Fix instruction order issue in scalarizer (microsoft#5001)

When scalarize shuffle inst which has extract element as the input element, create clone before the shuffle instead of after. Because later when replace the shuffle, new vector will be inserted before the shuffle. The elt define will be after its use if inserted before the shuffle earlier. Like %[[B:.+]] = load <2 x float>, <2 x float>* %b, align 4 %[[X:.+]] = insertelement <4 x float> undef, float %[[BX0]], i32 0 %[[Y:.+]] = insertelement <4 x float> %[[X]], float %[[BY0]], i32 1 %[[Z:.+]] = insertelement <4 x float> %[[Y]], float %[[BX1]], i32 2 %[[W:.+]] = insertelement <4 x float> %[[Z]], float %[[BY1]], i32 3 %[[BX0:.+]] = extractelement <2 x float> %[[B]], i32 0 %[[BY0:.+]] = extractelement <2 x float> %[[B]], i32 1 %[[BX1:.+]] = extractelement <2 x float> %[[B]], i32 0 %[[BY1:.+]] = extractelement <2 x float> %[[B]], i32 1
pow2clk · Feb 9, 2023 · b3dedc9 · b3dedc9
1 parent 40d5e1f
commit b3dedc9
Show file tree

Hide file tree

Showing 2 changed files with 34 additions and 1 deletion.
diff --git a/lib/Transforms/Scalar/Scalarizer.cpp b/lib/Transforms/Scalar/Scalarizer.cpp
@@ -633,7 +633,7 @@ bool Scalarizer::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
       // instruction is processed, it will be replaced without updating our
       // Gather entry.  This dead instruction will be accessed by finish(),
       // causing assert or crash.
-      Res[I] = IRBuilder<>(SVI.getNextNode()).Insert(EA->clone());
+      Res[I] = IRBuilder<>(&SVI).Insert(EA->clone());
     }
     // HLSL Change Ends
   }

diff --git a/tools/clang/test/HLSLFileCheckLit/passes/llvm/scalarizer/shuffle_use_extract_elt.ll b/tools/clang/test/HLSLFileCheckLit/passes/llvm/scalarizer/shuffle_use_extract_elt.ll
@@ -0,0 +1,33 @@
+; RUN: opt -S -scalarizer -dce %s | FileCheck %s
+
+; CHECK: %[[B:.+]] = load <2 x float>, <2 x float>* %b, align 4
+; CHECK: %[[BX0:.+]] = extractelement <2 x float> %[[B]], i32 0
+; CHECK: %[[BY0:.+]] = extractelement <2 x float> %[[B]], i32 1
+; CHECK: %[[BX1:.+]] = extractelement <2 x float> %[[B]], i32 0
+; CHECK: %[[BY1:.+]] = extractelement <2 x float> %[[B]], i32 1
+
+; CHECK: %[[X:.+]] = insertelement <4 x float> undef, float %[[BX0]], i32 0
+; CHECK: %[[Y:.+]] = insertelement <4 x float> %[[X]], float %[[BY0]], i32 1
+; CHECK: %[[Z:.+]] = insertelement <4 x float> %[[Y]], float %[[BX1]], i32 2
+; CHECK: %[[W:.+]] = insertelement <4 x float> %[[Z]], float %[[BY1]], i32 3
+; CHECK: ret <4 x float> %[[W]]
+
+declare void @foo(<2 x float>, <2 x float>* dereferenceable(8))
+
+; Function Attrs: noinline nounwind
+define internal <4 x float> @bar(<3 x float> %v) #0 {
+entry:
+  %0 = alloca <2 x float>
+  %b = alloca <2 x float>, align 4
+  store <2 x float> zeroinitializer, <2 x float>* %b, align 4
+  %1 = insertelement <3 x float> %v, float 1.000000e+00, i32 0
+  %2 = shufflevector <3 x float> %1, <3 x float> undef, <2 x i32> <i32 0, i32 1>
+  store <2 x float> %2, <2 x float>* %0
+  ;call void @foo(<2 x float>* dereferenceable(8) %0, <2 x float>* dereferenceable(8) %b)
+  %3 = load <2 x float>, <2 x float>* %b, align 4
+  %4 = shufflevector <2 x float> %3, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
+  ret <4 x float> %4
+}
+
+attributes #0 = { noinline nounwind }
+