patch for LLVM 3.3 and test correction at avx2

llvm_patches/3_3_PR17764_reverse_operands_avx2.patch

@@ -0,0 +1,51 @@
+From 13c33dd2931ae9d9c5c9f142677f025281fbefca Mon Sep 17 00:00:00 2001
+From: Michael Liao <michael.hliao@gmail.com>
+Date: Fri, 1 Nov 2013 11:08:08 -0700
+Subject: [PATCH] Fix PR17764
+
+- %ret = select %mask, %v1, %v2 is equivalent to
+
+    %ret = %mask ? %v1 : %v2
+
+  but VPBLENDVB %mask, %v1, %v2, %ret (operands are in Intel assembly
+  order) is equivalent to
+
+    %ret = %mask ? %v2 : %v1
+---
+ lib/Target/X86/X86InstrSSE.td |  2 +-
+ test/CodeGen/X86/pr17764.ll   | 10 ++++++++++
+ 2 files changed, 11 insertions(+), 1 deletion(-)
+ create mode 100644 test/CodeGen/X86/pr17764.ll
+
+diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
+index 7cae485..bac88f9 100644
+--- lib/Target/X86/X86InstrSSE.td
++++ lib/Target/X86/X86InstrSSE.td
+@@ -6965,7 +6965,7 @@ let Predicates = [HasAVX] in {
+ let Predicates = [HasAVX2] in {
+   def : Pat<(v32i8 (vselect (v32i8 VR256:$mask), (v32i8 VR256:$src1),
+                             (v32i8 VR256:$src2))),
+-            (VPBLENDVBYrr VR256:$src1, VR256:$src2, VR256:$mask)>;
++            (VPBLENDVBYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
+   def : Pat<(v16i16 (X86Blendi (v16i16 VR256:$src1), (v16i16 VR256:$src2),
+                                (imm:$mask))),
+             (VPBLENDWYrri VR256:$src1, VR256:$src2, imm:$mask)>;
+diff --git a/test/CodeGen/X86/pr17764.ll b/test/CodeGen/X86/pr17764.ll
+new file mode 100644
+index 0000000..7a3fd6d
+--- /dev/null
++++ test/CodeGen/X86/pr17764.ll
+@@ -0,0 +1,10 @@
++; RUN: llc < %s -mtriple=x86_64-linux -mcpu=core-avx2 | FileCheck %s
++
++define <16 x i16> @foo(<16 x i1> %mask, <16 x i16> %x, <16 x i16> %y) {
++  %ret = select <16 x i1> %mask, <16 x i16> %x, <16 x i16> %y
++  ret <16 x i16> %ret
++}
++
++; CHECK: foo
++; CHECK: vpblendvb %ymm0, %ymm1, %ymm2, %ymm0
++; CHECK: ret
+-- 
+1.8.1.2
+

tests/test-141.ispc

@@ -3,8 +3,9 @@ export uniform int width() { return programCount; }
 
 
 export void f_f(uniform float RET[], uniform float aFOO[]) {
-    float a = aFOO[programIndex]; 
-    RET[programIndex] = (exp(-log(1/a)) - a) < 1e-7 ? 1 : 0;
+    float a = aFOO[programIndex];
+    // calculation error 1e-6 is the same as in icc
+    RET[programIndex] = (exp(-log(1/a)) - a) < 1e-6 ? 1 : 0;
 }
 
 export void result(uniform float RET[4]) {