RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-23 11:13:28 +01:00
Code Issues Projects Releases Wiki Activity

[X86] Add support for commuting EVEX VCMP instructons with any immediate value.

Browse Source 
Previously we limited to the EQ/NE/TRUE/FALSE/ORD/UNORD immediates.

llvm-svn: 372067
This commit is contained in:
Craig Topper 2019-09-17 04:41:05 +00:00
parent 9a1ee6dff6
commit 6a4f10bb7e
2 changed files with 34 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

39
lib/Target/X86/X86InstrInfo.cpp
View File

@ -1720,6 +1720,27 @@ MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
OpIdx1, OpIdx2); OpIdx1, OpIdx2);
} }
case X86::VCMPSDZrr:
case X86::VCMPSSZrr:
case X86::VCMPPDZrri:
case X86::VCMPPSZrri:
case X86::VCMPPDZ128rri:
case X86::VCMPPSZ128rri:
case X86::VCMPPDZ256rri:
case X86::VCMPPSZ256rri:
case X86::VCMPPDZrrik:
case X86::VCMPPSZrrik:
case X86::VCMPPDZ128rrik:
case X86::VCMPPSZ128rrik:
case X86::VCMPPDZ256rrik:
case X86::VCMPPSZ256rrik: {
unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm() & 0x1f;
Imm = X86::getSwappedVCMPImm(Imm);
auto &WorkingMI = cloneIfNew(MI);
WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(Imm);
return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
OpIdx1, OpIdx2);
}
case X86::VPERM2F128rr: case X86::VPERM2F128rr:
case X86::VPERM2I128rr: { case X86::VPERM2I128rr: {
// Flip permute source immediate. // Flip permute source immediate.
@ -1963,17 +1984,23 @@ bool X86InstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
// Ordered/Unordered/Equal/NotEqual tests // Ordered/Unordered/Equal/NotEqual tests
unsigned Imm = MI.getOperand(3 + OpOffset).getImm() & 0x7; unsigned Imm = MI.getOperand(3 + OpOffset).getImm() & 0x7;
switch (Imm) { switch (Imm) {
default:
// EVEX versions can be commuted.
if ((Desc.TSFlags & X86II::EncodingMask) == X86II::EVEX)
break;
return false;
case 0x00: // EQUAL case 0x00: // EQUAL
case 0x03: // UNORDERED case 0x03: // UNORDERED
case 0x04: // NOT EQUAL case 0x04: // NOT EQUAL
case 0x07: // ORDERED case 0x07: // ORDERED
// The indices of the commutable operands are 1 and 2 (or 2 and 3 break;
// when masked).
// Assign them to the returned operand indices here.
return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1 + OpOffset,
2 + OpOffset);
} }
return false;
// The indices of the commutable operands are 1 and 2 (or 2 and 3
// when masked).
// Assign them to the returned operand indices here.
return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1 + OpOffset,
2 + OpOffset);
} }
case X86::MOVSSrr: case X86::MOVSSrr:
// X86::MOVSDrr is always commutable. MOVSS is only commutable if we can // X86::MOVSDrr is always commutable. MOVSS is only commutable if we can

3
test/CodeGen/X86/avx512-broadcast-unfold.ll
View File

@ -4457,8 +4457,7 @@ define void @bcast_unfold_cmp_v8f32_refold(float* nocapture %0) {
; CHECK-NEXT: vbroadcastss {{.*#+}} ymm1 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0] ; CHECK-NEXT: vbroadcastss {{.*#+}} ymm1 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
; CHECK-NEXT: .p2align 4, 0x90 ; CHECK-NEXT: .p2align 4, 0x90
; CHECK-NEXT: .LBB126_1: # =>This Inner Loop Header: Depth=1 ; CHECK-NEXT: .LBB126_1: # =>This Inner Loop Header: Depth=1
; CHECK-NEXT: vmovups 4096(%rdi,%rax), %ymm2 ; CHECK-NEXT: vcmpgtps 4096(%rdi,%rax), %ymm0, %k1
; CHECK-NEXT: vcmpltps %ymm0, %ymm2, %k1
; CHECK-NEXT: vmovaps %ymm1, %ymm2 ; CHECK-NEXT: vmovaps %ymm1, %ymm2
; CHECK-NEXT: vbroadcastss {{.*}}(%rip), %ymm2 {%k1} ; CHECK-NEXT: vbroadcastss {{.*}}(%rip), %ymm2 {%k1}
; CHECK-NEXT: vmovups %ymm2, 4096(%rdi,%rax) ; CHECK-NEXT: vmovups %ymm2, 4096(%rdi,%rax)