[llvm] [X86] Invalid fp16 comparison fix (PR #160304)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-x86

Author: None (azwolski)

<details>
<summary>Changes</summary>

Missing `VCMPPHZrrik`, `VCMPPHZ128rrik`, and `VCMPPHZ256rrik` opcodes in `commuteInstructionImpl` and `findCommutedOpIndices` led to improper handling of compare instruction during optimization. Operands were commuted, but swapping of the immediate was not called.
 
 Fixes: #<!-- -->159723
 

---
Full diff: https://github.com/llvm/llvm-project/pull/160304.diff


2 Files Affected:

- (modified) llvm/lib/Target/X86/X86InstrInfo.cpp (+7-1) 
- (modified) llvm/test/CodeGen/X86/pr159723.ll (+3-3) 


``````````diff
diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index 58d526269ff3c..497441d8b6f93 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -2573,10 +2573,13 @@ MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
   case X86::VCMPPSZ256rri:
   case X86::VCMPPDZrrik:
   case X86::VCMPPSZrrik:
+  case X86::VCMPPHZrrik:
   case X86::VCMPPDZ128rrik:
   case X86::VCMPPSZ128rrik:
+  case X86::VCMPPHZ128rrik:
   case X86::VCMPPDZ256rrik:
   case X86::VCMPPSZ256rrik:
+  case X86::VCMPPHZ256rrik:
     WorkingMI = CloneIfNew(MI);
     WorkingMI->getOperand(MI.getNumExplicitOperands() - 1)
         .setImm(X86::getSwappedVCMPImm(
@@ -2830,10 +2833,13 @@ bool X86InstrInfo::findCommutedOpIndices(const MachineInstr &MI,
   case X86::VCMPPSZ256rri:
   case X86::VCMPPDZrrik:
   case X86::VCMPPSZrrik:
+  case X86::VCMPPHZrrik:
   case X86::VCMPPDZ128rrik:
   case X86::VCMPPSZ128rrik:
+  case X86::VCMPPHZ128rrik:
   case X86::VCMPPDZ256rrik:
-  case X86::VCMPPSZ256rrik: {
+  case X86::VCMPPSZ256rrik:
+  case X86::VCMPPHZ256rrik: {
     unsigned OpOffset = X86II::isKMasked(Desc.TSFlags) ? 1 : 0;
 
     // Float comparison can be safely commuted for
diff --git a/llvm/test/CodeGen/X86/pr159723.ll b/llvm/test/CodeGen/X86/pr159723.ll
index cab4abb043639..c66b101fff990 100644
--- a/llvm/test/CodeGen/X86/pr159723.ll
+++ b/llvm/test/CodeGen/X86/pr159723.ll
@@ -17,7 +17,7 @@ define <8 x i1> @test_cmp_v8half_ogt(<8 x half> %rhs, <8 x i1> %mask) nounwind {
 ; CHECK-NEXT:    kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
 ; CHECK-NEXT:    callq test_call_8 at PLT
 ; CHECK-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
-; CHECK-NEXT:    vcmpltph {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %k0 {%k1} # 16-byte Folded Reload
+; CHECK-NEXT:    vcmpgtph {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %k0 {%k1} # 16-byte Folded Reload
 ; CHECK-NEXT:    vpmovm2w %k0, %xmm0
 ; CHECK-NEXT:    addq $40, %rsp
 ; CHECK-NEXT:    retq
@@ -79,7 +79,7 @@ define <16 x i1> @test_cmp_v16half_olt_commute(<16 x half> %rhs, <16 x i1> %mask
 ; CHECK-NEXT:    kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
 ; CHECK-NEXT:    callq test_call_16 at PLT
 ; CHECK-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
-; CHECK-NEXT:    vcmpltph {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %k0 {%k1} # 32-byte Folded Reload
+; CHECK-NEXT:    vcmpgtph {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %k0 {%k1} # 32-byte Folded Reload
 ; CHECK-NEXT:    vpmovm2b %k0, %xmm0
 ; CHECK-NEXT:    addq $56, %rsp
 ; CHECK-NEXT:    vzeroupper
@@ -100,7 +100,7 @@ define <32 x i1> @test_cmp_v32half_oge(<32 x half> %rhs, <32 x i1> %mask) nounwi
 ; CHECK-NEXT:    kmovd %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
 ; CHECK-NEXT:    callq test_call_32 at PLT
 ; CHECK-NEXT:    kmovd {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 4-byte Reload
-; CHECK-NEXT:    vcmpleph {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %k0 {%k1} # 64-byte Folded Reload
+; CHECK-NEXT:    vcmpgeph {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %k0 {%k1} # 64-byte Folded Reload
 ; CHECK-NEXT:    vpmovm2b %k0, %ymm0
 ; CHECK-NEXT:    addq $88, %rsp
 ; CHECK-NEXT:    retq

``````````

</details>


https://github.com/llvm/llvm-project/pull/160304