[X86][SSE] combineX86ShufflesRecursively - bail if constant folding fails due to oneuse limits.

Fixes issue reported on D105827 where a single shuffle of a constant (with multiple uses) was caught in an infinite loop where one shuffle (UNPCKL) used an undef arg but then that got recombined to SHUFPS as the constant value had its own undef that confused matching.....

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -37085,6 +37085,18 @@ static SDValue combineX86ShufflesRecursively(
           Ops, Mask, Root, HasVariableMask, DAG, Subtarget))
     return Cst;
 
+  // If constant fold failed and we only have constants - then we have
+  // multiple uses by a single non-variable shuffle - just bail.
+  if (Depth == 0 && llvm::all_of(Ops, [&](SDValue Op) {
+        APInt UndefElts;
+        SmallVector<APInt> RawBits;
+        unsigned EltSizeInBits = RootSizeInBits / Mask.size();
+        return getTargetConstantBitsFromNode(Op, EltSizeInBits, UndefElts,
+                                             RawBits);
+      })) {
+    return SDValue();
+  }
+
   // Canonicalize the combined shuffle mask chain with horizontal ops.
   // NOTE: This will update the Ops and Mask.
   if (SDValue HOp = canonicalizeShuffleMaskWithHorizOp(

llvm/test/CodeGen/X86/vector-shuffle-combining.ll

@@ -3287,3 +3287,131 @@ define void @PR45604(<32 x i16>* %dst, <8 x i16>* %src) {
   store <32 x i16> %v3, <32 x i16>* %dst, align 16
   ret void
 }
+
+; Test case reported on D105827
+define void @SpinningCube() {
+; SSE2-LABEL: SpinningCube:
+; SSE2:       # %bb.0: # %entry
+; SSE2-NEXT:    movl $1065353216, (%rax) # imm = 0x3F800000
+; SSE2-NEXT:    movaps {{.*#+}} xmm0 = <u,u,u,1.0E+0>
+; SSE2-NEXT:    movaps {{.*#+}} xmm1 = <0.0E+0,-2.0E+0,u,u>
+; SSE2-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE2-NEXT:    movaps %xmm2, %xmm3
+; SSE2-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm1[1,3]
+; SSE2-NEXT:    xorps %xmm4, %xmm4
+; SSE2-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm3[2,0]
+; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm0[2,3]
+; SSE2-NEXT:    addps %xmm4, %xmm2
+; SSE2-NEXT:    movaps %xmm2, (%rax)
+; SSE2-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0,0,0]
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0,1,3]
+; SSE2-NEXT:    mulps %xmm2, %xmm1
+; SSE2-NEXT:    addps %xmm0, %xmm1
+; SSE2-NEXT:    movaps %xmm1, (%rax)
+; SSE2-NEXT:    retq
+;
+; SSSE3-LABEL: SpinningCube:
+; SSSE3:       # %bb.0: # %entry
+; SSSE3-NEXT:    movl $1065353216, (%rax) # imm = 0x3F800000
+; SSSE3-NEXT:    movaps {{.*#+}} xmm0 = <u,u,u,1.0E+0>
+; SSSE3-NEXT:    movaps {{.*#+}} xmm1 = <0.0E+0,-2.0E+0,u,u>
+; SSSE3-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSSE3-NEXT:    movaps %xmm2, %xmm3
+; SSSE3-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm1[1,3]
+; SSSE3-NEXT:    xorps %xmm4, %xmm4
+; SSSE3-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm3[2,0]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm0[2,3]
+; SSSE3-NEXT:    addps %xmm4, %xmm2
+; SSSE3-NEXT:    movaps %xmm2, (%rax)
+; SSSE3-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0,1,3]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0,0,2]
+; SSSE3-NEXT:    mulps %xmm1, %xmm2
+; SSSE3-NEXT:    addps %xmm0, %xmm2
+; SSSE3-NEXT:    movaps %xmm2, (%rax)
+; SSSE3-NEXT:    retq
+;
+; SSE41-LABEL: SpinningCube:
+; SSE41:       # %bb.0: # %entry
+; SSE41-NEXT:    movl $1065353216, (%rax) # imm = 0x3F800000
+; SSE41-NEXT:    movaps {{.*#+}} xmm0 = <u,u,u,1.0E+0>
+; SSE41-NEXT:    movaps {{.*#+}} xmm1 = <0.0E+0,-2.0E+0,u,u>
+; SSE41-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0,1,3]
+; SSE41-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE41-NEXT:    movaps %xmm1, %xmm3
+; SSE41-NEXT:    insertps {{.*#+}} xmm3 = xmm3[0,1,2],xmm2[0]
+; SSE41-NEXT:    movaps %xmm0, %xmm4
+; SSE41-NEXT:    insertps {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[2,3]
+; SSE41-NEXT:    addps %xmm3, %xmm4
+; SSE41-NEXT:    movaps %xmm4, (%rax)
+; SSE41-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE41-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0,0,2]
+; SSE41-NEXT:    mulps %xmm1, %xmm2
+; SSE41-NEXT:    addps %xmm0, %xmm2
+; SSE41-NEXT:    movaps %xmm2, (%rax)
+; SSE41-NEXT:    retq
+;
+; AVX1-LABEL: SpinningCube:
+; AVX1:       # %bb.0: # %entry
+; AVX1-NEXT:    movl $1065353216, (%rax) # imm = 0x3F800000
+; AVX1-NEXT:    vmovaps {{.*#+}} xmm0 = <u,u,u,1.0E+0>
+; AVX1-NEXT:    vmovaps {{.*#+}} xmm1 = <0.0E+0,-2.0E+0,u,u>
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm2 = xmm1[0,0,1,3]
+; AVX1-NEXT:    vmovss {{.*#+}} xmm3 = mem[0],zero,zero,zero
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm2 = xmm2[0,1,2],xmm3[0]
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm3 = xmm0[0],xmm3[0],xmm0[2,3]
+; AVX1-NEXT:    vaddps %xmm3, %xmm2, %xmm2
+; AVX1-NEXT:    vmovaps %xmm2, (%rax)
+; AVX1-NEXT:    vbroadcastss (%rax), %xmm2
+; AVX1-NEXT:    vmulps %xmm1, %xmm2, %xmm1
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm1 = xmm1[0,0,1,3]
+; AVX1-NEXT:    vaddps %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, (%rax)
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: SpinningCube:
+; AVX2:       # %bb.0: # %entry
+; AVX2-NEXT:    movl $1065353216, (%rax) # imm = 0x3F800000
+; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm0 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
+; AVX2-NEXT:    vmovaps {{.*#+}} xmm1 = <0.0E+0,-2.0E+0,u,u>
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm2 = xmm1[0,0,1,3]
+; AVX2-NEXT:    vmovss {{.*#+}} xmm3 = mem[0],zero,zero,zero
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm2 = xmm2[0,1,2],xmm3[0]
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm3 = xmm0[0],xmm3[0],xmm0[2,3]
+; AVX2-NEXT:    vaddps %xmm3, %xmm2, %xmm2
+; AVX2-NEXT:    vmovaps %xmm2, (%rax)
+; AVX2-NEXT:    vbroadcastss (%rax), %xmm2
+; AVX2-NEXT:    vmulps %xmm1, %xmm2, %xmm1
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm1 = xmm1[0,0,1,3]
+; AVX2-NEXT:    vaddps %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, (%rax)
+; AVX2-NEXT:    retq
+entry:
+  store float 1.000000e+00, float* undef, align 4
+  %0 = load float, float* undef, align 4
+  %1 = fmul float undef, 0.000000e+00
+  %2 = insertelement <4 x float> poison, float %0, i32 3
+  %3 = load float, float* undef, align 4
+  %4 = insertelement <2 x float> poison, float %3, i32 0
+  %5 = shufflevector <2 x float> %4, <2 x float> poison, <2 x i32> zeroinitializer
+  %6 = fmul <2 x float> %5, <float 0.000000e+00, float -2.000000e+00>
+  %7 = fadd float %1, undef
+  %8 = shufflevector <2 x float> %6, <2 x float> poison, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
+  %9 = shufflevector <4 x float> undef, <4 x float> %8, <4 x i32> <i32 0, i32 4, i32 5, i32 undef>
+  %10 = insertelement <4 x float> %9, float %7, i32 3
+  %11 = insertelement <4 x float> %2, float 0x7FF8000000000000, i32 1
+  %12 = insertelement <4 x float> %11, float undef, i32 0
+  %13 = insertelement <4 x float> %12, float undef, i32 2
+  %14 = fadd <4 x float> %10, %13
+  store <4 x float> %14, <4 x float>* undef, align 16
+  %15 = load float, float* undef, align 4
+  %16 = insertelement <2 x float> poison, float %15, i32 0
+  %17 = shufflevector <2 x float> %16, <2 x float> poison, <2 x i32> zeroinitializer
+  %18 = fmul <2 x float> %17, <float 0.000000e+00, float -2.000000e+00>
+  %19 = shufflevector <2 x float> %18, <2 x float> poison, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
+  %20 = shufflevector <4 x float> undef, <4 x float> %19, <4 x i32> <i32 0, i32 4, i32 5, i32 undef>
+  %21 = fadd <4 x float> %20, %2
+  store <4 x float> %21, <4 x float>* undef, align 16
+  ret void
+}