[X86][CodeGen] Prefer KMOVkk_EVEX than KMOVkk when EGPR is supported #74048
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
In memory fold table, we have
```
{X86::KMOVDkk, X86::KMOVDkm, 0},
{X86::KMOVDkk_EVEX, X86::KMOVDkm_EVEX, 0}
```
where KMOVDkm_EVEX can use EGPR as base and index registers, while
KMOVDkm can't. Hence, though KMOVkk does not have any GPR operands,
we prefer to use KMOVDkk_EVEX to help memory fold.
It will be compressed to KMOVDkk in EVEX2VEX pass if memory folding
does not happen.
|
@llvm/pr-subscribers-backend-x86 Author: Shengchen Kan (KanRobert) ChangesIn memory fold table, we have where KMOVDkm_EVEX can use EGPR as base and index registers, while It will be compressed to KMOVDkk in EVEX2VEX pass if memory folding Full diff: https://github.com/llvm/llvm-project/pull/74048.diff 3 Files Affected:
diff --git a/llvm/lib/Target/X86/X86DomainReassignment.cpp b/llvm/lib/Target/X86/X86DomainReassignment.cpp
index be7e8db95b98ed8..bdd86e48fa54388 100644
--- a/llvm/lib/Target/X86/X86DomainReassignment.cpp
+++ b/llvm/lib/Target/X86/X86DomainReassignment.cpp
@@ -625,8 +625,10 @@ void X86DomainReassignment::initConverters() {
createReplacerDstCOPY(X86::MOVZX64rm16,
HasEGPR ? X86::KMOVWkm_EVEX : X86::KMOVWkm);
- createReplacerDstCOPY(X86::MOVZX32rr16, X86::KMOVWkk);
- createReplacerDstCOPY(X86::MOVZX64rr16, X86::KMOVWkk);
+ createReplacerDstCOPY(X86::MOVZX32rr16,
+ HasEGPR ? X86::KMOVWkk_EVEX : X86::KMOVWkk);
+ createReplacerDstCOPY(X86::MOVZX64rr16,
+ HasEGPR ? X86::KMOVWkk_EVEX : X86::KMOVWkk);
if (STI->hasDQI()) {
createReplacerDstCOPY(X86::MOVZX16rm8,
@@ -636,9 +638,12 @@ void X86DomainReassignment::initConverters() {
createReplacerDstCOPY(X86::MOVZX64rm8,
HasEGPR ? X86::KMOVBkm_EVEX : X86::KMOVBkm);
- createReplacerDstCOPY(X86::MOVZX16rr8, X86::KMOVBkk);
- createReplacerDstCOPY(X86::MOVZX32rr8, X86::KMOVBkk);
- createReplacerDstCOPY(X86::MOVZX64rr8, X86::KMOVBkk);
+ createReplacerDstCOPY(X86::MOVZX16rr8,
+ HasEGPR ? X86::KMOVBkk_EVEX : X86::KMOVBkk);
+ createReplacerDstCOPY(X86::MOVZX32rr8,
+ HasEGPR ? X86::KMOVBkk_EVEX : X86::KMOVBkk);
+ createReplacerDstCOPY(X86::MOVZX64rr8,
+ HasEGPR ? X86::KMOVBkk_EVEX : X86::KMOVBkk);
}
auto createReplacer = [&](unsigned From, unsigned To) {
@@ -647,7 +652,7 @@ void X86DomainReassignment::initConverters() {
createReplacer(X86::MOV16rm, HasEGPR ? X86::KMOVWkm_EVEX : X86::KMOVWkm);
createReplacer(X86::MOV16mr, HasEGPR ? X86::KMOVWmk_EVEX : X86::KMOVWmk);
- createReplacer(X86::MOV16rr, X86::KMOVWkk);
+ createReplacer(X86::MOV16rr, HasEGPR ? X86::KMOVWkk_EVEX : X86::KMOVWkk);
createReplacer(X86::SHR16ri, X86::KSHIFTRWri);
createReplacer(X86::SHL16ri, X86::KSHIFTLWri);
createReplacer(X86::NOT16r, X86::KNOTWrr);
@@ -662,8 +667,8 @@ void X86DomainReassignment::initConverters() {
createReplacer(X86::MOV32mr, HasEGPR ? X86::KMOVDmk_EVEX : X86::KMOVDmk);
createReplacer(X86::MOV64mr, HasEGPR ? X86::KMOVQmk_EVEX : X86::KMOVQmk);
- createReplacer(X86::MOV32rr, X86::KMOVDkk);
- createReplacer(X86::MOV64rr, X86::KMOVQkk);
+ createReplacer(X86::MOV32rr, HasEGPR ? X86::KMOVDkk_EVEX : X86::KMOVDkk);
+ createReplacer(X86::MOV64rr, HasEGPR ? X86::KMOVQkk_EVEX : X86::KMOVQkk);
createReplacer(X86::SHR32ri, X86::KSHIFTRDri);
createReplacer(X86::SHR64ri, X86::KSHIFTRQri);
@@ -703,7 +708,7 @@ void X86DomainReassignment::initConverters() {
createReplacer(X86::MOV8rm, HasEGPR ? X86::KMOVBkm_EVEX : X86::KMOVBkm);
createReplacer(X86::MOV8mr, HasEGPR ? X86::KMOVBmk_EVEX : X86::KMOVBmk);
- createReplacer(X86::MOV8rr, X86::KMOVBkk);
+ createReplacer(X86::MOV8rr, HasEGPR ? X86::KMOVBkk_EVEX : X86::KMOVBkk);
createReplacer(X86::NOT8r, X86::KNOTBrr);
diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index 583f8ec73a0361f..ea3bf1f101c1e08 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -4070,6 +4070,7 @@ void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
// First deal with the normal symmetric copies.
bool HasAVX = Subtarget.hasAVX();
bool HasVLX = Subtarget.hasVLX();
+ bool HasEGPR = Subtarget.hasEGPR();
unsigned Opc = 0;
if (X86::GR64RegClass.contains(DestReg, SrcReg))
Opc = X86::MOV64rr;
@@ -4124,7 +4125,8 @@ void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
// All KMASK RegClasses hold the same k registers, can be tested against
// anyone.
else if (X86::VK16RegClass.contains(DestReg, SrcReg))
- Opc = Subtarget.hasBWI() ? X86::KMOVQkk : X86::KMOVWkk;
+ Opc = Subtarget.hasBWI() ? (HasEGPR ? X86::KMOVQkk_EVEX : X86::KMOVQkk)
+ : (HasEGPR ? X86::KMOVQkk_EVEX : X86::KMOVWkk);
if (!Opc)
Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
diff --git a/llvm/test/CodeGen/X86/apx/kmov-kk.ll b/llvm/test/CodeGen/X86/apx/kmov-kk.ll
new file mode 100644
index 000000000000000..c9a4c5762872f45
--- /dev/null
+++ b/llvm/test/CodeGen/X86/apx/kmov-kk.ll
@@ -0,0 +1,14 @@
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx512f,+egpr -show-mc-encoding | FileCheck --check-prefix=EGPR %s
+
+define <16 x i32> @kmovkk(ptr %base, <16 x i32> %ind, i16 %mask) {
+; EGPR: kmovw %esi, %k1 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x92,0xce]
+ %broadcast.splatinsert = insertelement <16 x ptr> undef, ptr %base, i32 0
+ %broadcast.splat = shufflevector <16 x ptr> %broadcast.splatinsert, <16 x ptr> undef, <16 x i32> zeroinitializer
+ %gep.random = getelementptr i32, <16 x ptr> %broadcast.splat, <16 x i32> %ind
+ %imask = bitcast i16 %mask to <16 x i1>
+ %gt1 = call <16 x i32> @llvm.masked.gather.v16i32.v16p0(<16 x ptr> %gep.random, i32 4, <16 x i1> %imask, <16 x i32>undef)
+ %gt2 = call <16 x i32> @llvm.masked.gather.v16i32.v16p0(<16 x ptr> %gep.random, i32 4, <16 x i1> %imask, <16 x i32>%gt1)
+ %res = add <16 x i32> %gt1, %gt2
+ ret <16 x i32> %res
+}
+declare <16 x i32> @llvm.masked.gather.v16i32.v16p0(<16 x ptr>, i32, <16 x i1>, <16 x i32>)
|
RKSimon
reviewed
Dec 1, 2023
| createReplacerDstCOPY(X86::MOVZX32rr16, | ||
| HasEGPR ? X86::KMOVWkk_EVEX : X86::KMOVWkk); | ||
| createReplacerDstCOPY(X86::MOVZX64rr16, | ||
| HasEGPR ? X86::KMOVWkk_EVEX : X86::KMOVWkk); |
There was a problem hiding this comment.
Can we access + repurpose the EVEX->VEX tables to do this for us?
There was a problem hiding this comment.
Are you suggesting the usage like
createReplacerDstCOPY(X86::MOVZX64rr16,
getEVEXVariant(X86::KMOVWkk, /*UseEVEXVariant=*/HasEGPR));
There was a problem hiding this comment.
I think it's probably not worth it. First, KMOV may be the only instruction that needs to be promoted in CPP, and other instructions are implemented in TD through predicate HasEGPR/NoEGPR. Second, looking up the table will change the complexity from O(1) to O(lgN) and increase compile time.
There was a problem hiding this comment.
OK, no need to do it - I have found during debugging that tracking down all the places that we switch instruction encoding is not always easy. But maybe we should just improve debug output - the X86DomainReassignment pass for instance doesn't do a good job listing all the changes it makes.
phoebewang
reviewed
Dec 1, 2023
| ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx512f,+egpr -show-mc-encoding | FileCheck --check-prefix=EGPR %s | ||
|
|
||
| define <16 x i32> @kmovkk(ptr %base, <16 x i32> %ind, i16 %mask) { | ||
| ; EGPR: kmovq %k1, %k2 # EVEX TO VEX Compression encoding: [0xc4,0xe1,0xf8,0x90,0xd1] |
There was a problem hiding this comment.
Is it possible to add a test to cover the memory fold case?
There was a problem hiding this comment.
It's not straightforward to me. And I think the memory fold is already tested by ./llvm/test/TableGen/x86-fold-tables.inc
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
In memory fold table, we have
where
KMOVDkm_EVEXcan use EGPR as base and index registers, whileKMOVDkmcan't. Hence, thoughKMOVkkdoes not have any GPR operands,we prefer to use
KMOVDkk_EVEXto help register allocation.It will be compressed to
KMOVDkkin EVEX2VEX pass if memory foldingdoes not happen.