[llvm] [X86][FP16] Do not combine to ADDSUB if target doesn't support FP16 (PR #69109)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-x86

Author: Phoebe Wang (phoebewang)

<details>
<summary>Changes</summary>

Fix crash when build code with `-mattr=f16c,fma` or `-mattr=avx512vl`.

---
Full diff: https://github.com/llvm/llvm-project/pull/69109.diff


2 Files Affected:

- (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+1-1) 
- (added) llvm/test/CodeGen/X86/avx512fp16-combine-shuffle-fma.ll (+58) 


``````````diff
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 13684babb2385ea..66b6d8260b7c777 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -40685,7 +40685,7 @@ static SDValue combineShuffle(SDNode *N, SelectionDAG &DAG,
   SDLoc dl(N);
   EVT VT = N->getValueType(0);
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
-  if (TLI.isTypeLegal(VT))
+  if (TLI.isTypeLegal(VT) && !isSoftF16(VT, Subtarget))
     if (SDValue AddSub = combineShuffleToAddSubOrFMAddSub(N, Subtarget, DAG))
       return AddSub;
 
diff --git a/llvm/test/CodeGen/X86/avx512fp16-combine-shuffle-fma.ll b/llvm/test/CodeGen/X86/avx512fp16-combine-shuffle-fma.ll
new file mode 100644
index 000000000000000..54ccc23840f99cc
--- /dev/null
+++ b/llvm/test/CodeGen/X86/avx512fp16-combine-shuffle-fma.ll
@@ -0,0 +1,58 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2 | FileCheck %s --check-prefix=AVX2
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=f16c,fma | FileCheck %s --check-prefix=F16C
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512vl | FileCheck %s --check-prefix=F16C
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512fp16 | FileCheck %s --check-prefix=FP16
+
+define <2 x half> @foo(<2 x half> %0) "unsafe-fp-math"="true" nounwind {
+; AVX2-LABEL: foo:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    subq $40, %rsp
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2 at PLT
+; AVX2-NEXT:    vmulss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; AVX2-NEXT:    callq __truncsfhf2 at PLT
+; AVX2-NEXT:    vmovss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __extendhfsf2 at PLT
+; AVX2-NEXT:    vmovss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
+; AVX2-NEXT:    vmovss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
+; AVX2-NEXT:    # xmm0 = mem[0],zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2 at PLT
+; AVX2-NEXT:    vsubss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 4-byte Folded Reload
+; AVX2-NEXT:    callq __truncsfhf2 at PLT
+; AVX2-NEXT:    addq $40, %rsp
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: foo:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vpsrld $16, %xmm0, %xmm1
+; F16C-NEXT:    vcvtph2ps %xmm1, %ymm1
+; F16C-NEXT:    vmulps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; F16C-NEXT:    vcvtps2ph $4, %ymm1, %xmm1
+; F16C-NEXT:    vcvtph2ps %xmm0, %ymm0
+; F16C-NEXT:    vcvtph2ps %xmm1, %ymm1
+; F16C-NEXT:    vsubps %ymm0, %ymm1, %ymm2
+; F16C-NEXT:    vcvtps2ph $4, %ymm2, %xmm2
+; F16C-NEXT:    vaddps %ymm0, %ymm1, %ymm0
+; F16C-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; F16C-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2,3,4,5,6,7]
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    retq
+;
+; FP16-LABEL: foo:
+; FP16:       # %bb.0:
+; FP16-NEXT:    vpsrld $16, %xmm0, %xmm1
+; FP16-NEXT:    vfmaddsub231ph {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm0
+; FP16-NEXT:    retq
+  %2 = shufflevector <2 x half> %0, <2 x half> undef, <2 x i32> <i32 1, i32 2>
+  %3 = fmul fast <2 x half> %2, <half 0xH3D3A, half 0xH3854>
+  %4 = fsub fast <2 x half> %3, %0
+  %5 = fadd fast <2 x half> %3, %0
+  %6 = shufflevector <2 x half> %4, <2 x half> %5, <2 x i32> <i32 0, i32 3>
+  %7 = fadd fast <2 x half> %6, zeroinitializer
+  %8 = shufflevector <2 x half> undef, <2 x half> %7, <2 x i32> <i32 0, i32 3>
+  %9 = fsub fast <2 x half> %8, zeroinitializer
+  ret <2 x half> %9
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/69109