[llvm] r253611 - [X86][SSE4A] Fix issue with EXTRQI shuffles not starting at the correct start index.

[Simon Pilgrim via llvm-commits]

Author: rksimon
Date: Thu Nov 19 16:13:56 2015
New Revision: 253611

URL: http://llvm.org/viewvc/llvm-project?rev=253611&view=rev
Log:
[X86][SSE4A] Fix issue with EXTRQI shuffles not starting at the correct start index.

Found during stress testing.

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/test/CodeGen/X86/vector-shuffle-sse4a.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=253611&r1=253610&r2=253611&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Thu Nov 19 16:13:56 2015
@@ -7351,8 +7351,9 @@ static SDValue lowerVectorShuffleWithSSE
       SDValue &V = (M < Size ? V1 : V2);
       M = M % Size;
 
-      // All mask elements must be in the lower half.
-      if (M >= HalfSize)
+      // The extracted elements must start at a valid index and all mask
+      // elements must be in the lower half.
+      if (i > M || M >= HalfSize)
         return SDValue();
 
       if (Idx < 0 || (Src == V && Idx == (M - i))) {

Modified: llvm/trunk/test/CodeGen/X86/vector-shuffle-sse4a.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vector-shuffle-sse4a.ll?rev=253611&r1=253610&r2=253611&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/vector-shuffle-sse4a.ll (original)
+++ llvm/trunk/test/CodeGen/X86/vector-shuffle-sse4a.ll Thu Nov 19 16:13:56 2015
@@ -319,5 +319,42 @@ define <16 x i8> @shuffle_8_18_uuuuuuuuu
   ret <16 x i8> %1
 }
 
+define <16 x i8> @shuffle_uu_0_5_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu(<16 x i8> %v) {
+; BTVER1-LABEL: shuffle_uu_0_5_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu:
+; BTVER1:       # BB#0:
+; BTVER1-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[0,0,5,5,4,4,5,5,4,4,5,5,6,6,7,7]
+; BTVER1-NEXT:    retq
+;
+; BTVER2-LABEL: shuffle_uu_0_5_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu:
+; BTVER2:       # BB#0:
+; BTVER2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,0,5,5,4,4,5,5,4,4,5,5,6,6,7,7]
+; BTVER2-NEXT:    retq
+  %1 = shufflevector <16 x i8> %v, <16 x i8> zeroinitializer, <16 x i32> <i32 undef, i32 0, i32 5, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  ret <16 x i8> %1
+}
+
+define <16 x i8> @shuffle_uu_16_4_16_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu(<16 x i8> %v) {
+; BTVER1-LABEL: shuffle_uu_16_4_16_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu:
+; BTVER1:       # BB#0:
+; BTVER1-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[u],zero,xmm0[4],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
+; BTVER1-NEXT:    retq
+;
+; BTVER2-LABEL: shuffle_uu_16_4_16_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu:
+; BTVER2:       # BB#0:
+; BTVER2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[u],zero,xmm0[4],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
+; BTVER2-NEXT:    retq
+  %1 = shufflevector <16 x i8> %v, <16 x i8> zeroinitializer, <16 x i32> <i32 undef, i32 16, i32 4, i32 16, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  ret <16 x i8> %1
+}
+
+define <16 x i8> @shuffle_uu_uu_4_16_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu(<16 x i8> %v) {
+; ALL-LABEL: shuffle_uu_uu_4_16_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu_uu:
+; ALL:       # BB#0:
+; ALL-NEXT:    extrq {{.*#+}} xmm0 = xmm0[2,3,4],zero,zero,zero,zero,zero,xmm0[u,u,u,u,u,u,u,u]
+; ALL-NEXT:    retq
+  %1 = shufflevector <16 x i8> %v, <16 x i8> zeroinitializer, <16 x i32> <i32 undef, i32 undef, i32 4, i32 16, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  ret <16 x i8> %1
+}
+
 declare <2 x i64> @llvm.x86.sse4a.extrqi(<2 x i64>, i8, i8) nounwind
 declare <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64>, <2 x i64>, i8, i8) nounwind