|
|
0x0: CP0Control::di({{
|
|
|
|
StatusReg status = Status;
|
|
|
|
ConfigReg config = Config;
|
|
|
|
// Rev 2.0 or beyond?
|
|
|
|
if (config.ar >= 1) {
|
|
|
|
Rt = status;
|
|
|
|
status.ie = 0;
|
|
|
|
} else {
|
|
|
|
// Enable this else branch once we
|
|
|
|
// actually set values for Config on init
|
|
|
|
fault = new ReservedInstructionFault();
|
|
|
|
}
|
|
|
|
Status = status;
|
|
|
|
}});
|
|
|
|
0x1: CP0Control::ei({{
|
|
|
|
StatusReg status = Status;
|
|
|
|
ConfigReg config = Config;
|
|
|
|
if (config.ar >= 1) {
|
|
|
|
Rt = status;
|
|
|
|
status.ie = 1;
|
|
|
|
} else {
|
|
|
|
fault = new ReservedInstructionFault();
|
|
|
|
}
|
|
|
|
}});
|
|
|
|
default:CP0Unimpl::unknown();
|
|
|
|
}
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP0Unimpl::unknown();
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
format CP0Control {
|
|
|
|
0xA: rdpgpr({{
|
2009-07-21 05:14:15 +02:00
|
|
|
ConfigReg config = Config;
|
|
|
|
if (config.ar >= 1) {
|
|
|
|
// Rev 2 of the architecture
|
|
|
|
panic("Shadow Sets Not Fully Implemented.\n");
|
|
|
|
} else {
|
2007-06-23 01:03:42 +02:00
|
|
|
fault = new ReservedInstructionFault();
|
|
|
|
}
|
2009-07-21 05:14:15 +02:00
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
0xE: wrpgpr({{
|
2009-07-21 05:14:15 +02:00
|
|
|
ConfigReg config = Config;
|
|
|
|
if (config.ar >= 1) {
|
|
|
|
// Rev 2 of the architecture
|
|
|
|
panic("Shadow Sets Not Fully Implemented.\n");
|
|
|
|
} else {
|
|
|
|
fault = new ReservedInstructionFault();
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
2009-07-21 05:14:15 +02:00
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
//Table A-12 MIPS32 COP0 Encoding of Function Field When rs=CO
|
|
|
|
0x1: decode FUNCTION {
|
2009-07-22 10:51:10 +02:00
|
|
|
format CP0Control {
|
|
|
|
0x18: eret({{
|
|
|
|
StatusReg status = Status;
|
|
|
|
ConfigReg config = Config;
|
|
|
|
SRSCtlReg srsCtl = SRSCtl;
|
|
|
|
DPRINTF(MipsPRA,"Restoring PC - %x\n",EPC);
|
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
|
|
|
MipsISA::PCState pc = PCS;
|
2009-07-22 10:51:10 +02:00
|
|
|
if (status.erl == 1) {
|
|
|
|
status.erl = 0;
|
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.npc(ErrorEPC);
|
2009-07-22 10:51:10 +02:00
|
|
|
// Need to adjust NNPC, otherwise things break
|
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.nnpc(ErrorEPC + sizeof(MachInst));
|
2009-07-22 10:51:10 +02:00
|
|
|
} else {
|
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.npc(EPC);
|
2009-07-22 10:51:10 +02:00
|
|
|
// Need to adjust NNPC, otherwise things break
|
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.nnpc(EPC + sizeof(MachInst));
|
2009-07-22 10:51:10 +02:00
|
|
|
status.exl = 0;
|
|
|
|
if (config.ar >=1 &&
|
|
|
|
srsCtl.hss > 0 &&
|
|
|
|
status.bev == 0) {
|
|
|
|
srsCtl.css = srsCtl.pss;
|
|
|
|
//xc->setShadowSet(srsCtl.pss);
|
|
|
|
}
|
2009-07-21 05:14:15 +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
|
|
|
PCS = pc;
|
2009-07-22 10:51:10 +02:00
|
|
|
LLFlag = 0;
|
|
|
|
Status = status;
|
|
|
|
SRSCtl = srsCtl;
|
|
|
|
}}, IsReturn, IsSerializing, IsERET);
|
|
|
|
|
|
|
|
0x1F: deret({{
|
|
|
|
DebugReg debug = Debug;
|
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
|
|
|
MipsISA::PCState pc = PCS;
|
2009-07-22 10:51:10 +02:00
|
|
|
if (debug.dm == 1) {
|
|
|
|
debug.dm = 1;
|
|
|
|
debug.iexi = 0;
|
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.npc(DEPC);
|
2009-07-22 10:51:10 +02:00
|
|
|
} else {
|
|
|
|
// Undefined;
|
|
|
|
}
|
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
|
|
|
PCS = pc;
|
2009-07-22 10:51:10 +02:00
|
|
|
Debug = debug;
|
|
|
|
}}, IsReturn, IsSerializing, IsERET);
|
|
|
|
}
|
|
|
|
format CP0TLB {
|
|
|
|
0x01: tlbr({{
|
|
|
|
MipsISA::PTE *PTEntry =
|
|
|
|
xc->tcBase()->getITBPtr()->
|
|
|
|
getEntry(Index & 0x7FFFFFFF);
|
|
|
|
if (PTEntry == NULL) {
|
|
|
|
fatal("Invalid PTE Entry received on "
|
|
|
|
"a TLBR instruction\n");
|
|
|
|
}
|
|
|
|
/* Setup PageMask */
|
|
|
|
// If 1KB pages are not enabled, a read of PageMask
|
|
|
|
// must return 0b00 in bits 12, 11
|
|
|
|
PageMask = (PTEntry->Mask << 11);
|
|
|
|
/* Setup EntryHi */
|
|
|
|
EntryHi = ((PTEntry->VPN << 11) | (PTEntry->asid));
|
|
|
|
/* Setup Entry Lo0 */
|
|
|
|
EntryLo0 = ((PTEntry->PFN0 << 6) |
|
|
|
|
(PTEntry->C0 << 3) |
|
|
|
|
(PTEntry->D0 << 2) |
|
|
|
|
(PTEntry->V0 << 1) |
|
|
|
|
PTEntry->G);
|
|
|
|
/* Setup Entry Lo1 */
|
|
|
|
EntryLo1 = ((PTEntry->PFN1 << 6) |
|
|
|
|
(PTEntry->C1 << 3) |
|
|
|
|
(PTEntry->D1 << 2) |
|
|
|
|
(PTEntry->V1 << 1) |
|
|
|
|
PTEntry->G);
|
|
|
|
}}); // Need to hook up to TLB
|
|
|
|
|
|
|
|
0x02: tlbwi({{
|
|
|
|
//Create PTE
|
|
|
|
MipsISA::PTE newEntry;
|
|
|
|
//Write PTE
|
|
|
|
newEntry.Mask = (Addr)(PageMask >> 11);
|
|
|
|
newEntry.VPN = (Addr)(EntryHi >> 11);
|
|
|
|
/* PageGrain _ ESP Config3 _ SP */
|
|
|
|
if (bits(PageGrain, 28) == 0 || bits(Config3, 4) ==0) {
|
|
|
|
// If 1KB pages are *NOT* enabled, lowest bits of
|
|
|
|
// the mask are 0b11 for TLB writes
|
|
|
|
newEntry.Mask |= 0x3;
|
|
|
|
// Reset bits 0 and 1 if 1KB pages are not enabled
|
|
|
|
newEntry.VPN &= 0xFFFFFFFC;
|
|
|
|
}
|
|
|
|
newEntry.asid = (uint8_t)(EntryHi & 0xFF);
|
|
|
|
|
|
|
|
newEntry.PFN0 = (Addr)(EntryLo0 >> 6);
|
|
|
|
newEntry.PFN1 = (Addr)(EntryLo1 >> 6);
|
|
|
|
newEntry.D0 = (bool)((EntryLo0 >> 2) & 1);
|
|
|
|
newEntry.D1 = (bool)((EntryLo1 >> 2) & 1);
|
|
|
|
newEntry.V1 = (bool)((EntryLo1 >> 1) & 1);
|
|
|
|
newEntry.V0 = (bool)((EntryLo0 >> 1) & 1);
|
|
|
|
newEntry.G = (bool)((EntryLo0 & EntryLo1) & 1);
|
|
|
|
newEntry.C0 = (uint8_t)((EntryLo0 >> 3) & 0x7);
|
|
|
|
newEntry.C1 = (uint8_t)((EntryLo1 >> 3) & 0x7);
|
|
|
|
/* Now, compute the AddrShiftAmount and OffsetMask -
|
|
|
|
TLB optimizations */
|
|
|
|
/* Addr Shift Amount for 1KB or larger pages */
|
|
|
|
if ((newEntry.Mask & 0xFFFF) == 3) {
|
|
|
|
newEntry.AddrShiftAmount = 12;
|
|
|
|
} else if ((newEntry.Mask & 0xFFFF) == 0x0000) {
|
|
|
|
newEntry.AddrShiftAmount = 10;
|
|
|
|
} else if ((newEntry.Mask & 0xFFFC) == 0x000C) {
|
|
|
|
newEntry.AddrShiftAmount = 14;
|
|
|
|
} else if ((newEntry.Mask & 0xFFF0) == 0x0030) {
|
|
|
|
newEntry.AddrShiftAmount = 16;
|
|
|
|
} else if ((newEntry.Mask & 0xFFC0) == 0x00C0) {
|
|
|
|
newEntry.AddrShiftAmount = 18;
|
|
|
|
} else if ((newEntry.Mask & 0xFF00) == 0x0300) {
|
|
|
|
newEntry.AddrShiftAmount = 20;
|
|
|
|
} else if ((newEntry.Mask & 0xFC00) == 0x0C00) {
|
|
|
|
newEntry.AddrShiftAmount = 22;
|
|
|
|
} else if ((newEntry.Mask & 0xF000) == 0x3000) {
|
|
|
|
newEntry.AddrShiftAmount = 24;
|
|
|
|
} else if ((newEntry.Mask & 0xC000) == 0xC000) {
|
|
|
|
newEntry.AddrShiftAmount = 26;
|
|
|
|
} else if ((newEntry.Mask & 0x30000) == 0x30000) {
|
|
|
|
newEntry.AddrShiftAmount = 28;
|
|
|
|
} else {
|
|
|
|
fatal("Invalid Mask Pattern Detected!\n");
|
|
|
|
}
|
|
|
|
newEntry.OffsetMask =
|
|
|
|
(1 << newEntry.AddrShiftAmount) - 1;
|
|
|
|
|
|
|
|
MipsISA::TLB *Ptr = xc->tcBase()->getITBPtr();
|
|
|
|
Config3Reg config3 = Config3;
|
|
|
|
PageGrainReg pageGrain = PageGrain;
|
|
|
|
int SP = 0;
|
|
|
|
if (bits(config3, config3.sp) == 1 &&
|
|
|
|
bits(pageGrain, pageGrain.esp) == 1) {
|
|
|
|
SP = 1;
|
|
|
|
}
|
|
|
|
IndexReg index = Index;
|
|
|
|
Ptr->insertAt(newEntry, Index & 0x7FFFFFFF, SP);
|
|
|
|
}});
|
|
|
|
0x06: tlbwr({{
|
|
|
|
//Create PTE
|
|
|
|
MipsISA::PTE newEntry;
|
|
|
|
//Write PTE
|
|
|
|
newEntry.Mask = (Addr)(PageMask >> 11);
|
|
|
|
newEntry.VPN = (Addr)(EntryHi >> 11);
|
|
|
|
/* PageGrain _ ESP Config3 _ SP */
|
|
|
|
if (bits(PageGrain, 28) == 0 ||
|
|
|
|
bits(Config3, 4) == 0) {
|
|
|
|
// If 1KB pages are *NOT* enabled, lowest bits of
|
|
|
|
// the mask are 0b11 for TLB writes
|
|
|
|
newEntry.Mask |= 0x3;
|
|
|
|
// Reset bits 0 and 1 if 1KB pages are not enabled
|
|
|
|
newEntry.VPN &= 0xFFFFFFFC;
|
|
|
|
}
|
|
|
|
newEntry.asid = (uint8_t)(EntryHi & 0xFF);
|
|
|
|
|
|
|
|
newEntry.PFN0 = (Addr)(EntryLo0 >> 6);
|
|
|
|
newEntry.PFN1 = (Addr)(EntryLo1 >> 6);
|
|
|
|
newEntry.D0 = (bool)((EntryLo0 >> 2) & 1);
|
|
|
|
newEntry.D1 = (bool)((EntryLo1 >> 2) & 1);
|
|
|
|
newEntry.V1 = (bool)((EntryLo1 >> 1) & 1);
|
|
|
|
newEntry.V0 = (bool)((EntryLo0 >> 1) & 1);
|
|
|
|
newEntry.G = (bool)((EntryLo0 & EntryLo1) & 1);
|
|
|
|
newEntry.C0 = (uint8_t)((EntryLo0 >> 3) & 0x7);
|
|
|
|
newEntry.C1 = (uint8_t)((EntryLo1 >> 3) & 0x7);
|
|
|
|
/* Now, compute the AddrShiftAmount and OffsetMask -
|
|
|
|
TLB optimizations */
|
|
|
|
/* Addr Shift Amount for 1KB or larger pages */
|
|
|
|
if ((newEntry.Mask & 0xFFFF) == 3){
|
|
|
|
newEntry.AddrShiftAmount = 12;
|
|
|
|
} else if ((newEntry.Mask & 0xFFFF) == 0x0000) {
|
|
|
|
newEntry.AddrShiftAmount = 10;
|
|
|
|
} else if ((newEntry.Mask & 0xFFFC) == 0x000C) {
|
|
|
|
newEntry.AddrShiftAmount = 14;
|
|
|
|
} else if ((newEntry.Mask & 0xFFF0) == 0x0030) {
|
|
|
|
newEntry.AddrShiftAmount = 16;
|
|
|
|
} else if ((newEntry.Mask & 0xFFC0) == 0x00C0) {
|
|
|
|
newEntry.AddrShiftAmount = 18;
|
|
|
|
} else if ((newEntry.Mask & 0xFF00) == 0x0300) {
|
|
|
|
newEntry.AddrShiftAmount = 20;
|
|
|
|
} else if ((newEntry.Mask & 0xFC00) == 0x0C00) {
|
|
|
|
newEntry.AddrShiftAmount = 22;
|
|
|
|
} else if ((newEntry.Mask & 0xF000) == 0x3000) {
|
|
|
|
newEntry.AddrShiftAmount = 24;
|
|
|
|
} else if ((newEntry.Mask & 0xC000) == 0xC000) {
|
|
|
|
newEntry.AddrShiftAmount = 26;
|
|
|
|
} else if ((newEntry.Mask & 0x30000) == 0x30000) {
|
|
|
|
newEntry.AddrShiftAmount = 28;
|
|
|
|
} else {
|
|
|
|
fatal("Invalid Mask Pattern Detected!\n");
|
|
|
|
}
|
|
|
|
newEntry.OffsetMask =
|
|
|
|
(1 << newEntry.AddrShiftAmount) - 1;
|
|
|
|
|
|
|
|
MipsISA::TLB *Ptr = xc->tcBase()->getITBPtr();
|
|
|
|
Config3Reg config3 = Config3;
|
|
|
|
PageGrainReg pageGrain = PageGrain;
|
|
|
|
int SP = 0;
|
|
|
|
if (bits(config3, config3.sp) == 1 &&
|
|
|
|
bits(pageGrain, pageGrain.esp) == 1) {
|
|
|
|
SP = 1;
|
|
|
|
}
|
|
|
|
IndexReg index = Index;
|
|
|
|
Ptr->insertAt(newEntry, Random, SP);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
0x08: tlbp({{
|
|
|
|
Config3Reg config3 = Config3;
|
|
|
|
PageGrainReg pageGrain = PageGrain;
|
|
|
|
EntryHiReg entryHi = EntryHi;
|
|
|
|
int tlbIndex;
|
|
|
|
Addr vpn;
|
|
|
|
if (pageGrain.esp == 1 && config3.sp ==1) {
|
|
|
|
vpn = EntryHi >> 11;
|
|
|
|
} else {
|
|
|
|
// Mask off lower 2 bits
|
|
|
|
vpn = ((EntryHi >> 11) & 0xFFFFFFFC);
|
|
|
|
}
|
|
|
|
tlbIndex = xc->tcBase()->getITBPtr()->
|
2009-07-22 10:57:55 +02:00
|
|
|
probeEntry(vpn, entryHi.asid);
|
2009-07-22 10:51:10 +02:00
|
|
|
// Check TLB for entry matching EntryHi
|
|
|
|
if (tlbIndex != -1) {
|
|
|
|
Index = tlbIndex;
|
|
|
|
} else {
|
|
|
|
// else, set Index = 1 << 31
|
|
|
|
Index = (1 << 31);
|
|
|
|
}
|
|
|
|
}});
|
|
|
|
}
|
|
|
|
format CP0Unimpl {
|
|
|
|
0x20: wait();
|
|
|
|
}
|
|
|
|
default: CP0Unimpl::unknown();
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
//Table A-13 MIPS32 COP1 Encoding of rs Field
|
|
|
|
0x1: decode RS_MSB {
|
2006-02-08 00:36:08 +01:00
|
|
|
0x0: decode RS_HI {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x0: decode RS_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format CP1Control {
|
2006-06-11 21:38:40 +02:00
|
|
|
0x0: mfc1 ({{ Rt.uw = Fs.uw; }});
|
2006-04-10 18:23:17 +02:00
|
|
|
|
|
|
|
0x2: cfc1({{
|
2009-07-22 10:51:10 +02:00
|
|
|
switch (FS) {
|
2006-04-27 11:07:11 +02:00
|
|
|
case 0:
|
2006-05-12 08:57:32 +02:00
|
|
|
Rt = FIR;
|
2006-04-27 11:07:11 +02:00
|
|
|
break;
|
|
|
|
case 25:
|
2009-07-22 10:51:10 +02:00
|
|
|
Rt = (FCSR & 0xFE000000) >> 24 |
|
|
|
|
(FCSR & 0x00800000) >> 23;
|
2006-04-27 11:07:11 +02:00
|
|
|
break;
|
|
|
|
case 26:
|
2009-07-22 10:51:10 +02:00
|
|
|
Rt = (FCSR & 0x0003F07C);
|
2006-04-27 11:07:11 +02:00
|
|
|
break;
|
|
|
|
case 28:
|
2009-07-22 10:51:10 +02:00
|
|
|
Rt = (FCSR & 0x00000F80) |
|
|
|
|
(FCSR & 0x01000000) >> 21 |
|
|
|
|
(FCSR & 0x00000003);
|
2006-04-27 11:07:11 +02:00
|
|
|
break;
|
|
|
|
case 31:
|
2006-05-12 08:57:32 +02:00
|
|
|
Rt = FCSR;
|
2006-04-27 11:07:11 +02:00
|
|
|
break;
|
|
|
|
default:
|
2007-11-13 22:58:16 +01:00
|
|
|
warn("FP Control Value (%d) Not Valid");
|
2006-04-10 18:23:17 +02:00
|
|
|
}
|
|
|
|
}});
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
0x3: mfhc1({{ Rt.uw = Fs.ud<63:32>; }});
|
2006-06-09 09:57:25 +02:00
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: mtc1({{ Fs.uw = Rt.uw; }});
|
2006-06-09 09:57:25 +02:00
|
|
|
|
2006-04-10 18:23:17 +02:00
|
|
|
0x6: ctc1({{
|
2009-07-22 10:51:10 +02:00
|
|
|
switch (FS) {
|
2006-04-27 11:07:11 +02:00
|
|
|
case 25:
|
2009-07-22 10:51:10 +02:00
|
|
|
FCSR = (Rt.uw<7:1> << 25) | // move 31-25
|
|
|
|
(FCSR & 0x01000000) | // bit 24
|
|
|
|
(FCSR & 0x004FFFFF); // bit 22-0
|
2006-04-27 11:07:11 +02:00
|
|
|
break;
|
|
|
|
case 26:
|
2009-07-22 10:51:10 +02:00
|
|
|
FCSR = (FCSR & 0xFFFC0000) | // move 31-18
|
|
|
|
Rt.uw<17:12> << 12 | // bit 17-12
|
|
|
|
(FCSR & 0x00000F80) << 7 | // bit 11-7
|
|
|
|
Rt.uw<6:2> << 2 | // bit 6-2
|
|
|
|
(FCSR & 0x00000002); // bit 1-0
|
2006-04-27 11:07:11 +02:00
|
|
|
break;
|
|
|
|
case 28:
|
2009-07-22 10:51:10 +02:00
|
|
|
FCSR = (FCSR & 0xFE000000) | // move 31-25
|
|
|
|
Rt.uw<2:2> << 24 | // bit 24
|
|
|
|
(FCSR & 0x00FFF000) << 23 | // bit 23-12
|
|
|
|
Rt.uw<11:7> << 7 | // bit 24
|
|
|
|
(FCSR & 0x000007E) |
|
|
|
|
Rt.uw<1:0>; // bit 22-0
|
2006-04-27 11:07:11 +02:00
|
|
|
break;
|
|
|
|
case 31:
|
2009-07-22 10:51:10 +02:00
|
|
|
FCSR = Rt.uw;
|
2006-04-27 11:07:11 +02:00
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
2009-07-22 10:51:10 +02:00
|
|
|
panic("FP Control Value (%d) "
|
|
|
|
"Not Available. Ignoring Access "
|
|
|
|
"to Floating Control Status "
|
|
|
|
"Register", FS);
|
2006-04-27 11:07:11 +02:00
|
|
|
}
|
2006-04-10 18:23:17 +02:00
|
|
|
}});
|
2006-06-09 09:57:25 +02:00
|
|
|
|
|
|
|
0x7: mthc1({{
|
|
|
|
uint64_t fs_hi = Rt.uw;
|
|
|
|
uint64_t fs_lo = Fs.ud & 0x0FFFFFFFF;
|
|
|
|
Fs.ud = (fs_hi << 32) | fs_lo;
|
|
|
|
}});
|
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
format CP1Unimpl {
|
|
|
|
0x1: dmfc1();
|
|
|
|
0x5: dmtc1();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
0x1: decode RS_LO {
|
|
|
|
0x0: decode ND {
|
|
|
|
format Branch {
|
|
|
|
0x0: decode TF {
|
|
|
|
0x0: bc1f({{
|
|
|
|
cond = getCondCode(FCSR, BRANCH_CC) == 0;
|
|
|
|
}});
|
|
|
|
0x1: bc1t({{
|
|
|
|
cond = getCondCode(FCSR, BRANCH_CC) == 1;
|
|
|
|
}});
|
|
|
|
}
|
|
|
|
0x1: decode TF {
|
|
|
|
0x0: bc1fl({{
|
|
|
|
cond = getCondCode(FCSR, BRANCH_CC) == 0;
|
|
|
|
}}, Likely);
|
|
|
|
0x1: bc1tl({{
|
|
|
|
cond = getCondCode(FCSR, BRANCH_CC) == 1;
|
|
|
|
}}, Likely);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
format CP1Unimpl {
|
|
|
|
0x1: bc1any2();
|
|
|
|
0x2: bc1any4();
|
|
|
|
default: unknown();
|
|
|
|
}
|
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
0x1: decode RS_HI {
|
|
|
|
0x2: decode RS_LO {
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table A-14 MIPS32 COP1 Encoding of Function Field When
|
|
|
|
//rs=S (( single-precision floating point))
|
2006-04-26 22:13:47 +02:00
|
|
|
0x0: decode FUNCTION_HI {
|
|
|
|
0x0: decode FUNCTION_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
format FloatOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: add_s({{ Fd.sf = Fs.sf + Ft.sf; }});
|
|
|
|
0x1: sub_s({{ Fd.sf = Fs.sf - Ft.sf; }});
|
|
|
|
0x2: mul_s({{ Fd.sf = Fs.sf * Ft.sf; }});
|
|
|
|
0x3: div_s({{ Fd.sf = Fs.sf / Ft.sf; }});
|
|
|
|
0x4: sqrt_s({{ Fd.sf = sqrt(Fs.sf); }});
|
|
|
|
0x5: abs_s({{ Fd.sf = fabs(Fs.sf); }});
|
|
|
|
0x7: neg_s({{ Fd.sf = -Fs.sf; }});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
0x6: BasicOp::mov_s({{ Fd.sf = Fs.sf; }});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-04-26 22:13:47 +02:00
|
|
|
0x1: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format FloatConvertOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: round_l_s({{ val = Fs.sf; }},
|
|
|
|
ToLong, Round);
|
|
|
|
0x1: trunc_l_s({{ val = Fs.sf; }},
|
|
|
|
ToLong, Trunc);
|
|
|
|
0x2: ceil_l_s({{ val = Fs.sf;}},
|
|
|
|
ToLong, Ceil);
|
|
|
|
0x3: floor_l_s({{ val = Fs.sf; }},
|
|
|
|
ToLong, Floor);
|
|
|
|
0x4: round_w_s({{ val = Fs.sf; }},
|
|
|
|
ToWord, Round);
|
|
|
|
0x5: trunc_w_s({{ val = Fs.sf; }},
|
|
|
|
ToWord, Trunc);
|
|
|
|
0x6: ceil_w_s({{ val = Fs.sf; }},
|
|
|
|
ToWord, Ceil);
|
|
|
|
0x7: floor_w_s({{ val = Fs.sf; }},
|
|
|
|
ToWord, Floor);
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-04-26 22:13:47 +02:00
|
|
|
0x2: decode FUNCTION_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x1: decode MOVCF {
|
2006-06-09 09:57:25 +02:00
|
|
|
format BasicOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: movf_s({{
|
|
|
|
Fd = (getCondCode(FCSR,CC) == 0) ?
|
|
|
|
Fs : Fd;
|
|
|
|
}});
|
|
|
|
0x1: movt_s({{
|
|
|
|
Fd = (getCondCode(FCSR,CC) == 1) ?
|
|
|
|
Fs : Fd;
|
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
format BasicOp {
|
|
|
|
0x2: movz_s({{ Fd = (Rt == 0) ? Fs : Fd; }});
|
|
|
|
0x3: movn_s({{ Fd = (Rt != 0) ? Fs : Fd; }});
|
|
|
|
}
|
|
|
|
|
2006-05-09 20:39:45 +02:00
|
|
|
format FloatOp {
|
|
|
|
0x5: recip_s({{ Fd = 1 / Fs; }});
|
2009-07-22 10:51:10 +02:00
|
|
|
0x6: rsqrt_s({{ Fd = 1 / sqrt(Fs); }});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
format CP1Unimpl {
|
2009-07-22 10:51:10 +02:00
|
|
|
default: unknown();
|
2007-11-13 22:58:16 +01:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
0x3: CP1Unimpl::unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
|
2006-04-26 22:13:47 +02:00
|
|
|
0x4: decode FUNCTION_LO {
|
2006-05-09 21:18:36 +02:00
|
|
|
format FloatConvertOp {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x1: cvt_d_s({{ val = Fs.sf; }}, ToDouble);
|
|
|
|
0x4: cvt_w_s({{ val = Fs.sf; }}, ToWord);
|
|
|
|
0x5: cvt_l_s({{ val = Fs.sf; }}, ToLong);
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x6: FloatOp::cvt_ps_s({{
|
2009-07-22 10:51:10 +02:00
|
|
|
Fd.ud = (uint64_t) Fs.uw << 32 |
|
|
|
|
(uint64_t) Ft.uw;
|
|
|
|
}});
|
2007-11-13 22:58:16 +01:00
|
|
|
format CP1Unimpl {
|
2009-07-22 10:51:10 +02:00
|
|
|
default: unknown();
|
2007-11-13 22:58:16 +01:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
0x5: CP1Unimpl::unknown();
|
2006-05-09 20:39:45 +02:00
|
|
|
|
|
|
|
0x6: decode FUNCTION_LO {
|
2006-05-09 21:18:36 +02:00
|
|
|
format FloatCompareOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: c_f_s({{ cond = 0; }},
|
|
|
|
SinglePrecision, UnorderedFalse);
|
|
|
|
0x1: c_un_s({{ cond = 0; }},
|
|
|
|
SinglePrecision, UnorderedTrue);
|
2006-06-09 09:57:25 +02:00
|
|
|
0x2: c_eq_s({{ cond = (Fs.sf == Ft.sf); }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x3: c_ueq_s({{ cond = (Fs.sf == Ft.sf); }},
|
|
|
|
UnorderedTrue);
|
|
|
|
0x4: c_olt_s({{ cond = (Fs.sf < Ft.sf); }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x5: c_ult_s({{ cond = (Fs.sf < Ft.sf); }},
|
|
|
|
UnorderedTrue);
|
|
|
|
0x6: c_ole_s({{ cond = (Fs.sf <= Ft.sf); }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x7: c_ule_s({{ cond = (Fs.sf <= Ft.sf); }},
|
|
|
|
UnorderedTrue);
|
2006-05-09 20:39:45 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
0x7: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format FloatCompareOp {
|
|
|
|
0x0: c_sf_s({{ cond = 0; }}, SinglePrecision,
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x1: c_ngle_s({{ cond = 0; }}, SinglePrecision,
|
|
|
|
UnorderedTrue, QnanException);
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: c_seq_s({{ cond = (Fs.sf == Ft.sf); }},
|
2006-06-09 09:57:25 +02:00
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x3: c_ngl_s({{ cond = (Fs.sf == Ft.sf); }},
|
|
|
|
UnorderedTrue, QnanException);
|
|
|
|
0x4: c_lt_s({{ cond = (Fs.sf < Ft.sf); }},
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x5: c_nge_s({{ cond = (Fs.sf < Ft.sf); }},
|
|
|
|
UnorderedTrue, QnanException);
|
|
|
|
0x6: c_le_s({{ cond = (Fs.sf <= Ft.sf); }},
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x7: c_ngt_s({{ cond = (Fs.sf <= Ft.sf); }},
|
|
|
|
UnorderedTrue, QnanException);
|
2006-05-09 20:39:45 +02:00
|
|
|
}
|
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table A-15 MIPS32 COP1 Encoding of Function Field When
|
|
|
|
//rs=D
|
2006-04-26 22:13:47 +02:00
|
|
|
0x1: decode FUNCTION_HI {
|
|
|
|
0x0: decode FUNCTION_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
format FloatOp {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x0: add_d({{ Fd.df = Fs.df + Ft.df; }});
|
|
|
|
0x1: sub_d({{ Fd.df = Fs.df - Ft.df; }});
|
|
|
|
0x2: mul_d({{ Fd.df = Fs.df * Ft.df; }});
|
|
|
|
0x3: div_d({{ Fd.df = Fs.df / Ft.df; }});
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: sqrt_d({{ Fd.df = sqrt(Fs.df); }});
|
|
|
|
0x5: abs_d({{ Fd.df = fabs(Fs.df); }});
|
|
|
|
0x7: neg_d({{ Fd.df = -1 * Fs.df; }});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
0x6: BasicOp::mov_d({{ Fd.df = Fs.df; }});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-04-26 22:13:47 +02:00
|
|
|
0x1: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format FloatConvertOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: round_l_d({{ val = Fs.df; }},
|
|
|
|
ToLong, Round);
|
|
|
|
0x1: trunc_l_d({{ val = Fs.df; }},
|
|
|
|
ToLong, Trunc);
|
|
|
|
0x2: ceil_l_d({{ val = Fs.df; }},
|
|
|
|
ToLong, Ceil);
|
|
|
|
0x3: floor_l_d({{ val = Fs.df; }},
|
|
|
|
ToLong, Floor);
|
|
|
|
0x4: round_w_d({{ val = Fs.df; }},
|
|
|
|
ToWord, Round);
|
|
|
|
0x5: trunc_w_d({{ val = Fs.df; }},
|
|
|
|
ToWord, Trunc);
|
|
|
|
0x6: ceil_w_d({{ val = Fs.df; }},
|
|
|
|
ToWord, Ceil);
|
|
|
|
0x7: floor_w_d({{ val = Fs.df; }},
|
|
|
|
ToWord, Floor);
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-04-26 22:13:47 +02:00
|
|
|
0x2: decode FUNCTION_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x1: decode MOVCF {
|
2006-06-09 09:57:25 +02:00
|
|
|
format BasicOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: movf_d({{
|
|
|
|
Fd.df = (getCondCode(FCSR,CC) == 0) ?
|
2006-06-09 09:57:25 +02:00
|
|
|
Fs.df : Fd.df;
|
2009-07-22 10:51:10 +02:00
|
|
|
}});
|
|
|
|
0x1: movt_d({{
|
|
|
|
Fd.df = (getCondCode(FCSR,CC) == 1) ?
|
2006-06-09 09:57:25 +02:00
|
|
|
Fs.df : Fd.df;
|
2009-07-22 10:51:10 +02:00
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
format BasicOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: movz_d({{
|
|
|
|
Fd.df = (Rt == 0) ? Fs.df : Fd.df;
|
|
|
|
}});
|
|
|
|
0x3: movn_d({{
|
|
|
|
Fd.df = (Rt != 0) ? Fs.df : Fd.df;
|
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
|
2006-05-10 22:52:27 +02:00
|
|
|
format FloatOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x5: recip_d({{ Fd.df = 1 / Fs.df; }});
|
|
|
|
0x6: rsqrt_d({{ Fd.df = 1 / sqrt(Fs.df); }});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
format CP1Unimpl {
|
2009-07-22 10:51:10 +02:00
|
|
|
default: unknown();
|
2007-11-13 22:58:16 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2007-11-13 22:58:16 +01:00
|
|
|
}
|
2006-04-26 22:13:47 +02:00
|
|
|
0x4: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format FloatConvertOp {
|
|
|
|
0x0: cvt_s_d({{ val = Fs.df; }}, ToSingle);
|
|
|
|
0x4: cvt_w_d({{ val = Fs.df; }}, ToWord);
|
|
|
|
0x5: cvt_l_d({{ val = Fs.df; }}, ToLong);
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP1Unimpl::unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-05-09 20:39:45 +02:00
|
|
|
|
|
|
|
0x6: decode FUNCTION_LO {
|
2006-05-10 14:33:52 +02:00
|
|
|
format FloatCompareOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: c_f_d({{ cond = 0; }},
|
|
|
|
DoublePrecision, UnorderedFalse);
|
|
|
|
0x1: c_un_d({{ cond = 0; }},
|
|
|
|
DoublePrecision, UnorderedTrue);
|
2006-06-09 09:57:25 +02:00
|
|
|
0x2: c_eq_d({{ cond = (Fs.df == Ft.df); }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x3: c_ueq_d({{ cond = (Fs.df == Ft.df); }},
|
|
|
|
UnorderedTrue);
|
|
|
|
0x4: c_olt_d({{ cond = (Fs.df < Ft.df); }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x5: c_ult_d({{ cond = (Fs.df < Ft.df); }},
|
|
|
|
UnorderedTrue);
|
|
|
|
0x6: c_ole_d({{ cond = (Fs.df <= Ft.df); }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x7: c_ule_d({{ cond = (Fs.df <= Ft.df); }},
|
|
|
|
UnorderedTrue);
|
2006-05-09 20:39:45 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
0x7: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format FloatCompareOp {
|
|
|
|
0x0: c_sf_d({{ cond = 0; }}, DoublePrecision,
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x1: c_ngle_d({{ cond = 0; }}, DoublePrecision,
|
|
|
|
UnorderedTrue, QnanException);
|
|
|
|
0x2: c_seq_d({{ cond = (Fs.df == Ft.df); }},
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x3: c_ngl_d({{ cond = (Fs.df == Ft.df); }},
|
|
|
|
UnorderedTrue, QnanException);
|
|
|
|
0x4: c_lt_d({{ cond = (Fs.df < Ft.df); }},
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x5: c_nge_d({{ cond = (Fs.df < Ft.df); }},
|
|
|
|
UnorderedTrue, QnanException);
|
|
|
|
0x6: c_le_d({{ cond = (Fs.df <= Ft.df); }},
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x7: c_ngt_d({{ cond = (Fs.df <= Ft.df); }},
|
|
|
|
UnorderedTrue, QnanException);
|
2006-05-09 20:39:45 +02:00
|
|
|
}
|
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP1Unimpl::unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
0x2: CP1Unimpl::unknown();
|
|
|
|
0x3: CP1Unimpl::unknown();
|
|
|
|
0x7: CP1Unimpl::unknown();
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table A-16 MIPS32 COP1 Encoding of Function
|
|
|
|
//Field When rs=W
|
2006-02-18 09:12:04 +01:00
|
|
|
0x4: decode FUNCTION {
|
2006-05-10 22:52:27 +02:00
|
|
|
format FloatConvertOp {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x20: cvt_s_w({{ val = Fs.uw; }}, ToSingle);
|
|
|
|
0x21: cvt_d_w({{ val = Fs.uw; }}, ToDouble);
|
2007-11-13 22:58:16 +01:00
|
|
|
0x26: CP1Unimpl::cvt_ps_w();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP1Unimpl::unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table A-16 MIPS32 COP1 Encoding of Function Field
|
|
|
|
//When rs=L1
|
|
|
|
//Note: "1. Format type L is legal only if 64-bit
|
|
|
|
//floating point operations are enabled."
|
2006-02-18 09:12:04 +01:00
|
|
|
0x5: decode FUNCTION_HI {
|
2006-06-09 09:57:25 +02:00
|
|
|
format FloatConvertOp {
|
|
|
|
0x20: cvt_s_l({{ val = Fs.ud; }}, ToSingle);
|
|
|
|
0x21: cvt_d_l({{ val = Fs.ud; }}, ToDouble);
|
2007-11-13 22:58:16 +01:00
|
|
|
0x26: CP1Unimpl::cvt_ps_l();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP1Unimpl::unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table A-17 MIPS64 COP1 Encoding of Function Field
|
|
|
|
//When rs=PS1
|
|
|
|
//Note: "1. Format type PS is legal only if 64-bit
|
|
|
|
//floating point operations are enabled. "
|
2006-04-26 22:13:47 +02:00
|
|
|
0x6: decode FUNCTION_HI {
|
|
|
|
0x0: decode FUNCTION_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
format Float64Op {
|
2006-05-10 22:52:27 +02:00
|
|
|
0x0: add_ps({{
|
2006-05-11 02:54:03 +02:00
|
|
|
Fd1.sf = Fs1.sf + Ft2.sf;
|
|
|
|
Fd2.sf = Fs2.sf + Ft2.sf;
|
2006-02-18 09:12:04 +01:00
|
|
|
}});
|
2006-05-10 22:52:27 +02:00
|
|
|
0x1: sub_ps({{
|
2006-05-11 02:54:03 +02:00
|
|
|
Fd1.sf = Fs1.sf - Ft2.sf;
|
|
|
|
Fd2.sf = Fs2.sf - Ft2.sf;
|
2006-02-18 09:12:04 +01:00
|
|
|
}});
|
2006-05-10 22:52:27 +02:00
|
|
|
0x2: mul_ps({{
|
2006-05-11 02:54:03 +02:00
|
|
|
Fd1.sf = Fs1.sf * Ft2.sf;
|
|
|
|
Fd2.sf = Fs2.sf * Ft2.sf;
|
2006-02-18 09:12:04 +01:00
|
|
|
}});
|
2006-05-10 22:52:27 +02:00
|
|
|
0x5: abs_ps({{
|
2006-05-11 02:54:03 +02:00
|
|
|
Fd1.sf = fabs(Fs1.sf);
|
|
|
|
Fd2.sf = fabs(Fs2.sf);
|
2006-02-18 09:12:04 +01:00
|
|
|
}});
|
2006-05-10 22:52:27 +02:00
|
|
|
0x6: mov_ps({{
|
2006-05-11 02:54:03 +02:00
|
|
|
Fd1.sf = Fs1.sf;
|
|
|
|
Fd2.sf = Fs2.sf;
|
2006-02-18 09:12:04 +01:00
|
|
|
}});
|
2006-05-10 22:52:27 +02:00
|
|
|
0x7: neg_ps({{
|
2006-06-09 09:57:25 +02:00
|
|
|
Fd1.sf = -(Fs1.sf);
|
|
|
|
Fd2.sf = -(Fs2.sf);
|
2006-02-18 09:12:04 +01:00
|
|
|
}});
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP1Unimpl::unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
0x1: CP1Unimpl::unknown();
|
2006-04-26 22:13:47 +02:00
|
|
|
0x2: decode FUNCTION_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x1: decode MOVCF {
|
|
|
|
format Float64Op {
|
2006-05-11 02:54:03 +02:00
|
|
|
0x0: movf_ps({{
|
2006-06-09 09:57:25 +02:00
|
|
|
Fd1 = (getCondCode(FCSR, CC) == 0) ?
|
|
|
|
Fs1 : Fd1;
|
|
|
|
Fd2 = (getCondCode(FCSR, CC+1) == 0) ?
|
|
|
|
Fs2 : Fd2;
|
2006-05-11 02:54:03 +02:00
|
|
|
}});
|
|
|
|
0x1: movt_ps({{
|
2006-06-09 09:57:25 +02:00
|
|
|
Fd2 = (getCondCode(FCSR, CC) == 1) ?
|
|
|
|
Fs1 : Fd1;
|
|
|
|
Fd2 = (getCondCode(FCSR, CC+1) == 1) ?
|
|
|
|
Fs2 : Fd2;
|
2006-05-11 02:54:03 +02:00
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-05-10 22:52:27 +02:00
|
|
|
format Float64Op {
|
2006-05-11 02:54:03 +02:00
|
|
|
0x2: movz_ps({{
|
2006-06-09 09:57:25 +02:00
|
|
|
Fd1 = (getCondCode(FCSR, CC) == 0) ?
|
|
|
|
Fs1 : Fd1;
|
|
|
|
Fd2 = (getCondCode(FCSR, CC) == 0) ?
|
|
|
|
Fs2 : Fd2;
|
2006-05-11 02:54:03 +02:00
|
|
|
}});
|
|
|
|
0x3: movn_ps({{
|
2006-06-09 09:57:25 +02:00
|
|
|
Fd1 = (getCondCode(FCSR, CC) == 1) ?
|
|
|
|
Fs1 : Fd1;
|
|
|
|
Fd2 = (getCondCode(FCSR, CC) == 1) ?
|
|
|
|
Fs2 : Fd2;
|
2006-05-11 02:54:03 +02:00
|
|
|
}});
|
2006-02-22 09:33:35 +01:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP1Unimpl::unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
0x3: CP1Unimpl::unknown();
|
2006-04-26 22:13:47 +02:00
|
|
|
0x4: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x0: FloatOp::cvt_s_pu({{ Fd.sf = Fs2.sf; }});
|
2007-11-13 22:58:16 +01:00
|
|
|
default: CP1Unimpl::unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-04-26 22:13:47 +02:00
|
|
|
0x5: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x0: FloatOp::cvt_s_pl({{ Fd.sf = Fs1.sf; }});
|
|
|
|
format Float64Op {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: pll({{
|
|
|
|
Fd.ud = (uint64_t)Fs1.uw << 32 | Ft1.uw;
|
|
|
|
}});
|
|
|
|
0x5: plu({{
|
|
|
|
Fd.ud = (uint64_t)Fs1.uw << 32 | Ft2.uw;
|
|
|
|
}});
|
|
|
|
0x6: pul({{
|
|
|
|
Fd.ud = (uint64_t)Fs2.uw << 32 | Ft1.uw;
|
|
|
|
}});
|
|
|
|
0x7: puu({{
|
|
|
|
Fd.ud = (uint64_t)Fs2.uw << 32 | Ft2.uw;
|
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
default: CP1Unimpl::unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-05-09 20:39:45 +02:00
|
|
|
|
|
|
|
0x6: decode FUNCTION_LO {
|
2006-05-11 09:26:19 +02:00
|
|
|
format FloatPSCompareOp {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x0: c_f_ps({{ cond1 = 0; }}, {{ cond2 = 0; }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x1: c_un_ps({{ cond1 = 0; }}, {{ cond2 = 0; }},
|
|
|
|
UnorderedTrue);
|
|
|
|
0x2: c_eq_ps({{ cond1 = (Fs1.sf == Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf == Ft2.sf); }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x3: c_ueq_ps({{ cond1 = (Fs1.sf == Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf == Ft2.sf); }},
|
|
|
|
UnorderedTrue);
|
|
|
|
0x4: c_olt_ps({{ cond1 = (Fs1.sf < Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf < Ft2.sf); }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x5: c_ult_ps({{ cond1 = (Fs.sf < Ft.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf < Ft2.sf); }},
|
|
|
|
UnorderedTrue);
|
|
|
|
0x6: c_ole_ps({{ cond1 = (Fs.sf <= Ft.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf <= Ft2.sf); }},
|
|
|
|
UnorderedFalse);
|
|
|
|
0x7: c_ule_ps({{ cond1 = (Fs1.sf <= Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf <= Ft2.sf); }},
|
|
|
|
UnorderedTrue);
|
2006-05-09 20:39:45 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
0x7: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format FloatPSCompareOp {
|
|
|
|
0x0: c_sf_ps({{ cond1 = 0; }}, {{ cond2 = 0; }},
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x1: c_ngle_ps({{ cond1 = 0; }},
|
|
|
|
{{ cond2 = 0; }},
|
|
|
|
UnorderedTrue, QnanException);
|
|
|
|
0x2: c_seq_ps({{ cond1 = (Fs1.sf == Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf == Ft2.sf); }},
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x3: c_ngl_ps({{ cond1 = (Fs1.sf == Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf == Ft2.sf); }},
|
|
|
|
UnorderedTrue, QnanException);
|
|
|
|
0x4: c_lt_ps({{ cond1 = (Fs1.sf < Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf < Ft2.sf); }},
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x5: c_nge_ps({{ cond1 = (Fs1.sf < Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf < Ft2.sf); }},
|
|
|
|
UnorderedTrue, QnanException);
|
|
|
|
0x6: c_le_ps({{ cond1 = (Fs1.sf <= Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf <= Ft2.sf); }},
|
|
|
|
UnorderedFalse, QnanException);
|
|
|
|
0x7: c_ngt_ps({{ cond1 = (Fs1.sf <= Ft1.sf); }},
|
|
|
|
{{ cond2 = (Fs2.sf <= Ft2.sf); }},
|
|
|
|
UnorderedTrue, QnanException);
|
2006-05-09 20:39:45 +02:00
|
|
|
}
|
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default: CP1Unimpl::unknown();
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
//Table A-19 MIPS32 COP2 Encoding of rs Field
|
|
|
|
0x2: decode RS_MSB {
|
2007-11-13 22:58:16 +01:00
|
|
|
format CP2Unimpl {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x0: decode RS_HI {
|
|
|
|
0x0: decode RS_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x0: mfc2();
|
|
|
|
0x2: cfc2();
|
|
|
|
0x3: mfhc2();
|
|
|
|
0x4: mtc2();
|
|
|
|
0x6: ctc2();
|
|
|
|
0x7: mftc2();
|
2009-07-22 10:51:10 +02:00
|
|
|
default: unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x1: decode ND {
|
|
|
|
0x0: decode TF {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x0: bc2f();
|
|
|
|
0x1: bc2t();
|
2009-07-22 10:51:10 +02:00
|
|
|
default: unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x1: decode TF {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x0: bc2fl();
|
|
|
|
0x1: bc2tl();
|
2009-07-22 10:51:10 +02:00
|
|
|
default: unknown();
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
default: unknown();
|
2007-11-13 22:58:16 +01:00
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
}
|
|
|
|
default: unknown();
|
|
|
|
}
|
|
|
|
default: unknown();
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
//Table A-20 MIPS64 COP1X Encoding of Function Field 1
|
|
|
|
//Note: "COP1X instructions are legal only if 64-bit floating point
|
|
|
|
//operations are enabled."
|
|
|
|
0x3: decode FUNCTION_HI {
|
|
|
|
0x0: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format LoadIndexedMemory {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: lwxc1({{ Fd.uw = Mem.uw; }});
|
|
|
|
0x1: ldxc1({{ Fd.ud = Mem.ud; }});
|
|
|
|
0x5: luxc1({{ Fd.ud = Mem.ud; }},
|
2006-06-11 21:38:40 +02:00
|
|
|
{{ EA = (Rs + Rt) & ~7; }});
|
2006-02-08 20:50:07 +01:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
0x1: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format StoreIndexedMemory {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: swxc1({{ Mem.uw = Fs.uw; }});
|
|
|
|
0x1: sdxc1({{ Mem.ud = Fs.ud; }});
|
|
|
|
0x5: suxc1({{ Mem.ud = Fs.ud; }},
|
2006-06-11 21:38:40 +02:00
|
|
|
{{ EA = (Rs + Rt) & ~7; }});
|
2006-02-08 20:50:07 +01:00
|
|
|
}
|
2006-06-09 09:57:25 +02:00
|
|
|
0x7: Prefetch::prefx({{ EA = Rs + Rt; }});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x3: decode FUNCTION_LO {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x6: Float64Op::alnv_ps({{
|
|
|
|
if (Rs<2:0> == 0) {
|
|
|
|
Fd.ud = Fs.ud;
|
|
|
|
} else if (Rs<2:0> == 4) {
|
|
|
|
#if BYTE_ORDER == BIG_ENDIAN
|
|
|
|
Fd.ud = Fs.ud<31:0> << 32 | Ft.ud<63:32>;
|
|
|
|
#elif BYTE_ORDER == LITTLE_ENDIAN
|
|
|
|
Fd.ud = Ft.ud<31:0> << 32 | Fs.ud<63:32>;
|
|
|
|
#endif
|
|
|
|
} else {
|
|
|
|
Fd.ud = Fd.ud;
|
|
|
|
}
|
|
|
|
}});
|
2006-06-09 09:57:25 +02:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
format FloatAccOp {
|
|
|
|
0x4: decode FUNCTION_LO {
|
|
|
|
0x0: madd_s({{ Fd.sf = (Fs.sf * Ft.sf) + Fr.sf; }});
|
|
|
|
0x1: madd_d({{ Fd.df = (Fs.df * Ft.df) + Fr.df; }});
|
|
|
|
0x6: madd_ps({{
|
|
|
|
Fd1.sf = (Fs1.df * Ft1.df) + Fr1.df;
|
|
|
|
Fd2.sf = (Fs2.df * Ft2.df) + Fr2.df;
|
|
|
|
}});
|
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x5: decode FUNCTION_LO {
|
|
|
|
0x0: msub_s({{ Fd.sf = (Fs.sf * Ft.sf) - Fr.sf; }});
|
|
|
|
0x1: msub_d({{ Fd.df = (Fs.df * Ft.df) - Fr.df; }});
|
|
|
|
0x6: msub_ps({{
|
|
|
|
Fd1.sf = (Fs1.df * Ft1.df) - Fr1.df;
|
|
|
|
Fd2.sf = (Fs2.df * Ft2.df) - Fr2.df;
|
|
|
|
}});
|
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x6: decode FUNCTION_LO {
|
|
|
|
0x0: nmadd_s({{ Fd.sf = (-1 * Fs.sf * Ft.sf) - Fr.sf; }});
|
|
|
|
0x1: nmadd_d({{ Fd.df = (-1 * Fs.df * Ft.df) + Fr.df; }});
|
|
|
|
0x6: nmadd_ps({{
|
|
|
|
Fd1.sf = -((Fs1.df * Ft1.df) + Fr1.df);
|
|
|
|
Fd2.sf = -((Fs2.df * Ft2.df) + Fr2.df);
|
|
|
|
}});
|
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x7: decode FUNCTION_LO {
|
|
|
|
0x0: nmsub_s({{ Fd.sf = (-1 * Fs.sf * Ft.sf) - Fr.sf; }});
|
|
|
|
0x1: nmsub_d({{ Fd.df = (-1 * Fs.df * Ft.df) - Fr.df; }});
|
|
|
|
0x6: nmsub_ps({{
|
|
|
|
Fd1.sf = -((Fs1.df * Ft1.df) - Fr1.df);
|
|
|
|
Fd2.sf = -((Fs2.df * Ft2.df) - Fr2.df);
|
|
|
|
}});
|
2006-02-08 20:50:07 +01:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
format Branch {
|
|
|
|
0x4: beql({{ cond = (Rs.sw == Rt.sw); }}, Likely);
|
|
|
|
0x5: bnel({{ cond = (Rs.sw != Rt.sw); }}, Likely);
|
|
|
|
0x6: blezl({{ cond = (Rs.sw <= 0); }}, Likely);
|
|
|
|
0x7: bgtzl({{ cond = (Rs.sw > 0); }}, Likely);
|
2006-02-08 20:50:07 +01:00
|
|
|
}
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x3: decode OPCODE_LO {
|
2006-02-08 00:36:08 +01:00
|
|
|
//Table A-5 MIPS32 SPECIAL2 Encoding of Function Field
|
|
|
|
0x4: decode FUNCTION_HI {
|
|
|
|
0x0: decode FUNCTION_LO {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: IntOp::mul({{
|
|
|
|
int64_t temp1 = Rs.sd * Rt.sd;
|
|
|
|
Rd.sw = temp1<31:0>;
|
|
|
|
}}, IntMultOp);
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2007-06-23 01:03:42 +02:00
|
|
|
format HiLoRdSelValOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: madd({{
|
|
|
|
val = ((int64_t)HI_RD_SEL << 32 | LO_RD_SEL) +
|
|
|
|
(Rs.sd * Rt.sd);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x1: maddu({{
|
|
|
|
val = ((uint64_t)HI_RD_SEL << 32 | LO_RD_SEL) +
|
|
|
|
(Rs.ud * Rt.ud);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x4: msub({{
|
|
|
|
val = ((int64_t)HI_RD_SEL << 32 | LO_RD_SEL) -
|
|
|
|
(Rs.sd * Rt.sd);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x5: msubu({{
|
|
|
|
val = ((uint64_t)HI_RD_SEL << 32 | LO_RD_SEL) -
|
|
|
|
(Rs.ud * Rt.ud);
|
|
|
|
}}, IntMultOp);
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
0x4: decode FUNCTION_LO {
|
2006-02-18 09:12:04 +01:00
|
|
|
format BasicOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: clz({{
|
|
|
|
int cnt = 32;
|
|
|
|
for (int idx = 31; idx >= 0; idx--) {
|
|
|
|
if (Rs<idx:idx> == 1) {
|
|
|
|
cnt = 31 - idx;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
Rd.uw = cnt;
|
|
|
|
}});
|
|
|
|
0x1: clo({{
|
|
|
|
int cnt = 32;
|
|
|
|
for (int idx = 31; idx >= 0; idx--) {
|
|
|
|
if (Rs<idx:idx> == 0) {
|
|
|
|
cnt = 31 - idx;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
Rd.uw = cnt;
|
|
|
|
}});
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
0x7: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
0x7: FailUnimpl::sdbbp();
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
//Table A-6 MIPS32 SPECIAL3 Encoding of Function Field for Release 2
|
|
|
|
//of the Architecture
|
2006-02-08 00:36:08 +01:00
|
|
|
0x7: decode FUNCTION_HI {
|
2006-02-18 09:12:04 +01:00
|
|
|
0x0: decode FUNCTION_LO {
|
2006-06-09 09:57:25 +02:00
|
|
|
format BasicOp {
|
2006-06-11 21:38:40 +02:00
|
|
|
0x0: ext({{ Rt.uw = bits(Rs.uw, MSB+LSB, LSB); }});
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: ins({{
|
|
|
|
Rt.uw = bits(Rt.uw, 31, MSB+1) << (MSB+1) |
|
|
|
|
bits(Rs.uw, MSB-LSB, 0) << LSB |
|
|
|
|
bits(Rt.uw, LSB-1, 0);
|
|
|
|
}});
|
2006-02-08 20:50:07 +01:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
0x1: decode FUNCTION_LO {
|
2007-06-23 01:03:42 +02:00
|
|
|
format MT_Control {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: fork({{
|
|
|
|
forkThread(xc->tcBase(), fault, RD, Rs, Rt);
|
|
|
|
}}, UserMode);
|
|
|
|
0x1: yield({{
|
|
|
|
Rd.sw = yieldThread(xc->tcBase(), fault, Rs.sw,
|
|
|
|
YQMask);
|
|
|
|
}}, UserMode);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
//Table 5-9 MIPS32 LX Encoding of the op Field (DSP ASE MANUAL)
|
|
|
|
0x2: decode OP_HI {
|
|
|
|
0x0: decode OP_LO {
|
|
|
|
format LoadIndexedMemory {
|
|
|
|
0x0: lwx({{ Rd.sw = Mem.sw; }});
|
|
|
|
0x4: lhx({{ Rd.sw = Mem.sh; }});
|
|
|
|
0x6: lbux({{ Rd.uw = Mem.ub; }});
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: DspIntOp::insv({{
|
|
|
|
int pos = dspctl<5:0>;
|
|
|
|
int size = dspctl<12:7> - 1;
|
|
|
|
Rt.uw = insertBits(Rt.uw, pos+size,
|
|
|
|
pos, Rs.uw<size:0>);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
0x2: decode FUNCTION_LO {
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table 5-5 MIPS32 ADDU.QB Encoding of the op Field
|
|
|
|
//(DSP ASE MANUAL)
|
2007-06-23 01:03:42 +02:00
|
|
|
0x0: decode OP_HI {
|
|
|
|
0x0: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: addu_qb({{
|
|
|
|
Rd.uw = dspAdd(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
NOSATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x1: subu_qb({{
|
|
|
|
Rd.uw = dspSub(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
NOSATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x4: addu_s_qb({{
|
|
|
|
Rd.uw = dspAdd(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
SATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x5: subu_s_qb({{
|
|
|
|
Rd.uw = dspSub(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
SATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x6: muleu_s_ph_qbl({{
|
|
|
|
Rd.uw = dspMuleu(Rs.uw, Rt.uw,
|
|
|
|
MODE_L, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x7: muleu_s_ph_qbr({{
|
|
|
|
Rd.uw = dspMuleu(Rs.uw, Rt.uw,
|
|
|
|
MODE_R, &dspctl);
|
|
|
|
}}, IntMultOp);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x1: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: addu_ph({{
|
|
|
|
Rd.uw = dspAdd(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
NOSATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x1: subu_ph({{
|
|
|
|
Rd.uw = dspSub(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
NOSATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x2: addq_ph({{
|
|
|
|
Rd.uw = dspAdd(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
NOSATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x3: subq_ph({{
|
|
|
|
Rd.uw = dspSub(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
NOSATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x4: addu_s_ph({{
|
|
|
|
Rd.uw = dspAdd(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
SATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x5: subu_s_ph({{
|
|
|
|
Rd.uw = dspSub(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
SATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x6: addq_s_ph({{
|
|
|
|
Rd.uw = dspAdd(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
SATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x7: subq_s_ph({{
|
|
|
|
Rd.uw = dspSub(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
SATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x2: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: addsc({{
|
|
|
|
int64_t dresult;
|
|
|
|
dresult = Rs.ud + Rt.ud;
|
|
|
|
Rd.sw = dresult<31:0>;
|
|
|
|
dspctl = insertBits(dspctl, 13, 13,
|
|
|
|
dresult<32:32>);
|
|
|
|
}});
|
|
|
|
0x1: addwc({{
|
|
|
|
int64_t dresult;
|
|
|
|
dresult = Rs.sd + Rt.sd + dspctl<13:13>;
|
|
|
|
Rd.sw = dresult<31:0>;
|
|
|
|
if (dresult<32:32> != dresult<31:31>)
|
|
|
|
dspctl = insertBits(dspctl, 20, 20, 1);
|
|
|
|
}});
|
|
|
|
0x2: modsub({{
|
|
|
|
Rd.sw = (Rs.sw == 0) ? Rt.sw<23:8> :
|
|
|
|
Rs.sw - Rt.sw<7:0>;
|
|
|
|
}});
|
|
|
|
0x4: raddu_w_qb({{
|
|
|
|
Rd.uw = Rs.uw<31:24> + Rs.uw<23:16> +
|
|
|
|
Rs.uw<15:8> + Rs.uw<7:0>;
|
|
|
|
}});
|
|
|
|
0x6: addq_s_w({{
|
|
|
|
Rd.sw = dspAdd(Rs.sw, Rt.sw, SIMD_FMT_W,
|
|
|
|
SATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x7: subq_s_w({{
|
|
|
|
Rd.sw = dspSub(Rs.sw, Rt.sw, SIMD_FMT_W,
|
|
|
|
SATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x3: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: muleq_s_w_phl({{
|
|
|
|
Rd.sw = dspMuleq(Rs.sw, Rt.sw,
|
|
|
|
MODE_L, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x5: muleq_s_w_phr({{
|
|
|
|
Rd.sw = dspMuleq(Rs.sw, Rt.sw,
|
|
|
|
MODE_R, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x6: mulq_s_ph({{
|
|
|
|
Rd.sw = dspMulq(Rs.sw, Rt.sw, SIMD_FMT_PH,
|
|
|
|
SATURATE, NOROUND, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x7: mulq_rs_ph({{
|
|
|
|
Rd.sw = dspMulq(Rs.sw, Rt.sw, SIMD_FMT_PH,
|
|
|
|
SATURATE, ROUND, &dspctl);
|
|
|
|
}}, IntMultOp);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table 5-6 MIPS32 CMPU_EQ_QB Encoding of the op Field
|
|
|
|
//(DSP ASE MANUAL)
|
2007-06-23 01:03:42 +02:00
|
|
|
0x1: decode OP_HI {
|
|
|
|
0x0: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: cmpu_eq_qb({{
|
|
|
|
dspCmp(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, CMP_EQ, &dspctl);
|
|
|
|
}});
|
|
|
|
0x1: cmpu_lt_qb({{
|
|
|
|
dspCmp(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, CMP_LT, &dspctl);
|
|
|
|
}});
|
|
|
|
0x2: cmpu_le_qb({{
|
|
|
|
dspCmp(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, CMP_LE, &dspctl);
|
|
|
|
}});
|
|
|
|
0x3: pick_qb({{
|
|
|
|
Rd.uw = dspPick(Rs.uw, Rt.uw,
|
|
|
|
SIMD_FMT_QB, &dspctl);
|
|
|
|
}});
|
|
|
|
0x4: cmpgu_eq_qb({{
|
|
|
|
Rd.uw = dspCmpg(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, CMP_EQ );
|
|
|
|
}});
|
|
|
|
0x5: cmpgu_lt_qb({{
|
|
|
|
Rd.uw = dspCmpg(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, CMP_LT);
|
|
|
|
}});
|
|
|
|
0x6: cmpgu_le_qb({{
|
|
|
|
Rd.uw = dspCmpg(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, CMP_LE);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x1: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: cmp_eq_ph({{
|
|
|
|
dspCmp(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
SIGNED, CMP_EQ, &dspctl);
|
|
|
|
}});
|
|
|
|
0x1: cmp_lt_ph({{
|
|
|
|
dspCmp(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
SIGNED, CMP_LT, &dspctl);
|
|
|
|
}});
|
|
|
|
0x2: cmp_le_ph({{
|
|
|
|
dspCmp(Rs.uw, Rt.uw, SIMD_FMT_PH,
|
|
|
|
SIGNED, CMP_LE, &dspctl);
|
|
|
|
}});
|
|
|
|
0x3: pick_ph({{
|
|
|
|
Rd.uw = dspPick(Rs.uw, Rt.uw,
|
|
|
|
SIMD_FMT_PH, &dspctl);
|
|
|
|
}});
|
|
|
|
0x4: precrq_qb_ph({{
|
|
|
|
Rd.uw = Rs.uw<31:24> << 24 |
|
|
|
|
Rs.uw<15:8> << 16 |
|
|
|
|
Rt.uw<31:24> << 8 |
|
|
|
|
Rt.uw<15:8>;
|
|
|
|
}});
|
|
|
|
0x5: precr_qb_ph({{
|
|
|
|
Rd.uw = Rs.uw<23:16> << 24 |
|
|
|
|
Rs.uw<7:0> << 16 |
|
|
|
|
Rt.uw<23:16> << 8 |
|
|
|
|
Rt.uw<7:0>;
|
|
|
|
}});
|
|
|
|
0x6: packrl_ph({{
|
|
|
|
Rd.uw = dspPack(Rs.uw, Rt.uw, SIMD_FMT_PH);
|
|
|
|
}});
|
|
|
|
0x7: precrqu_s_qb_ph({{
|
|
|
|
Rd.uw = dspPrecrqu(Rs.uw, Rt.uw, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x2: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: precrq_ph_w({{
|
|
|
|
Rd.uw = Rs.uw<31:16> << 16 | Rt.uw<31:16>;
|
|
|
|
}});
|
|
|
|
0x5: precrq_rs_ph_w({{
|
|
|
|
Rd.uw = dspPrecrq(Rs.uw, Rt.uw,
|
|
|
|
SIMD_FMT_W, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x3: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: cmpgdu_eq_qb({{
|
|
|
|
Rd.uw = dspCmpgd(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, CMP_EQ, &dspctl);
|
|
|
|
}});
|
|
|
|
0x1: cmpgdu_lt_qb({{
|
|
|
|
Rd.uw = dspCmpgd(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, CMP_LT, &dspctl);
|
|
|
|
}});
|
|
|
|
0x2: cmpgdu_le_qb({{
|
|
|
|
Rd.uw = dspCmpgd(Rs.uw, Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, CMP_LE, &dspctl);
|
|
|
|
}});
|
|
|
|
0x6: precr_sra_ph_w({{
|
|
|
|
Rt.uw = dspPrecrSra(Rt.uw, Rs.uw, RD,
|
|
|
|
SIMD_FMT_W, NOROUND);
|
|
|
|
}});
|
|
|
|
0x7: precr_sra_r_ph_w({{
|
|
|
|
Rt.uw = dspPrecrSra(Rt.uw, Rs.uw, RD,
|
|
|
|
SIMD_FMT_W, ROUND);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table 5-7 MIPS32 ABSQ_S.PH Encoding of the op Field
|
|
|
|
//(DSP ASE MANUAL)
|
2007-06-23 01:03:42 +02:00
|
|
|
0x2: decode OP_HI {
|
|
|
|
0x0: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x1: absq_s_qb({{
|
|
|
|
Rd.sw = dspAbs(Rt.sw, SIMD_FMT_QB, &dspctl);
|
|
|
|
}});
|
|
|
|
0x2: repl_qb({{
|
|
|
|
Rd.uw = RS_RT<7:0> << 24 |
|
|
|
|
RS_RT<7:0> << 16 |
|
|
|
|
RS_RT<7:0> << 8 |
|
|
|
|
RS_RT<7:0>;
|
|
|
|
}});
|
|
|
|
0x3: replv_qb({{
|
|
|
|
Rd.sw = Rt.uw<7:0> << 24 |
|
|
|
|
Rt.uw<7:0> << 16 |
|
|
|
|
Rt.uw<7:0> << 8 |
|
|
|
|
Rt.uw<7:0>;
|
|
|
|
}});
|
|
|
|
0x4: precequ_ph_qbl({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_QB, UNSIGNED,
|
|
|
|
SIMD_FMT_PH, SIGNED, MODE_L);
|
|
|
|
}});
|
|
|
|
0x5: precequ_ph_qbr({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_QB, UNSIGNED,
|
|
|
|
SIMD_FMT_PH, SIGNED, MODE_R);
|
|
|
|
}});
|
|
|
|
0x6: precequ_ph_qbla({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_QB, UNSIGNED,
|
|
|
|
SIMD_FMT_PH, SIGNED, MODE_LA);
|
|
|
|
}});
|
|
|
|
0x7: precequ_ph_qbra({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_QB, UNSIGNED,
|
|
|
|
SIMD_FMT_PH, SIGNED, MODE_RA);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x1: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x1: absq_s_ph({{
|
|
|
|
Rd.sw = dspAbs(Rt.sw, SIMD_FMT_PH, &dspctl);
|
|
|
|
}});
|
|
|
|
0x2: repl_ph({{
|
|
|
|
Rd.uw = (sext<10>(RS_RT))<15:0> << 16 |
|
|
|
|
(sext<10>(RS_RT))<15:0>;
|
|
|
|
}});
|
|
|
|
0x3: replv_ph({{
|
|
|
|
Rd.uw = Rt.uw<15:0> << 16 |
|
|
|
|
Rt.uw<15:0>;
|
|
|
|
}});
|
|
|
|
0x4: preceq_w_phl({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_PH, SIGNED,
|
|
|
|
SIMD_FMT_W, SIGNED, MODE_L);
|
|
|
|
}});
|
|
|
|
0x5: preceq_w_phr({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_PH, SIGNED,
|
|
|
|
SIMD_FMT_W, SIGNED, MODE_R);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x2: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x1: absq_s_w({{
|
|
|
|
Rd.sw = dspAbs(Rt.sw, SIMD_FMT_W, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x3: decode OP_LO {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x3: IntOp::bitrev({{
|
|
|
|
Rd.uw = bitrev( Rt.uw<15:0> );
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: preceu_ph_qbl({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, SIMD_FMT_PH,
|
|
|
|
UNSIGNED, MODE_L);
|
|
|
|
}});
|
|
|
|
0x5: preceu_ph_qbr({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, SIMD_FMT_PH,
|
|
|
|
UNSIGNED, MODE_R );
|
|
|
|
}});
|
|
|
|
0x6: preceu_ph_qbla({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, SIMD_FMT_PH,
|
|
|
|
UNSIGNED, MODE_LA );
|
|
|
|
}});
|
|
|
|
0x7: preceu_ph_qbra({{
|
|
|
|
Rd.uw = dspPrece(Rt.uw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED, SIMD_FMT_PH,
|
|
|
|
UNSIGNED, MODE_RA);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table 5-8 MIPS32 SHLL.QB Encoding of the op Field
|
|
|
|
//(DSP ASE MANUAL)
|
2007-06-23 01:03:42 +02:00
|
|
|
0x3: decode OP_HI {
|
|
|
|
0x0: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: shll_qb({{
|
|
|
|
Rd.sw = dspShll(Rt.sw, RS, SIMD_FMT_QB,
|
|
|
|
NOSATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x1: shrl_qb({{
|
|
|
|
Rd.sw = dspShrl(Rt.sw, RS, SIMD_FMT_QB,
|
|
|
|
UNSIGNED);
|
|
|
|
}});
|
|
|
|
0x2: shllv_qb({{
|
|
|
|
Rd.sw = dspShll(Rt.sw, Rs.sw, SIMD_FMT_QB,
|
|
|
|
NOSATURATE, UNSIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x3: shrlv_qb({{
|
|
|
|
Rd.sw = dspShrl(Rt.sw, Rs.sw, SIMD_FMT_QB,
|
|
|
|
UNSIGNED);
|
|
|
|
}});
|
|
|
|
0x4: shra_qb({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, RS, SIMD_FMT_QB,
|
|
|
|
NOROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x5: shra_r_qb({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, RS, SIMD_FMT_QB,
|
|
|
|
ROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x6: shrav_qb({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, Rs.sw, SIMD_FMT_QB,
|
|
|
|
NOROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x7: shrav_r_qb({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, Rs.sw, SIMD_FMT_QB,
|
|
|
|
ROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x1: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: shll_ph({{
|
|
|
|
Rd.uw = dspShll(Rt.uw, RS, SIMD_FMT_PH,
|
|
|
|
NOSATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x1: shra_ph({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, RS, SIMD_FMT_PH,
|
|
|
|
NOROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x2: shllv_ph({{
|
|
|
|
Rd.sw = dspShll(Rt.sw, Rs.sw, SIMD_FMT_PH,
|
|
|
|
NOSATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x3: shrav_ph({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, Rs.sw, SIMD_FMT_PH,
|
|
|
|
NOROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x4: shll_s_ph({{
|
|
|
|
Rd.sw = dspShll(Rt.sw, RS, SIMD_FMT_PH,
|
|
|
|
SATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x5: shra_r_ph({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, RS, SIMD_FMT_PH,
|
|
|
|
ROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x6: shllv_s_ph({{
|
|
|
|
Rd.sw = dspShll(Rt.sw, Rs.sw, SIMD_FMT_PH,
|
|
|
|
SATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x7: shrav_r_ph({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, Rs.sw, SIMD_FMT_PH,
|
|
|
|
ROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x2: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x4: shll_s_w({{
|
|
|
|
Rd.sw = dspShll(Rt.sw, RS, SIMD_FMT_W,
|
|
|
|
SATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x5: shra_r_w({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, RS, SIMD_FMT_W,
|
|
|
|
ROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x6: shllv_s_w({{
|
|
|
|
Rd.sw = dspShll(Rt.sw, Rs.sw, SIMD_FMT_W,
|
|
|
|
SATURATE, SIGNED, &dspctl);
|
|
|
|
}});
|
|
|
|
0x7: shrav_r_w({{
|
|
|
|
Rd.sw = dspShra(Rt.sw, Rs.sw, SIMD_FMT_W,
|
|
|
|
ROUND, SIGNED, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x3: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x1: shrl_ph({{
|
|
|
|
Rd.sw = dspShrl(Rt.sw, RS, SIMD_FMT_PH,
|
|
|
|
UNSIGNED);
|
|
|
|
}});
|
|
|
|
0x3: shrlv_ph({{
|
|
|
|
Rd.sw = dspShrl(Rt.sw, Rs.sw, SIMD_FMT_PH,
|
|
|
|
UNSIGNED);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
0x3: decode FUNCTION_LO {
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table 3.12 MIPS32 ADDUH.QB Encoding of the op Field
|
|
|
|
//(DSP ASE Rev2 Manual)
|
2007-06-23 01:03:42 +02:00
|
|
|
0x0: decode OP_HI {
|
|
|
|
0x0: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: adduh_qb({{
|
|
|
|
Rd.uw = dspAddh(Rs.sw, Rt.sw, SIMD_FMT_QB,
|
|
|
|
NOROUND, UNSIGNED);
|
|
|
|
}});
|
|
|
|
0x1: subuh_qb({{
|
|
|
|
Rd.uw = dspSubh(Rs.sw, Rt.sw, SIMD_FMT_QB,
|
|
|
|
NOROUND, UNSIGNED);
|
|
|
|
}});
|
|
|
|
0x2: adduh_r_qb({{
|
|
|
|
Rd.uw = dspAddh(Rs.sw, Rt.sw, SIMD_FMT_QB,
|
|
|
|
ROUND, UNSIGNED);
|
|
|
|
}});
|
|
|
|
0x3: subuh_r_qb({{
|
|
|
|
Rd.uw = dspSubh(Rs.sw, Rt.sw, SIMD_FMT_QB,
|
|
|
|
ROUND, UNSIGNED);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x1: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: addqh_ph({{
|
|
|
|
Rd.uw = dspAddh(Rs.sw, Rt.sw, SIMD_FMT_PH,
|
|
|
|
NOROUND, SIGNED);
|
|
|
|
}});
|
|
|
|
0x1: subqh_ph({{
|
|
|
|
Rd.uw = dspSubh(Rs.sw, Rt.sw, SIMD_FMT_PH,
|
|
|
|
NOROUND, SIGNED);
|
|
|
|
}});
|
|
|
|
0x2: addqh_r_ph({{
|
|
|
|
Rd.uw = dspAddh(Rs.sw, Rt.sw, SIMD_FMT_PH,
|
|
|
|
ROUND, SIGNED);
|
|
|
|
}});
|
|
|
|
0x3: subqh_r_ph({{
|
|
|
|
Rd.uw = dspSubh(Rs.sw, Rt.sw, SIMD_FMT_PH,
|
|
|
|
ROUND, SIGNED);
|
|
|
|
}});
|
|
|
|
0x4: mul_ph({{
|
|
|
|
Rd.sw = dspMul(Rs.sw, Rt.sw, SIMD_FMT_PH,
|
|
|
|
NOSATURATE, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x6: mul_s_ph({{
|
|
|
|
Rd.sw = dspMul(Rs.sw, Rt.sw, SIMD_FMT_PH,
|
|
|
|
SATURATE, &dspctl);
|
|
|
|
}}, IntMultOp);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x2: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: addqh_w({{
|
|
|
|
Rd.uw = dspAddh(Rs.sw, Rt.sw, SIMD_FMT_W,
|
|
|
|
NOROUND, SIGNED);
|
|
|
|
}});
|
|
|
|
0x1: subqh_w({{
|
|
|
|
Rd.uw = dspSubh(Rs.sw, Rt.sw, SIMD_FMT_W,
|
|
|
|
NOROUND, SIGNED);
|
|
|
|
}});
|
|
|
|
0x2: addqh_r_w({{
|
|
|
|
Rd.uw = dspAddh(Rs.sw, Rt.sw, SIMD_FMT_W,
|
|
|
|
ROUND, SIGNED);
|
|
|
|
}});
|
|
|
|
0x3: subqh_r_w({{
|
|
|
|
Rd.uw = dspSubh(Rs.sw, Rt.sw, SIMD_FMT_W,
|
|
|
|
ROUND, SIGNED);
|
|
|
|
}});
|
|
|
|
0x6: mulq_s_w({{
|
|
|
|
Rd.sw = dspMulq(Rs.sw, Rt.sw, SIMD_FMT_W,
|
|
|
|
SATURATE, NOROUND, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x7: mulq_rs_w({{
|
|
|
|
Rd.sw = dspMulq(Rs.sw, Rt.sw, SIMD_FMT_W,
|
|
|
|
SATURATE, ROUND, &dspctl);
|
|
|
|
}}, IntMultOp);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
2006-02-14 08:03:14 +01:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-14 08:03:14 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
//Table A-10 MIPS32 BSHFL Encoding of sa Field
|
|
|
|
0x4: decode SA {
|
2006-02-08 20:50:07 +01:00
|
|
|
format BasicOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x02: wsbh({{
|
|
|
|
Rd.uw = Rt.uw<23:16> << 24 |
|
|
|
|
Rt.uw<31:24> << 16 |
|
|
|
|
Rt.uw<7:0> << 8 |
|
|
|
|
Rt.uw<15:8>;
|
2006-06-09 09:57:25 +02:00
|
|
|
}});
|
2006-06-11 21:38:40 +02:00
|
|
|
0x10: seb({{ Rd.sw = Rt.sb; }});
|
|
|
|
0x18: seh({{ Rd.sw = Rt.sh; }});
|
2006-02-08 20:50:07 +01:00
|
|
|
}
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
0x6: decode FUNCTION_LO {
|
2007-06-23 01:03:42 +02:00
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table 5-10 MIPS32 DPAQ.W.PH Encoding of the op Field
|
|
|
|
//(DSP ASE MANUAL)
|
2007-06-23 01:03:42 +02:00
|
|
|
0x0: decode OP_HI {
|
|
|
|
0x0: decode OP_LO {
|
|
|
|
format DspHiLoOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: dpa_w_ph({{
|
|
|
|
dspac = dspDpa(dspac, Rs.sw, Rt.sw, ACDST,
|
|
|
|
SIMD_FMT_PH, SIGNED, MODE_L);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x1: dps_w_ph({{
|
|
|
|
dspac = dspDps(dspac, Rs.sw, Rt.sw, ACDST,
|
|
|
|
SIMD_FMT_PH, SIGNED, MODE_L);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x2: mulsa_w_ph({{
|
|
|
|
dspac = dspMulsa(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_PH );
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x3: dpau_h_qbl({{
|
|
|
|
dspac = dspDpa(dspac, Rs.sw, Rt.sw, ACDST,
|
|
|
|
SIMD_FMT_QB, UNSIGNED, MODE_L);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x4: dpaq_s_w_ph({{
|
|
|
|
dspac = dspDpaq(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
SIMD_FMT_W, NOSATURATE,
|
|
|
|
MODE_L, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x5: dpsq_s_w_ph({{
|
|
|
|
dspac = dspDpsq(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
SIMD_FMT_W, NOSATURATE,
|
|
|
|
MODE_L, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x6: mulsaq_s_w_ph({{
|
|
|
|
dspac = dspMulsaq(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
&dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x7: dpau_h_qbr({{
|
|
|
|
dspac = dspDpa(dspac, Rs.sw, Rt.sw, ACDST,
|
|
|
|
SIMD_FMT_QB, UNSIGNED, MODE_R);
|
|
|
|
}}, IntMultOp);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x1: decode OP_LO {
|
|
|
|
format DspHiLoOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: dpax_w_ph({{
|
|
|
|
dspac = dspDpa(dspac, Rs.sw, Rt.sw, ACDST,
|
|
|
|
SIMD_FMT_PH, SIGNED, MODE_X);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x1: dpsx_w_ph({{
|
|
|
|
dspac = dspDps(dspac, Rs.sw, Rt.sw, ACDST,
|
|
|
|
SIMD_FMT_PH, SIGNED, MODE_X);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x3: dpsu_h_qbl({{
|
|
|
|
dspac = dspDps(dspac, Rs.sw, Rt.sw, ACDST,
|
|
|
|
SIMD_FMT_QB, UNSIGNED, MODE_L);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x4: dpaq_sa_l_w({{
|
|
|
|
dspac = dspDpaq(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_W,
|
|
|
|
SIMD_FMT_L, SATURATE,
|
|
|
|
MODE_L, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x5: dpsq_sa_l_w({{
|
|
|
|
dspac = dspDpsq(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_W,
|
|
|
|
SIMD_FMT_L, SATURATE,
|
|
|
|
MODE_L, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x7: dpsu_h_qbr({{
|
|
|
|
dspac = dspDps(dspac, Rs.sw, Rt.sw, ACDST,
|
|
|
|
SIMD_FMT_QB, UNSIGNED, MODE_R);
|
|
|
|
}}, IntMultOp);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x2: decode OP_LO {
|
|
|
|
format DspHiLoOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: maq_sa_w_phl({{
|
|
|
|
dspac = dspMaq(dspac, Rs.uw, Rt.uw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
MODE_L, SATURATE, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x2: maq_sa_w_phr({{
|
|
|
|
dspac = dspMaq(dspac, Rs.uw, Rt.uw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
MODE_R, SATURATE, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x4: maq_s_w_phl({{
|
|
|
|
dspac = dspMaq(dspac, Rs.uw, Rt.uw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
MODE_L, NOSATURATE, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x6: maq_s_w_phr({{
|
|
|
|
dspac = dspMaq(dspac, Rs.uw, Rt.uw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
MODE_R, NOSATURATE, &dspctl);
|
|
|
|
}}, IntMultOp);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x3: decode OP_LO {
|
|
|
|
format DspHiLoOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: dpaqx_s_w_ph({{
|
|
|
|
dspac = dspDpaq(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
SIMD_FMT_W, NOSATURATE,
|
|
|
|
MODE_X, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x1: dpsqx_s_w_ph({{
|
|
|
|
dspac = dspDpsq(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
SIMD_FMT_W, NOSATURATE,
|
|
|
|
MODE_X, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x2: dpaqx_sa_w_ph({{
|
|
|
|
dspac = dspDpaq(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
SIMD_FMT_W, SATURATE,
|
|
|
|
MODE_X, &dspctl);
|
|
|
|
}}, IntMultOp);
|
|
|
|
0x3: dpsqx_sa_w_ph({{
|
|
|
|
dspac = dspDpsq(dspac, Rs.sw, Rt.sw,
|
|
|
|
ACDST, SIMD_FMT_PH,
|
|
|
|
SIMD_FMT_W, SATURATE,
|
|
|
|
MODE_X, &dspctl);
|
|
|
|
}}, IntMultOp);
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
//Table 3.3 MIPS32 APPEND Encoding of the op Field
|
|
|
|
0x1: decode OP_HI {
|
|
|
|
0x0: decode OP_LO {
|
|
|
|
format IntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: append({{
|
|
|
|
Rt.uw = (Rt.uw << RD) | bits(Rs.uw, RD - 1, 0);
|
|
|
|
}});
|
|
|
|
0x1: prepend({{
|
|
|
|
Rt.uw = (Rt.uw >> RD) |
|
|
|
|
(bits(Rs.uw, RD - 1, 0) << (32 - RD));
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x2: decode OP_LO {
|
|
|
|
format IntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: balign({{
|
|
|
|
Rt.uw = (Rt.uw << (8 * BP)) |
|
|
|
|
(Rs.uw >> (8 * (4 - BP)));
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2006-02-18 09:12:04 +01:00
|
|
|
}
|
2007-06-23 01:03:42 +02:00
|
|
|
0x7: decode FUNCTION_LO {
|
|
|
|
|
2009-07-22 10:51:10 +02:00
|
|
|
//Table 5-11 MIPS32 EXTR.W Encoding of the op Field
|
|
|
|
//(DSP ASE MANUAL)
|
2007-06-23 01:03:42 +02:00
|
|
|
0x0: decode OP_HI {
|
|
|
|
0x0: decode OP_LO {
|
|
|
|
format DspHiLoOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: extr_w({{
|
|
|
|
Rt.uw = dspExtr(dspac, SIMD_FMT_W, RS,
|
|
|
|
NOROUND, NOSATURATE, &dspctl);
|
|
|
|
}});
|
|
|
|
0x1: extrv_w({{
|
|
|
|
Rt.uw = dspExtr(dspac, SIMD_FMT_W, Rs.uw,
|
|
|
|
NOROUND, NOSATURATE, &dspctl);
|
|
|
|
}});
|
|
|
|
0x2: extp({{
|
|
|
|
Rt.uw = dspExtp(dspac, RS, &dspctl);
|
|
|
|
}});
|
|
|
|
0x3: extpv({{
|
|
|
|
Rt.uw = dspExtp(dspac, Rs.uw, &dspctl);
|
|
|
|
}});
|
|
|
|
0x4: extr_r_w({{
|
|
|
|
Rt.uw = dspExtr(dspac, SIMD_FMT_W, RS,
|
|
|
|
ROUND, NOSATURATE, &dspctl);
|
|
|
|
}});
|
|
|
|
0x5: extrv_r_w({{
|
|
|
|
Rt.uw = dspExtr(dspac, SIMD_FMT_W, Rs.uw,
|
|
|
|
ROUND, NOSATURATE, &dspctl);
|
|
|
|
}});
|
|
|
|
0x6: extr_rs_w({{
|
|
|
|
Rt.uw = dspExtr(dspac, SIMD_FMT_W, RS,
|
|
|
|
ROUND, SATURATE, &dspctl);
|
|
|
|
}});
|
|
|
|
0x7: extrv_rs_w({{
|
|
|
|
Rt.uw = dspExtr(dspac, SIMD_FMT_W, Rs.uw,
|
|
|
|
ROUND, SATURATE, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x1: decode OP_LO {
|
|
|
|
format DspHiLoOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: extpdp({{
|
|
|
|
Rt.uw = dspExtpd(dspac, RS, &dspctl);
|
|
|
|
}});
|
|
|
|
0x3: extpdpv({{
|
|
|
|
Rt.uw = dspExtpd(dspac, Rs.uw, &dspctl);
|
|
|
|
}});
|
|
|
|
0x6: extr_s_h({{
|
|
|
|
Rt.uw = dspExtr(dspac, SIMD_FMT_PH, RS,
|
|
|
|
NOROUND, SATURATE, &dspctl);
|
|
|
|
}});
|
|
|
|
0x7: extrv_s_h({{
|
|
|
|
Rt.uw = dspExtr(dspac, SIMD_FMT_PH, Rs.uw,
|
|
|
|
NOROUND, SATURATE, &dspctl);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x2: decode OP_LO {
|
|
|
|
format DspIntOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: rddsp({{
|
|
|
|
Rd.uw = readDSPControl(&dspctl, RDDSPMASK);
|
|
|
|
}});
|
|
|
|
0x3: wrdsp({{
|
|
|
|
writeDSPControl(&dspctl, Rs.uw, WRDSPMASK);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
0x3: decode OP_LO {
|
|
|
|
format DspHiLoOp {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: shilo({{
|
|
|
|
if (sext<6>(HILOSA) < 0) {
|
|
|
|
dspac = (uint64_t)dspac <<
|
|
|
|
-sext<6>(HILOSA);
|
|
|
|
} else {
|
|
|
|
dspac = (uint64_t)dspac >>
|
|
|
|
sext<6>(HILOSA);
|
|
|
|
}
|
|
|
|
}});
|
|
|
|
0x3: shilov({{
|
|
|
|
if (sext<6>(Rs.sw<5:0>) < 0) {
|
|
|
|
dspac = (uint64_t)dspac <<
|
|
|
|
-sext<6>(Rs.sw<5:0>);
|
|
|
|
} else {
|
|
|
|
dspac = (uint64_t)dspac >>
|
|
|
|
sext<6>(Rs.sw<5:0>);
|
|
|
|
}
|
|
|
|
}});
|
|
|
|
0x7: mthlip({{
|
|
|
|
dspac = dspac << 32;
|
|
|
|
dspac |= Rs.uw;
|
|
|
|
dspctl = insertBits(dspctl, 5, 0,
|
|
|
|
dspctl<5:0> + 32);
|
|
|
|
}});
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2009-12-31 21:30:51 +01:00
|
|
|
0x3: decode OP {
|
2009-12-31 21:30:51 +01:00
|
|
|
#if FULL_SYSTEM
|
2009-12-31 21:30:51 +01:00
|
|
|
0x0: FailUnimpl::rdhwr();
|
2009-12-31 21:30:51 +01:00
|
|
|
#else
|
|
|
|
0x0: decode RD {
|
|
|
|
29: BasicOp::rdhwr({{ Rt = TpValue; }});
|
|
|
|
}
|
|
|
|
#endif
|
2007-11-13 22:58:16 +01:00
|
|
|
}
|
2007-06-23 01:03:42 +02:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x4: decode OPCODE_LO {
|
2006-02-20 20:30:23 +01:00
|
|
|
format LoadMemory {
|
2010-10-16 09:00:54 +02:00
|
|
|
0x0: lb({{ Rt.sw = Mem.sb; }});
|
|
|
|
0x1: lh({{ Rt.sw = Mem.sh; }});
|
2006-03-17 00:39:54 +01:00
|
|
|
0x3: lw({{ Rt.sw = Mem.sw; }});
|
2010-10-16 09:00:54 +02:00
|
|
|
0x4: lbu({{ Rt.uw = Mem.ub;}});
|
|
|
|
0x5: lhu({{ Rt.uw = Mem.uh; }});
|
2006-06-09 09:57:25 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
format LoadUnalignedMemory {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: lwl({{
|
|
|
|
uint32_t mem_shift = 24 - (8 * byte_offset);
|
|
|
|
Rt.uw = mem_word << mem_shift | (Rt.uw & mask(mem_shift));
|
|
|
|
}});
|
|
|
|
0x6: lwr({{
|
|
|
|
uint32_t mem_shift = 8 * byte_offset;
|
|
|
|
Rt.uw = (Rt.uw & (mask(mem_shift) << (32 - mem_shift))) |
|
|
|
|
(mem_word >> mem_shift);
|
|
|
|
}});
|
|
|
|
}
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x5: decode OPCODE_LO {
|
2006-02-20 20:30:23 +01:00
|
|
|
format StoreMemory {
|
2010-10-16 09:00:54 +02:00
|
|
|
0x0: sb({{ Mem.ub = Rt<7:0>; }});
|
|
|
|
0x1: sh({{ Mem.uh = Rt<15:0>; }});
|
2006-03-17 00:39:54 +01:00
|
|
|
0x3: sw({{ Mem.uw = Rt<31:0>; }});
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
format StoreUnalignedMemory {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x2: swl({{
|
|
|
|
uint32_t reg_shift = 24 - (8 * byte_offset);
|
|
|
|
uint32_t mem_shift = 32 - reg_shift;
|
|
|
|
mem_word = (mem_word & (mask(reg_shift) << mem_shift)) |
|
|
|
|
(Rt.uw >> reg_shift);
|
|
|
|
}});
|
|
|
|
0x6: swr({{
|
|
|
|
uint32_t reg_shift = 8 * byte_offset;
|
|
|
|
mem_word = Rt.uw << reg_shift |
|
|
|
|
(mem_word & (mask(reg_shift)));
|
|
|
|
}});
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
format CP0Control {
|
|
|
|
0x7: cache({{
|
2007-11-15 20:21:01 +01:00
|
|
|
//Addr CacheEA = Rs.uw + OFFSET;
|
2009-07-22 10:51:10 +02:00
|
|
|
//fault = xc->CacheOp((uint8_t)CACHE_OP,(Addr) CacheEA);
|
|
|
|
}});
|
2007-11-13 22:58:16 +01:00
|
|
|
}
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x6: decode OPCODE_LO {
|
|
|
|
format LoadMemory {
|
2009-04-19 13:25:01 +02:00
|
|
|
0x0: ll({{ Rt.uw = Mem.uw; }}, mem_flags=LLSC);
|
2006-06-09 09:57:25 +02:00
|
|
|
0x1: lwc1({{ Ft.uw = Mem.uw; }});
|
2006-04-27 11:07:11 +02:00
|
|
|
0x5: ldc1({{ Ft.ud = Mem.ud; }});
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
0x2: CP2Unimpl::lwc2();
|
|
|
|
0x6: CP2Unimpl::ldc2();
|
2006-06-09 09:57:25 +02:00
|
|
|
0x3: Prefetch::pref();
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2006-02-08 00:36:08 +01:00
|
|
|
|
2006-03-08 08:05:38 +01:00
|
|
|
|
2006-06-09 09:57:25 +02:00
|
|
|
0x7: decode OPCODE_LO {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x0: StoreCond::sc({{ Mem.uw = Rt.uw; }},
|
2006-06-09 09:57:25 +02:00
|
|
|
{{ uint64_t tmp = write_result;
|
|
|
|
Rt.uw = (tmp == 0 || tmp == 1) ? tmp : Rt.uw;
|
2009-07-22 10:51:10 +02:00
|
|
|
}}, mem_flags=LLSC,
|
|
|
|
inst_flags = IsStoreConditional);
|
2006-06-09 09:57:25 +02:00
|
|
|
format StoreMemory {
|
2009-07-22 10:51:10 +02:00
|
|
|
0x1: swc1({{ Mem.uw = Ft.uw; }});
|
|
|
|
0x5: sdc1({{ Mem.ud = Ft.ud; }});
|
2006-02-15 04:43:14 +01:00
|
|
|
}
|
2007-11-13 22:58:16 +01:00
|
|
|
0x2: CP2Unimpl::swc2();
|
|
|
|
0x6: CP2Unimpl::sdc2();
|
2006-02-08 00:36:08 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|