From d36ab4cc3c3cc66ca9511cc37dfbb320ce260031 Mon Sep 17 00:00:00 2001
From: Dmitry Babokin <babokin@gmail.com>
Date: Thu, 25 Apr 2013 20:39:01 +0400
Subject: [PATCH] Adding noalias attribute to malloc return

---
 builtins/util.m4 | 10 +++++-----
 ctx.cpp          | 14 ++++++++++++++
 2 files changed, 19 insertions(+), 5 deletions(-)

diff --git a/builtins/util.m4 b/builtins/util.m4
index 765b5587..87a1fd68 100644
--- a/builtins/util.m4
+++ b/builtins/util.m4
@@ -2536,11 +2536,11 @@ ok:
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; new/delete
 
-declare i8 * @malloc(i64)
+declare noalias i8 * @malloc(i64)
 declare void @free(i8 *)
 
-define i8 * @__new_uniform(i64 %size) {
-  %a = call i8 * @malloc(i64 %size)
+define noalias i8 * @__new_uniform(i64 %size) {
+  %a = call noalias i8 * @malloc(i64 %size)
   ret i8 * %a
 }
 
@@ -2552,7 +2552,7 @@ define <WIDTH x i64> @__new_varying32(<WIDTH x i32> %size, <WIDTH x MASK> %mask)
   per_lane(WIDTH, <WIDTH x MASK> %mask, `
     %sz_LANE_ID = extractelement <WIDTH x i32> %size, i32 LANE
     %sz64_LANE_ID = zext i32 %sz_LANE_ID to i64
-    %ptr_LANE_ID = call i8 * @malloc(i64 %sz64_LANE_ID)
+    %ptr_LANE_ID = call noalias i8 * @malloc(i64 %sz64_LANE_ID)
     %ptr_int_LANE_ID = ptrtoint i8 * %ptr_LANE_ID to i64
     %store_LANE_ID = getelementptr i64 * %ret64, i32 LANE
     store i64 %ptr_int_LANE_ID, i64 * %store_LANE_ID')
@@ -2568,7 +2568,7 @@ define <WIDTH x i64> @__new_varying64(<WIDTH x i64> %size, <WIDTH x MASK> %mask)
 
   per_lane(WIDTH, <WIDTH x MASK> %mask, `
     %sz_LANE_ID = extractelement <WIDTH x i64> %size, i32 LANE
-    %ptr_LANE_ID = call i8 * @malloc(i64 %sz_LANE_ID)
+    %ptr_LANE_ID = call noalias i8 * @malloc(i64 %sz_LANE_ID)
     %ptr_int_LANE_ID = ptrtoint i8 * %ptr_LANE_ID to i64
     %store_LANE_ID = getelementptr i64 * %ret64, i32 LANE
     store i64 %ptr_int_LANE_ID, i64 * %store_LANE_ID')
diff --git a/ctx.cpp b/ctx.cpp
index 11e25492..4d207ee2 100644
--- a/ctx.cpp
+++ b/ctx.cpp
@@ -3285,6 +3285,20 @@ FunctionEmitContext::CallInst(llvm::Value *func, const FunctionType *funcType,
         // pointer, so just emit the IR directly.
         llvm::Instruction *ci =
             llvm::CallInst::Create(func, argVals, name ? name : "", bblock);
+
+        // Copy noalias attribute to call instruction, to enable better
+        // alias analysis.
+        // TODO: what other attributes needs to be copied?
+        // TODO: do the same for varing path.
+#if defined (LLVM_3_3)
+        llvm::CallInst *cc = llvm::dyn_cast<llvm::CallInst>(ci);
+        if (cc &&
+            cc->getCalledFunction() &&
+            cc->getCalledFunction()->doesNotAlias(0)) {
+            cc->addAttribute(0, llvm::Attribute::NoAlias);
+        }
+#endif
+
         AddDebugPos(ci);
         return ci;
     }