[X86] Prevent shuffle combining from creating an identical X86ISD::SHUF128.

This can cause an infinite loop if SimplifiedDemandedElts asks
for the node to replace itself.

A similar protection exists in other places in shuffle combining.

Fixes ISPC https://github.com/ispc/ispc/issues/1864

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 517e6c0..1212585 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -34909,6 +34909,8 @@ static SDValue combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root,
         (Mask[1] < 0 || Mask[3] < 0 || Mask[1] == (Mask[3] % 2));
 
     if (!isAnyZero(Mask) && !PreferPERMQ) {
+      if (Depth == 0 && Root.getOpcode() == X86ISD::SHUF128)
+        return SDValue(); // Nothing to do!
       if (SDValue V = MatchSHUF128(ShuffleVT, DL, Mask, V1, V2, DAG))
         return DAG.getBitcast(RootVT, V);
     }

diff --git a/llvm/test/CodeGen/X86/vector-shuffle-512-v16.ll b/llvm/test/CodeGen/X86/vector-shuffle-512-v16.ll
index e9f4aa9..4fce1a3 100644 (file)
--- a/llvm/test/CodeGen/X86/vector-shuffle-512-v16.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-512-v16.ll
@@ -764,3 +764,47 @@ define <16 x float> @mask_shuffle_v4f32_v16f32_00_01_02_03_00_01_02_03_00_01_02_
   %res = shufflevector <4 x float> %a, <4 x float> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
   ret <16 x float> %res
 }
+
+%struct.foo = type { [4 x double], [3 x [4 x double]], [4 x double] }
+
+; This test previously hung in shuffle combining. https://github.com/ispc/ispc/issues/1864
+define void @ispc_1864(<16 x float>* %arg) {
+; ALL-LABEL: ispc_1864:
+; ALL:       # %bb.0: # %bb
+; ALL-NEXT:    pushq %rbp
+; ALL-NEXT:    .cfi_def_cfa_offset 16
+; ALL-NEXT:    .cfi_offset %rbp, -16
+; ALL-NEXT:    movq %rsp, %rbp
+; ALL-NEXT:    .cfi_def_cfa_register %rbp
+; ALL-NEXT:    andq $-64, %rsp
+; ALL-NEXT:    subq $4864, %rsp # imm = 0x1300
+; ALL-NEXT:    vbroadcastss {{.*#+}} ymm0 = [-5.0E+0,-5.0E+0,-5.0E+0,-5.0E+0,-5.0E+0,-5.0E+0,-5.0E+0,-5.0E+0]
+; ALL-NEXT:    vmulps 32(%rdi), %ymm0, %ymm0
+; ALL-NEXT:    vcvtps2pd %ymm0, %zmm0
+; ALL-NEXT:    vshuff64x2 {{.*#+}} zmm0 = zmm0[2,3,4,5,0,1,0,1]
+; ALL-NEXT:    vmovapd %ymm0, {{[0-9]+}}(%rsp)
+; ALL-NEXT:    movq %rbp, %rsp
+; ALL-NEXT:    popq %rbp
+; ALL-NEXT:    .cfi_def_cfa %rsp, 8
+; ALL-NEXT:    vzeroupper
+; ALL-NEXT:    retq
+bb:
+  %tmp = alloca [30 x %struct.foo], align 64
+  %tmp1 = load <16 x float>, <16 x float>* %arg, align 4
+  %tmp2 = fmul <16 x float> %tmp1, <float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00>
+  %tmp3 = fpext <16 x float> %tmp2 to <16 x double>
+  %tmp4 = getelementptr inbounds [30 x %struct.foo], [30 x %struct.foo]* %tmp, i64 0, i64 3, i32 2, i64 0
+  %tmp5 = extractelement <16 x double> %tmp3, i32 10
+  store double %tmp5, double* %tmp4, align 32
+  %tmp6 = getelementptr inbounds [30 x %struct.foo], [30 x %struct.foo]* %tmp, i64 0, i64 3, i32 2, i64 1
+  %tmp7 = extractelement <16 x double> %tmp3, i32 11
+  store double %tmp7, double* %tmp6, align 8
+  %tmp8 = getelementptr inbounds [30 x %struct.foo], [30 x %struct.foo]* %tmp, i64 0, i64 3, i32 2, i64 2
+  %tmp9 = extractelement <16 x double> %tmp3, i32 12
+  store double %tmp9, double* %tmp8, align 16
+  %tmp10 = getelementptr inbounds [30 x %struct.foo], [30 x %struct.foo]* %tmp, i64 0, i64 3, i32 2, i64 3
+  %tmp11 = extractelement <16 x double> %tmp3, i32 13
+  store double %tmp11, double* %tmp10, align 8
+  ret void
+}
+