From da0fd933156e20c02e276f62bfefb27e40b0b360 Mon Sep 17 00:00:00 2001
From: Matt Pharr <matt.pharr@intel.com>
Date: Fri, 22 Jul 2011 12:36:44 +0100
Subject: [PATCH] AVX fixes: add missing 8/16-bit gathers and scatters, set
 features string appropriately when AVX is enabled.

---
 module.cpp    | 12 +++++++-----
 stdlib-avx.ll |  6 +++++-
 2 files changed, 12 insertions(+), 6 deletions(-)

diff --git a/module.cpp b/module.cpp
index cc9e0b7c..56162ba9 100644
--- a/module.cpp
+++ b/module.cpp
@@ -949,10 +949,12 @@ Module::writeObjectFileOrAssembly(OutputType outputType, const char *outFileName
     }
 
     std::string featuresString;
+    llvm::TargetMachine *targetMachine = NULL;
 #if defined LLVM_3_0svn || defined LLVM_3_0
-    llvm::TargetMachine *targetMachine = 
-        target->createTargetMachine(triple.getTriple(), g->target.cpu,
-                                    featuresString);
+    if (g->target.isa == Target::AVX)
+        featuresString = "+avx";
+    targetMachine = target->createTargetMachine(triple.getTriple(), g->target.cpu,
+                                                featuresString);
 #else
     if (g->target.cpu.size()) {
         llvm::SubtargetFeatures features;
@@ -960,8 +962,8 @@ Module::writeObjectFileOrAssembly(OutputType outputType, const char *outFileName
         featuresString = features.getString();
     }
 
-    llvm::TargetMachine *targetMachine = 
-        target->createTargetMachine(triple.getTriple(), featuresString);
+    targetMachine = target->createTargetMachine(triple.getTriple(), 
+                                                featuresString);
 #endif
     if (targetMachine == NULL) {
         fprintf(stderr, "Unable to create target machine for target \"%s\"!",
diff --git a/stdlib-avx.ll b/stdlib-avx.ll
index de7f87fa..897c55a0 100644
--- a/stdlib-avx.ll
+++ b/stdlib-avx.ll
@@ -698,12 +698,16 @@ define void @__masked_store_blend_64(<8 x i64>* nocapture %ptr, <8 x i64> %new,
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; gather/scatter
 
+gen_gather(8, i8)
+gen_gather(8, i16)
 gen_gather(8, i32)
 gen_gather(8, i64)
+
+gen_scatter(8, i8)
+gen_scatter(8, i16)
 gen_scatter(8, i32)
 gen_scatter(8, i64)
 
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; double precision sqrt