[X86] getTargetConstantBitsFromNode - ensure we truncate broadcasts from wider constants.

We might be broadcasting the lower portion of a wider constant pool entry, ensure we truncate the full width of the constant back to the broadcast width.

(This also exposes an issue with how asm comments print these constants - I'll address that in a followup).

Fixes #63030

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 1126870..9a201b7 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -7542,6 +7542,8 @@ static bool getTargetConstantBitsFromNode(SDValue Op, unsigned EltSizeInBits,
       if (CollectConstantBits(C, SrcEltBits[0], UndefSrcElts, 0)) {
         if (UndefSrcElts[0])
           UndefSrcElts.setBits(0, NumSrcElts);
+        if (SrcEltBits[0].getBitWidth() != SrcEltSizeInBits)
+          SrcEltBits[0] = SrcEltBits[0].trunc(SrcEltSizeInBits);
         SrcEltBits.append(NumSrcElts - 1, SrcEltBits[0]);
         return CastBitData(UndefSrcElts, SrcEltBits);
       }

diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
index a6dbbfb..91f550b 100644 (file)
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
@@ -1,8 +1,8 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=i686-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=CHECK,X86,AVX2
-; RUN: llc < %s -mtriple=i686-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=CHECK,X86,AVX512
-; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=CHECK,X64,AVX2
-; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=CHECK,X64,AVX512
+; RUN: llc < %s -mtriple=i686-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=CHECK,X86,AVX2,X86-AVX2
+; RUN: llc < %s -mtriple=i686-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=CHECK,X86,AVX512,X86-AVX512
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=CHECK,X64,AVX2,X64-AVX2
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=CHECK,X64,AVX512,X64-AVX512
 
 declare <8 x i32> @llvm.x86.avx2.permd(<8 x i32>, <8 x i32>)
 declare <8 x float> @llvm.x86.avx2.permps(<8 x float>, <8 x i32>)
@@ -869,6 +869,62 @@ define <32 x i8> @PR52122(<32 x i8> %0, <32 x i8> %1) {
   ret <32 x i8> %5
 }
 
+define void @PR63030(ptr %p0) {
+; X86-AVX2-LABEL: PR63030:
+; X86-AVX2:       # %bb.0:
+; X86-AVX2-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-AVX2-NEXT:    vmovaps (%eax), %xmm0
+; X86-AVX2-NEXT:    vmovddup {{.*#+}} xmm1 = [3,0,2,0,3,0,2,0]
+; X86-AVX2-NEXT:    # xmm1 = mem[0,0]
+; X86-AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm0[1,1,0,0]
+; X86-AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
+; X86-AVX2-NEXT:    vmovaps {{.*#+}} xmm2 = [3,0,2,0]
+; X86-AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,1]
+; X86-AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
+; X86-AVX2-NEXT:    vmovaps %ymm0, (%eax)
+; X86-AVX2-NEXT:    vmovaps %ymm1, (%eax)
+; X86-AVX2-NEXT:    vzeroupper
+; X86-AVX2-NEXT:    retl
+;
+; X86-AVX512-LABEL: PR63030:
+; X86-AVX512:       # %bb.0:
+; X86-AVX512-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512-NEXT:    vmovdqa (%eax), %xmm0
+; X86-AVX512-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [1,0,8,0,0,0,0,0,0,0,9,0,1,0,1,0]
+; X86-AVX512-NEXT:    vpermi2q {{\.?LCPI[0-9]+_[0-9]+}}, %zmm0, %zmm1
+; X86-AVX512-NEXT:    vmovdqa64 %zmm1, (%eax)
+; X86-AVX512-NEXT:    vzeroupper
+; X86-AVX512-NEXT:    retl
+;
+; X64-AVX2-LABEL: PR63030:
+; X64-AVX2:       # %bb.0:
+; X64-AVX2-NEXT:    vmovaps (%rdi), %xmm0
+; X64-AVX2-NEXT:    vmovddup {{.*#+}} xmm1 = [3,2,3,2]
+; X64-AVX2-NEXT:    # xmm1 = mem[0,0]
+; X64-AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm0[1,1,0,0]
+; X64-AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
+; X64-AVX2-NEXT:    vmovaps {{.*#+}} xmm2 = [3,2]
+; X64-AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,1]
+; X64-AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
+; X64-AVX2-NEXT:    vmovaps %ymm0, (%rax)
+; X64-AVX2-NEXT:    vmovaps %ymm1, (%rax)
+; X64-AVX2-NEXT:    vzeroupper
+; X64-AVX2-NEXT:    retq
+;
+; X64-AVX512-LABEL: PR63030:
+; X64-AVX512:       # %bb.0:
+; X64-AVX512-NEXT:    vmovdqa (%rdi), %xmm0
+; X64-AVX512-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [1,8,0,0,0,9,1,1]
+; X64-AVX512-NEXT:    vpermi2q {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
+; X64-AVX512-NEXT:    vmovdqa64 %zmm1, (%rax)
+; X64-AVX512-NEXT:    vzeroupper
+; X64-AVX512-NEXT:    retq
+  %load = load <2 x i64>, ptr %p0, align 16
+  %shuffle = shufflevector <2 x i64> <i64 3, i64 2>, <2 x i64> %load, <8 x i32> <i32 3, i32 0, i32 2, i32 2, i32 2, i32 1, i32 3, i32 3>
+  store volatile <8 x i64> %shuffle, ptr poison, align 64
+  ret void
+}
+
 define void @packss_zext_v8i1() {
 ; X86-LABEL: packss_zext_v8i1:
 ; X86:       # %bb.0: