[PATCH] D108612: [X86] Building constant vector which element type is half will cause assertion fail.

[LiuChen via Phabricator via llvm-commits]

LiuChen3 created this revision.
Herald added subscribers: pengfei, hiraditya.
LiuChen3 requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Fix assertion fail when building con constant vector which element type is half.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D108612

Files:
  llvm/lib/Target/X86/X86ISelLowering.cpp
  llvm/test/CodeGen/X86/build_fp16_constant_vector.ll


Index: llvm/test/CodeGen/X86/build_fp16_constant_vector.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/X86/build_fp16_constant_vector.ll
@@ -0,0 +1,30 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512fp16 | FileCheck %s --check-prefixes=CHECK-32
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512fp16 | FileCheck %s --check-prefixes=CHECK-64
+
+define dso_local <32 x half> @foo(<32 x half> %a, <32 x half> %b, <32 x half> %c) {
+; CHECK-32-LABEL: foo:
+; CHECK-32:       # %bb.0: # %entry
+; CHECK-32-NEXT:    vbroadcastf64x4 {{.*#+}} zmm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0]
+; CHECK-32-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]
+; CHECK-32-NEXT:    vfmadd213ph %zmm2, %zmm3, %zmm0
+; CHECK-32-NEXT:    vfmadd213ph %zmm2, %zmm3, %zmm1
+; CHECK-32-NEXT:    vaddph %zmm1, %zmm0, %zmm0
+; CHECK-32-NEXT:    retl
+;
+; CHECK-64-LABEL: foo:
+; CHECK-64:       # %bb.0: # %entry
+; CHECK-64-NEXT:    vbroadcastf64x4 {{.*#+}} zmm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0,-1.0E+0]
+; CHECK-64-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]
+; CHECK-64-NEXT:    vfmadd213ph %zmm2, %zmm3, %zmm0
+; CHECK-64-NEXT:    vfmadd213ph %zmm2, %zmm3, %zmm1
+; CHECK-64-NEXT:    vaddph %zmm1, %zmm0, %zmm0
+; CHECK-64-NEXT:    retq
+entry:
+  %0 = tail call fast <32 x half> @llvm.fma.v32f16(<32 x half> %a, <32 x half> <half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00>, <32 x half> %c)
+  %1 = tail call fast <32 x half> @llvm.fma.v32f16(<32 x half> %b, <32 x half> <half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00>, <32 x half> %c)
+  %2 = fadd <32 x half> %0, %1
+  ret <32 x half> %2
+}
+
+declare <32 x half> @llvm.fma.v32f16(<32 x half>, <32 x half>, <32 x half>)
Index: llvm/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -9006,7 +9006,9 @@
     APInt Val = SplatValue.extractBits(ScalarSize, ScalarSize * i);
     Constant *Const;
     if (VT.isFloatingPoint()) {
-      if (ScalarSize == 32) {
+      if (ScalarSize == 16)
+        Const = ConstantFP::get(C, APFloat(APFloat::IEEEhalf(), Val));
+      else if (ScalarSize == 32) {
         Const = ConstantFP::get(C, APFloat(APFloat::IEEEsingle(), Val));
       } else {
         assert(ScalarSize == 64 && "Unsupported floating point scalar size");


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D108612.368279.patch
Type: text/x-patch
Size: 3695 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210824/1c2c8346/attachment.bin>