Skip to content

[X86][CodeGen] Support EVEX compression: NDD to nonNDD #77731

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Jan 12, 2024
Merged

[X86][CodeGen] Support EVEX compression: NDD to nonNDD #77731

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
dc7c1d1
[X86][CodeGen] Support EVEX compression: NDD to nonNDD
KanRobert Jan 10, 2024
cb8e045
address review comment
KanRobert Jan 12, 2024
cb54770
remove empty line
KanRobert Jan 12, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
42 changes: 37 additions & 5 deletions llvm/lib/Target/X86/X86CompressEVEX.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -220,6 +220,29 @@ static bool performCustomAdjustments(MachineInstr &MI, unsigned NewOpc) {
return true;
}

static bool isRedundantNewDataDest(MachineInstr &MI, const X86Subtarget &ST) {
// $rbx = ADD64rr_ND $rbx, $rax / $rbx = ADD64rr_ND $rax, $rbx
// ->
// $rbx = ADD64rr $rbx, $rax
const MCInstrDesc &Desc = MI.getDesc();
Register Reg0 = MI.getOperand(0).getReg();
const MachineOperand &Op1 = MI.getOperand(1);
if (!Op1.isReg())
return false;
Register Reg1 = Op1.getReg();
if (Reg1 == Reg0)
return true;

// Op1 and Op2 may be commutable for ND instructions.
if (!Desc.isCommutable() || Desc.getNumOperands() < 3 ||
!MI.getOperand(2).isReg() || MI.getOperand(2).getReg() != Reg0)
return false;
// Opcode may change after commute, e.g. SHRD -> SHLD
// TODO: Add test for this after ND SHRD/SHLD is supported
ST.getInstrInfo()->commuteInstruction(MI, false, 1, 2);
return true;
}

static bool CompressEVEXImpl(MachineInstr &MI, const X86Subtarget &ST) {
uint64_t TSFlags = MI.getDesc().TSFlags;

Expand All @@ -241,26 +264,35 @@ static bool CompressEVEXImpl(MachineInstr &MI, const X86Subtarget &ST) {
//
// For AVX512 cases, EVEX prefix is needed in order to carry this information
// thus preventing the transformation to VEX encoding.
bool IsND = X86II::hasNewDataDest(TSFlags);
if (TSFlags & X86II::EVEX_B)
return false;
if (!IsND || !isRedundantNewDataDest(MI, ST))
return false;

ArrayRef<X86CompressEVEXTableEntry> Table = ArrayRef(X86CompressEVEXTable);

unsigned Opc = MI.getOpcode();
const auto *I = llvm::lower_bound(Table, Opc);
if (I == Table.end() || I->OldOpc != Opc)
if (I == Table.end() || I->OldOpc != Opc) {
assert(!IsND && "Missing entry for ND instruction");
return false;
}

if (usesExtendedRegister(MI) || !checkVEXInstPredicate(Opc, ST) ||
!performCustomAdjustments(MI, I->NewOpc))
return false;
if (!IsND) {
if (usesExtendedRegister(MI) || !checkVEXInstPredicate(Opc, ST) ||
!performCustomAdjustments(MI, I->NewOpc))
return false;
}

const MCInstrDesc &NewDesc = ST.getInstrInfo()->get(I->NewOpc);
MI.setDesc(NewDesc);
uint64_t Encoding = NewDesc.TSFlags & X86II::EncodingMask;
auto AsmComment =
(Encoding == X86II::VEX) ? X86::AC_EVEX_2_VEX : X86::AC_EVEX_2_LEGACY;
MI.setAsmPrinterFlag(AsmComment);
if (IsND)
MI.tieOperands(0, 1);

return true;
}

Expand Down
10 changes: 5 additions & 5 deletions llvm/test/CodeGen/X86/apx/adc.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,7 +114,7 @@ define i16 @adc16ri8(i16 %a, i16 %x, i16 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subw %si, %dx, %ax
; CHECK-NEXT: adcw $0, %di, %ax
; CHECK-NEXT: addl $123, %eax, %eax
; CHECK-NEXT: addl $123, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
%s = add i16 %a, 123
Expand Down Expand Up @@ -168,7 +168,7 @@ define i16 @adc16ri(i16 %a, i16 %x, i16 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subw %si, %dx, %ax
; CHECK-NEXT: adcw $0, %di, %ax
; CHECK-NEXT: addl $1234, %eax, %eax # imm = 0x4D2
; CHECK-NEXT: addl $1234, %eax # imm = 0x4D2
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
%s = add i16 %a, 1234
Expand Down Expand Up @@ -265,7 +265,7 @@ define i16 @adc16mi8(ptr %ptr, i16 %x, i16 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subw %si, %dx, %ax
; CHECK-NEXT: adcw $0, (%rdi), %ax
; CHECK-NEXT: addl $123, %eax, %eax
; CHECK-NEXT: addl $123, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
%a = load i16, ptr %ptr
Expand Down Expand Up @@ -323,7 +323,7 @@ define i16 @adc16mi(ptr %ptr, i16 %x, i16 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subw %si, %dx, %ax
; CHECK-NEXT: adcw $0, (%rdi), %ax
; CHECK-NEXT: addl $1234, %eax, %eax # imm = 0x4D2
; CHECK-NEXT: addl $1234, %eax # imm = 0x4D2
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
%a = load i16, ptr %ptr
Expand Down Expand Up @@ -442,7 +442,7 @@ define void @adc16mi_legacy(ptr %ptr, i16 %x, i16 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subw %si, %dx, %ax
; CHECK-NEXT: adcw $0, (%rdi), %ax
; CHECK-NEXT: addl $1234, %eax, %eax # imm = 0x4D2
; CHECK-NEXT: addl $1234, %eax # imm = 0x4D2
; CHECK-NEXT: movw %ax, (%rdi)
; CHECK-NEXT: retq
%a = load i16, ptr %ptr
Expand Down
8 changes: 4 additions & 4 deletions llvm/test/CodeGen/X86/apx/add.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -206,7 +206,7 @@ define i16 @add16mi8(ptr %a) {
; CHECK-LABEL: add16mi8:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: movzwl (%rdi), %eax
; CHECK-NEXT: addl $123, %eax, %eax
; CHECK-NEXT: addl $123, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
entry:
Expand Down Expand Up @@ -252,7 +252,7 @@ define i16 @add16mi(ptr %a) {
; CHECK-LABEL: add16mi:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: movzwl (%rdi), %eax
; CHECK-NEXT: addl $1234, %eax, %eax # imm = 0x4D2
; CHECK-NEXT: addl $1234, %eax # imm = 0x4D2
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
entry:
Expand Down Expand Up @@ -489,8 +489,8 @@ define i1 @add64ri_reloc(i16 %k) {
; CHECK: # %bb.0:
; CHECK-NEXT: # kill: def $edi killed $edi def $rdi
; CHECK-NEXT: movswq %di, %rax
; CHECK-NEXT: addq %rax, %rax, %rax
; CHECK-NEXT: addq $val, %rax, %rax
; CHECK-NEXT: addq %rax, %rax
; CHECK-NEXT: addq $val, %rax
; CHECK-NEXT: setne %al
; CHECK-NEXT: retq
%g = getelementptr inbounds i16, ptr @val, i16 %k
Expand Down
21 changes: 14 additions & 7 deletions llvm/test/CodeGen/X86/apx/compress-evex.mir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,38 +1,35 @@
# RUN: llc %s -mtriple=x86_64-unknown -mattr=+ndd -start-before=x86-compress-evex -show-mc-encoding -o - | FileCheck %s
# RUN: llc %s -mtriple=x86_64-unknown -mattr=+ndd,+egpr -start-before=x86-compress-evex -show-mc-encoding -o - | FileCheck %s

...
---
name: ndd_2_non_ndd_xor
body: |
bb.0.entry:
liveins: $rdi, $rsi
; CHECK: xorq %rsi, %rax, %rax # encoding: [0x62,0xf4,0xfc,0x18,0x31,0xf0]
; CHECK: xorq %rsi, %rax # EVEX TO LEGACY Compression encoding: [0x48,0x31,0xf0]
renamable $rax = ADD64rr_ND killed renamable $rdi, renamable $rsi, implicit-def dead $eflags
renamable $rax = XOR64rr_ND killed renamable $rax, killed renamable $rsi, implicit-def dead $eflags
RET64 $rax

...
---
name: ndd_2_non_ndd_sub
body: |
bb.0.entry:
liveins: $rdi, $rsi
; CHECK: subq %rsi, %rax, %rax # encoding: [0x62,0xf4,0xfc,0x18,0x29,0xf0]
; CHECK: subq %rsi, %rax # EVEX TO LEGACY Compression encoding: [0x48,0x29,0xf0]
renamable $rax = ADD64rr_ND killed renamable $rdi, renamable $rsi, implicit-def dead $eflags
renamable $rax = SUB64rr_ND killed renamable $rax, killed renamable $rsi, implicit-def dead $eflags
RET64 $rax

...
---
name: ndd_2_non_ndd_commutable
body: |
bb.0.entry:
liveins: $rdi, $rsi
; CHECK: xorq %rax, %rsi, %rax # encoding: [0x62,0xf4,0xfc,0x18,0x31,0xc6]
; CHECK: xorq %rsi, %rax # EVEX TO LEGACY Compression encoding: [0x48,0x31,0xf0]
renamable $rax = ADD64rr_ND killed renamable $rdi, renamable $rsi, implicit-def dead $eflags
renamable $rax = XOR64rr_ND killed renamable $rsi, killed renamable $rax, implicit-def dead $eflags
RET64 $rax

...
---
name: ndd_2_non_ndd_incommutable
Expand All @@ -44,3 +41,13 @@ body: |
renamable $rax = SUB64rr_ND killed renamable $rsi, killed renamable $rax, implicit-def dead $eflags
RET64 $rax
...
---
name: ndd_2_non_ndd_egpr
body: |
bb.0.entry:
liveins: $rdi, $r16
; CHECK: xorq %r16, %rax # EVEX TO LEGACY Compression encoding: [0xd5,0x48,0x31,0xc0]
renamable $rax = ADD64rr_ND killed renamable $rdi, renamable $r16, implicit-def dead $eflags
renamable $rax = XOR64rr_ND killed renamable $rax, killed renamable $r16, implicit-def dead $eflags
RET64 $rax
...
2 changes: 1 addition & 1 deletion llvm/test/CodeGen/X86/apx/dec.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,7 @@ define i16 @dec16m(ptr %ptr) {
; CHECK-LABEL: dec16m:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: movzwl (%rdi), %eax
; CHECK-NEXT: decl %eax, %eax
; CHECK-NEXT: decl %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
entry:
Expand Down
2 changes: 1 addition & 1 deletion llvm/test/CodeGen/X86/apx/inc.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,7 @@ define i16 @inc16m(ptr %ptr) {
; CHECK-LABEL: inc16m:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: movzwl (%rdi), %eax
; CHECK-NEXT: incl %eax, %eax
; CHECK-NEXT: incl %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
entry:
Expand Down
4 changes: 2 additions & 2 deletions llvm/test/CodeGen/X86/apx/or.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -207,7 +207,7 @@ define i16 @or16mi8(ptr %a) {
; CHECK-LABEL: or16mi8:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: movzwl (%rdi), %eax
; CHECK-NEXT: orl $123, %eax, %eax
; CHECK-NEXT: orl $123, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
entry:
Expand Down Expand Up @@ -253,7 +253,7 @@ define i16 @or16mi(ptr %a) {
; CHECK-LABEL: or16mi:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: movzwl (%rdi), %eax
; CHECK-NEXT: orl $1234, %eax, %eax # imm = 0x4D2
; CHECK-NEXT: orl $1234, %eax # imm = 0x4D2
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
entry:
Expand Down
28 changes: 14 additions & 14 deletions llvm/test/CodeGen/X86/apx/sbb.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,7 +114,7 @@ define i16 @sbb16ri8(i16 %a, i16 %x, i16 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subw %si, %dx, %ax
; CHECK-NEXT: sbbw $0, %di, %ax
; CHECK-NEXT: addl $-123, %eax, %eax
; CHECK-NEXT: addl $-123, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
%s = sub i16 %a, 123
Expand All @@ -129,7 +129,7 @@ define i32 @sbb32ri8(i32 %a, i32 %x, i32 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subl %esi, %edx, %eax
; CHECK-NEXT: sbbl $0, %edi, %eax
; CHECK-NEXT: addl $-123, %eax, %eax
; CHECK-NEXT: addl $-123, %eax
; CHECK-NEXT: retq
%s = sub i32 %a, 123
%k = icmp ugt i32 %x, %y
Expand All @@ -143,7 +143,7 @@ define i64 @sbb64ri8(i64 %a, i64 %x, i64 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subq %rsi, %rdx, %rax
; CHECK-NEXT: sbbq $0, %rdi, %rax
; CHECK-NEXT: addq $-123, %rax, %rax
; CHECK-NEXT: addq $-123, %rax
; CHECK-NEXT: retq
%s = sub i64 %a, 123
%k = icmp ugt i64 %x, %y
Expand All @@ -157,7 +157,7 @@ define i8 @sbb8ri(i8 %a, i8 %x, i8 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subb %sil, %dl, %al
; CHECK-NEXT: sbbb $0, %dil, %al
; CHECK-NEXT: addb $-123, %al, %al
; CHECK-NEXT: addb $-123, %al
; CHECK-NEXT: retq
%s = sub i8 %a, 123
%k = icmp ugt i8 %x, %y
Expand All @@ -171,7 +171,7 @@ define i16 @sbb16ri(i16 %a, i16 %x, i16 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subw %si, %dx, %ax
; CHECK-NEXT: sbbw $0, %di, %ax
; CHECK-NEXT: addl $-1234, %eax, %eax # imm = 0xFB2E
; CHECK-NEXT: addl $-1234, %eax # imm = 0xFB2E
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
%s = sub i16 %a, 1234
Expand All @@ -186,7 +186,7 @@ define i32 @sbb32ri(i32 %a, i32 %x, i32 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subl %esi, %edx, %eax
; CHECK-NEXT: sbbl $0, %edi, %eax
; CHECK-NEXT: addl $-123456, %eax, %eax # imm = 0xFFFE1DC0
; CHECK-NEXT: addl $-123456, %eax # imm = 0xFFFE1DC0
; CHECK-NEXT: retq
%s = sub i32 %a, 123456
%k = icmp ugt i32 %x, %y
Expand All @@ -200,7 +200,7 @@ define i64 @sbb64ri(i64 %a, i64 %x, i64 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subq %rsi, %rdx, %rax
; CHECK-NEXT: sbbq $0, %rdi, %rax
; CHECK-NEXT: addq $-123456, %rax, %rax # imm = 0xFFFE1DC0
; CHECK-NEXT: addq $-123456, %rax # imm = 0xFFFE1DC0
; CHECK-NEXT: retq
%s = sub i64 %a, 123456
%k = icmp ugt i64 %x, %y
Expand Down Expand Up @@ -270,7 +270,7 @@ define i16 @sbb16mi8(ptr %ptr, i16 %x, i16 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subw %si, %dx, %ax
; CHECK-NEXT: sbbw $0, (%rdi), %ax
; CHECK-NEXT: addl $-123, %eax, %eax
; CHECK-NEXT: addl $-123, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
%a = load i16, ptr %ptr
Expand All @@ -286,7 +286,7 @@ define i32 @sbb32mi8(ptr %ptr, i32 %x, i32 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subl %esi, %edx, %eax
; CHECK-NEXT: sbbl $0, (%rdi), %eax
; CHECK-NEXT: addl $-123, %eax, %eax
; CHECK-NEXT: addl $-123, %eax
; CHECK-NEXT: retq
%a = load i32, ptr %ptr
%s = sub i32 %a, 123
Expand All @@ -301,7 +301,7 @@ define i64 @sbb64mi8(ptr %ptr, i64 %x, i64 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subq %rsi, %rdx, %rax
; CHECK-NEXT: sbbq $0, (%rdi), %rax
; CHECK-NEXT: addq $-123, %rax, %rax
; CHECK-NEXT: addq $-123, %rax
; CHECK-NEXT: retq
%a = load i64, ptr %ptr
%s = sub i64 %a, 123
Expand All @@ -316,7 +316,7 @@ define i8 @sbb8mi(ptr %ptr, i8 %x, i8 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subb %sil, %dl, %al
; CHECK-NEXT: sbbb $0, (%rdi), %al
; CHECK-NEXT: addb $-123, %al, %al
; CHECK-NEXT: addb $-123, %al
; CHECK-NEXT: retq
%a = load i8, ptr %ptr
%s = sub i8 %a, 123
Expand All @@ -331,7 +331,7 @@ define i16 @sbb16mi(ptr %ptr, i16 %x, i16 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subw %si, %dx, %ax
; CHECK-NEXT: sbbw $0, (%rdi), %ax
; CHECK-NEXT: addl $-1234, %eax, %eax # imm = 0xFB2E
; CHECK-NEXT: addl $-1234, %eax # imm = 0xFB2E
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
%a = load i16, ptr %ptr
Expand All @@ -347,7 +347,7 @@ define i32 @sbb32mi(ptr %ptr, i32 %x, i32 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subl %esi, %edx, %eax
; CHECK-NEXT: sbbl $0, (%rdi), %eax
; CHECK-NEXT: addl $-123456, %eax, %eax # imm = 0xFFFE1DC0
; CHECK-NEXT: addl $-123456, %eax # imm = 0xFFFE1DC0
; CHECK-NEXT: retq
%a = load i32, ptr %ptr
%s = sub i32 %a, 123456
Expand All @@ -362,7 +362,7 @@ define i64 @sbb64mi(ptr %ptr, i64 %x, i64 %y) nounwind {
; CHECK: # %bb.0:
; CHECK-NEXT: subq %rsi, %rdx, %rax
; CHECK-NEXT: sbbq $0, (%rdi), %rax
; CHECK-NEXT: addq $-123456, %rax, %rax # imm = 0xFFFE1DC0
; CHECK-NEXT: addq $-123456, %rax # imm = 0xFFFE1DC0
; CHECK-NEXT: retq
%a = load i64, ptr %ptr
%s = sub i64 %a, 123456
Expand Down
6 changes: 3 additions & 3 deletions llvm/test/CodeGen/X86/apx/sub.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,7 +173,7 @@ define i16 @sub16mr(ptr %a, i16 noundef %b) {
; CHECK-LABEL: sub16mr:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: movzwl (%rdi), %eax
; CHECK-NEXT: subl %esi, %eax, %eax
; CHECK-NEXT: subl %esi, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
entry:
Expand Down Expand Up @@ -208,7 +208,7 @@ define i16 @sub16mi8(ptr %a) {
; CHECK-LABEL: sub16mi8:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: movzwl (%rdi), %eax
; CHECK-NEXT: addl $-123, %eax, %eax
; CHECK-NEXT: addl $-123, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
entry:
Expand Down Expand Up @@ -254,7 +254,7 @@ define i16 @sub16mi(ptr %a) {
; CHECK-LABEL: sub16mi:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: movzwl (%rdi), %eax
; CHECK-NEXT: addl $-1234, %eax, %eax # imm = 0xFB2E
; CHECK-NEXT: addl $-1234, %eax # imm = 0xFB2E
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
; CHECK-NEXT: retq
entry:
Expand Down
Loading