r242678 - [CodeGen] Flip lanes when lowering __builtin_palignr with one lane

[Benjamin Kramer]

Author: d0k
Date: Mon Jul 20 10:31:17 2015
New Revision: 242678

URL: http://llvm.org/viewvc/llvm-project?rev=242678&view=rev
Log:
[CodeGen] Flip lanes when lowering __builtin_palignr with one lane

Otherwise we'd pick the wrong lane for the resulting shuffle and
miscompile code. PR24187.

Modified:
    cfe/trunk/lib/CodeGen/CGBuiltin.cpp
    cfe/trunk/test/CodeGen/palignr.c

Modified: cfe/trunk/lib/CodeGen/CGBuiltin.cpp
URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/CodeGen/CGBuiltin.cpp?rev=242678&r1=242677&r2=242678&view=diff
==============================================================================
--- cfe/trunk/lib/CodeGen/CGBuiltin.cpp (original)
+++ cfe/trunk/lib/CodeGen/CGBuiltin.cpp Mon Jul 20 10:31:17 2015
@@ -6238,6 +6238,7 @@ Value *CodeGenFunction::EmitX86BuiltinEx
     // but less than two lanes, convert to shifting in zeroes.
     if (ShiftVal > NumLaneElts) {
       ShiftVal -= NumLaneElts;
+      Ops[1] = Ops[0];
       Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType());
     }
 

Modified: cfe/trunk/test/CodeGen/palignr.c
URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/test/CodeGen/palignr.c?rev=242678&r1=242677&r2=242678&view=diff
==============================================================================
--- cfe/trunk/test/CodeGen/palignr.c (original)
+++ cfe/trunk/test/CodeGen/palignr.c Mon Jul 20 10:31:17 2015
@@ -4,13 +4,13 @@
 #define _mm_alignr_epi8(a, b, n) (__builtin_ia32_palignr128((a), (b), (n)))
 typedef __attribute__((vector_size(16))) int int4;
 
-// CHECK: palignr
+// CHECK: palignr $15, %xmm1, %xmm0
 int4 align1(int4 a, int4 b) { return _mm_alignr_epi8(a, b, 15); }
 // CHECK: ret
 // CHECK: ret
 // CHECK-NOT: palignr
 int4 align2(int4 a, int4 b) { return _mm_alignr_epi8(a, b, 16); }
-// CHECK: psrldq
+// CHECK: psrldq $1, %xmm0
 int4 align3(int4 a, int4 b) { return _mm_alignr_epi8(a, b, 17); }
 // CHECK: xor
 int4 align4(int4 a, int4 b) { return _mm_alignr_epi8(a, b, 32); }