From 6de494cfdb3ee7f7d62c8b08bcaf59685867f0bf Mon Sep 17 00:00:00 2001
From: Matt Pharr <matt.pharr@intel.com>
Date: Mon, 29 Aug 2011 16:50:59 -0700
Subject: [PATCH] Fix AVX bug introduced in
 4ab982bc16d509fd6ec5905c2d04f8a9e8ef41bc

---
 builtins-avx-x2.ll | 22 ++++++++++++++++++----
 builtins-avx.ll    | 24 ++++++++++++++++++++----
 2 files changed, 38 insertions(+), 8 deletions(-)

diff --git a/builtins-avx-x2.ll b/builtins-avx-x2.ll
index 3c380e24..d5696097 100644
--- a/builtins-avx-x2.ll
+++ b/builtins-avx-x2.ll
@@ -524,10 +524,24 @@ define void @__masked_store_64(<16 x i64>* nocapture, <16 x i64>,
 
 
 ;; FIXME: various code elsewhere in the builtins implementations makes
-;; calls to these, basically assuming that doing so is faster than doing
-;; a full call to an actual masked store, which isn't likely to be the
-;; case on AVX.  So here we provide those functions but then don't actually
-;; do what the caller asked for...
+;; calls to the 32/64 bit versions of these, basically assuming that doing
+;; so is faster than doing a full call to an actual masked store, which
+;; isn't likely to be the case on AVX.  So here we provide those functions
+;; but then don't actually do what the caller asked for...
+
+declare void @llvm.trap()
+
+define void @__masked_store_blend_8(<8 x i8>* nocapture, <8 x i8>, 
+                                    <8 x i32>) nounwind alwaysinline {
+  call void @llvm.trap()
+  ret void
+}
+
+define void @__masked_store_blend_16(<8 x i16>* nocapture, <8 x i16>, 
+                                     <8 x i32>) nounwind alwaysinline {
+  call void @llvm.trap()
+  ret void
+}
 
 define void @__masked_store_blend_32(<16 x i32>* nocapture, <16 x i32>, 
                                      <16 x i32>) nounwind alwaysinline {
diff --git a/builtins-avx.ll b/builtins-avx.ll
index e06bd87b..6b8faf39 100644
--- a/builtins-avx.ll
+++ b/builtins-avx.ll
@@ -450,10 +450,26 @@ define void @__masked_store_64(<8 x i64>* nocapture, <8 x i64>,
 
 
 ;; FIXME: various code elsewhere in the builtins implementations makes
-;; calls to these, basically assuming that doing so is faster than doing
-;; a full call to an actual masked store, which isn't likely to be the
-;; case on AVX.  So here we provide those functions but then don't actually
-;; do what the caller asked for...
+;; calls to the 32/64 bit versions of these, basically assuming that doing
+;; so is faster than doing a full call to an actual masked store, which
+;; isn't likely to be the case on AVX.  So here we provide those functions
+;; but then don't actually do what the caller asked for...
+
+declare void @llvm.trap()
+
+define void @__masked_store_blend_8(<8 x i8>* nocapture, <8 x i8>, 
+                                    <8 x i32>) nounwind alwaysinline {
+  call void @llvm.trap()
+  ret void
+}
+
+
+define void @__masked_store_blend_16(<8 x i16>* nocapture, <8 x i16>, 
+                                     <8 x i32>) nounwind alwaysinline {
+  call void @llvm.trap()
+  ret void
+}
+
 
 define void @__masked_store_blend_32(<8 x i32>* nocapture, <8 x i32>, 
                                      <8 x i32>) nounwind alwaysinline {