Summary:
It seems many CPUs don't implement this instruction as well as the other vector multiplies. Often using a multi uop flow. Silvermont in particular has a 7 uop flow with 11 cycle throughput. Sandy Bridge implements it as a single uop with 5 cycle latency and 1 cycle throughput. But Haswell and later use 2 uops with 10 cycle latency and 2 cycle throughput.
This patch adds a new X86SchedWritePair we can use to tag this instruction separately. I've provided correct information for Silvermont, Btver2, and Sandy Bridge. I've removed the InstRWs for SandyBridge. I've left Haswell/Broadwell/Skylake InstRWs in place because I wasn't sure how to account for the different load latency between 128 and 256 bits. I also left Znver1 InstRWs in place because the existing values don't match Agner's spreadsheet.
I also left a FIXME in the SandyBridge model because it being used for the "generic" model is too optimistic for the 256/512-bit versions since those are multiple uops on all known CPUs.
Reviewers: RKSimon, GGanesh, courbet
Reviewed By: RKSimon
Subscribers: gchatelet, gbedwell, andreadb, llvm-commits
Differential Revision: https://reviews.llvm.org/D44972
llvm-svn: 328914
These instructions have the memory operand before the register operand. So we need to put ReadDefault for all the load ops first. Then the ReadAfterLd
Differential Revision: https://reviews.llvm.org/D44838
llvm-svn: 328823
Currently MOVMSK instructions use the WriteVecLogic class, which is a very poor choice given that MOVMSK involves a SSE->GPR transfer.
Differential Revision: https://reviews.llvm.org/D44924
llvm-svn: 328664
Currently CRC32 instructions use the WriteFAdd class, this patch splits them off into their own, at the moment it is still mostly just a duplicate of WriteFAdd but it can now be tweaked on a target by target basis.
Differential Revision: https://reviews.llvm.org/D44647
llvm-svn: 328582
Give the bit count instructions their own scheduler classes instead of forcing them into existing classes.
These were mostly overridden anyway, but I had to add in costs from Agner for silvermont and znver1 and the Fam16h SoG for btver2 (Jaguar).
Differential Revision: https://reviews.llvm.org/D44879
llvm-svn: 328566
The Intel models need an extra 1uop for memory folded instructions, plus a lot of instructions take a non-default memory latency which should allow us to use the multiclass a lot more to tidy things up.
Differential Revision: https://reviews.llvm.org/D44840
llvm-svn: 328446
This makes the Y position consistent with other instructions.
This should have been NFC, but while refactoring the multiclass I noticed that VROUNDPD memory forms were using the register itinerary.
llvm-svn: 328254
Models were completely overriding all CLMUL instructions when the WriteCLMUL default classes could be used for exactly the same coverage.
llvm-svn: 328194
Models were completely overriding all AES instructions when the WriteAES default classes could be used for exactly the same coverage.
Removes 6 unnecessary scheduler classes from every model.
Note: Still looking for a way for tblgen to warn when this is happening - often the override is more complete than the default.
llvm-svn: 328192
JRCXZ was already present, but not the others.
We never codegen this instruction so this doesn't affect much just trying to get them all into a single generated scheduler class in the output.
llvm-svn: 327881
Currently the WriteResPair style multi-classes take a single pipeline stage and latency, this patch generalizes this to make it easier to create complex schedules with ResourceCycles and NumMicroOps be overriden from their defaults.
This has already been done for the Jaguar scheduler to remove a number of custom schedule classes and adding it to the other x86 targets will make it much tidier as we add additional classes in the future to try and replace so many custom cases.
I've converted some instructions but a lot of the models need a bit of cleanup after the patch has been committed - memory latencies not being consistent, the class not actually being used when we could remove some/all customs, etc. I'd prefer to keep this as NFC as possible so later patches can be smaller and target specific.
Differential Revision: https://reviews.llvm.org/D44612
llvm-svn: 327855
Also move ADC8i8 and SBB8i8 in the Sandy Bridge model to the same class as ADC8ri and SBB8ri. That seems more accurate since its the 8i8 is just the register forced to AL instead of coming from modrm.
llvm-svn: 327820
The information was so wildly inaccurate and incomplete its better to just remove it.
MMX_MASKMOVQ64 showed up twice in several scheduler models. In Haswell and Broadwell they were on adjacent lines. On Skylake the copies had different information.
MMX_MASKMOVQ and MASKMOVDQU were completely missing.
MMX_MASKMOVQ64 was listed on Haswell/Broadwell as 1 cycle on port 1 despite it being a store instruction.
Filed PR36780 to track fixing this right.
llvm-svn: 327783
As discussed on D44428 and PR36726, this patch splits off WriteFMove/WriteVecMove, WriteFLoad/WriteVecLoad and WriteFStore/WriteVecStore scheduler classes to permit vectors to be handled separately from gpr/scalar types.
I've minimised the diff here by only moving various basic SSE/AVX vector instructions across - we can fix the rest when called for. This does fix the MOVDQA vs MOVAPS/MOVAPD discrepancies mentioned on D44428.
Differential Revision: https://reviews.llvm.org/D44471
llvm-svn: 327630
Summary:
Only IMUL16rri uses an extra P0156. IMUL32* and IMUL16rr only use
P1.
This was computed using https://github.com/google/EXEgesis/blob/master/exegesis/tools/compute_itineraries.cc
This can easily be validated by running perf on the following code:
```
int main(int argc, char**argv) {
int a = argc;
int b = argc;
int c = argc;
int d = argc;
for (int i = 0; i < LOOP_ITERATIONS; ++i) {
asm volatile(
R"(
.rept 10000
imull $0x2, %%edx, %%eax
imull $0x2, %%ecx, %%ebx
imull $0x2, %%eax, %%edx
imull $0x2, %%ebx, %%ecx
.endr
)"
: "+a"(a), "+b"(b), "+c"(c), "+d"(d)
:
:);
}
return a+b+c+d;
}
```
-> test.cc
perf stat -x, -e cycles --pfm-events=uops_executed_port:port_0:u,uops_executed_port:port_1:u,uops_executed_port:port_2:u,uops_executed_port:port_3:u,uops_executed_port:port_4:u,uops_executed_port:port_5:u,uops_executed_port:port_6:u,uops_executed_port:port_7:u test
Reviewers: craig.topper, RKSimon, gadi.haber
Subscribers: llvm-commits, gchatelet, chandlerc
Differential Revision: https://reviews.llvm.org/D43460
llvm-svn: 326877
The regular expressions and the imul names caused some instructions to be matched by multiple regexs creating unpredictable results.
This changes them all to use explicit instrs instead.
While doing this I also found that some instructions in Skylake were missing load latency so I fixed that too.
llvm-svn: 323406
MMX instrutions all start with MMX_ so the 64 isn't needed for disambigutation.
SSE/AVX1 instructions are assumed 128-bit so we don't need to say 128.
AVX2 instructions should use a Y to indicate 256-bits.
llvm-svn: 323402
These were treated as optional suffixes, but the regular expressions are already prefix matches so this is unnecessary. It breaks the binary search optimization in tablegen due to the top level question mark.
llvm-svn: 323401
The regexs are treated as a prefix match already so the checking for optional text at the end provides no value. Instead it prevents the binary search optimization in tablegen from kicking in due to the top level question mark.
llvm-svn: 323351
We had a lot of separate 32 and 64 instructions that had the same scheduling data. This merges them into the same regular expression. This is pretty consistent with a lot of other instructions.
llvm-svn: 320924
I've hopefully sidestepped the MSVC issue that caused it to be reverted. We no longer include the Sched enum from X86GenInstrInfo.inc on the X86 target. So hopefully MSVC's preprocessor will skip over it and nothing will notice the 11000 character enum name.
Original commit message:
When the scheduler tables are generated by tablegen, the instructions are divided up into groups based on their default scheduling information and how they are referenced by groups for each processor. For any set of instructions that are matched by a specific InstRW line, that group of instructions is guaranteed to not be in a group with any other instructions. So in general, the more InstRW class definitions are created, the more groups we end up with in the generated files. Particularly if a lot of the InstRW lines only match to single instructions, which is true of a large number of the Intel scheduler models.
This change alone reduces the number of instructions groups from ~6000 to ~5500. And there's lots more we could do.
llvm-svn: 320655
This reverts commit r320508, in effect re-applying r320308. Simon has already
reverted the parts that caused the crash that motivated the revert in r320492.
llvm-svn: 320512
The checks we have for complete models are not great and miss many cases - e.g. in PR35636 it failed to recognise that only the first output (of 2) was actually tagged by the InstRW
Raised PR35639 and PR35643 as examples
llvm-svn: 320492
[X86] Use regular expressions more aggressively to reduce the number of scheduler entries needed for FMA3 instructions.
When the scheduler tables are generated by tablegen, the instructions are divided up into groups based on their default scheduling information and how they are referenced by groups for each processor. For any set of instructions that are matched by a specific InstRW line, that group of instructions is guaranteed to not be in a group with any other instructions. So in general, the more InstRW class definitions are created, the more groups we end up with in the generated files. Particularly if a lot of the InstRW lines only match to single instructions, which is true of a large number of the Intel scheduler models.
This change alone reduces the number of instructions groups from ~6000 to ~5500. And there's lots more we could do.
llvm-svn: 320470
When the scheduler tables are generated by tablegen, the instructions are divided up into groups based on their default scheduling information and how they are referenced by groups for each processor. For any set of instructions that are matched by a specific InstRW line, that group of instructions is guaranteed to not be in a group with any other instructions. So in general, the more InstRW class definitions are created, the more groups we end up with in the generated files. Particularly if a lot of the InstRW lines only match to single instructions, which is true of a large number of the Intel scheduler models.
This change alone reduces the number of instructions groups from ~6000 to ~5500. And there's lots more we could do.
llvm-svn: 320461
This matches AVX512 version and is more consistent overall. And improves our scheduler models.
In some cases this adds _Int to instructions that didn't have any Int_ before. It's a side effect of the adjustments made to some of the multiclasses.
llvm-svn: 320325
Craig Topper