AVX-512: fixed a bug in lowering VSELECT for 512-bit vector

https://llvm.org/bugs/show_bug.cgi?id=23634

llvm-svn: 238195

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index ed9e101..c85676e 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -21710,7 +21710,7 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
   // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
   if ((N->getOpcode() == ISD::VSELECT ||
        N->getOpcode() == X86ISD::SHRUNKBLEND) &&
-      !DCI.isBeforeLegalize()) {
+      !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
     SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
     if (Shuffle.getNode())
       return Shuffle;

diff --git a/llvm/test/CodeGen/X86/avx512-bugfix-23634.ll b/llvm/test/CodeGen/X86/avx512-bugfix-23634.ll
new file mode 100644 (file)
index 0000000..c31a13a
--- /dev/null
+++ b/llvm/test/CodeGen/X86/avx512-bugfix-23634.ll
@@ -0,0 +1,35 @@
+; RUN: llc < %s -march=x86-64 -mtriple=x86_64-apple-darwin -mcpu=knl | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; CHECK-LABEL: f_fu
+; CHECK-NOT: vpblend
+; CHECK: vmovdqa32 {{.*}} {%k1}
+
+define void @f_fu(float* %ret, float*  %aa, float %b) {
+allocas:
+  %ptr_cast_for_load = bitcast float* %aa to <16 x float>*
+  %ptr_masked_load.39 = load <16 x float>, <16 x float>* %ptr_cast_for_load, align 4
+  %b_load_to_int32 = fptosi float %b to i32
+  %b_load_to_int32_broadcast_init = insertelement <16 x i32> undef, i32 %b_load_to_int32, i32 0
+  %b_load_to_int32_broadcast = shufflevector <16 x i32> %b_load_to_int32_broadcast_init, <16 x i32> undef, <16 x i32> zeroinitializer
+  %b_to_int32 = fptosi float %b to i32
+  %b_to_int32_broadcast_init = insertelement <16 x i32> undef, i32 %b_to_int32, i32 0
+  %b_to_int32_broadcast = shufflevector <16 x i32> %b_to_int32_broadcast_init, <16 x i32> undef, <16 x i32> zeroinitializer
+
+  %a_load_to_int32 = fptosi <16 x float> %ptr_masked_load.39 to <16 x i32>
+  %div_v019_load_ = sdiv <16 x i32> %b_to_int32_broadcast, <i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2>
+
+  %v1.i = select <16 x i1> <i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true>, <16 x i32> <i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17>, <16 x i32> %a_load_to_int32
+
+  %foo_test = add <16 x i32> %div_v019_load_, %b_load_to_int32_broadcast 
+
+
+  %add_struct_offset_y_struct_offset33_x = add <16 x i32> %foo_test, %v1.i 
+
+  %val = sitofp <16 x i32> %add_struct_offset_y_struct_offset33_x to <16 x float>
+  %ptrcast = bitcast float* %ret to <16 x float>*
+  store <16 x float> %val, <16 x float>* %ptrcast, align 4
+  ret void
+}
\ No newline at end of file