From e31a61eeca5bb2893a8a683de8e4cf2189a13050 Mon Sep 17 00:00:00 2001
From: Dan Gohman <gohman@apple.com>
Date: Fri, 20 Jul 2007 16:34:21 +0000
Subject: [PATCH] Optimize alignment of loads and stores.

llvm-svn: 40102
---
 .../Scalar/InstructionCombining.cpp           | 10 +++
 .../InstCombine/loadstore-alignment.ll        | 66 +++++++++++++++++++
 2 files changed, 76 insertions(+)
 create mode 100644 llvm/test/Transforms/InstCombine/loadstore-alignment.ll

diff --git a/llvm/lib/Transforms/Scalar/InstructionCombining.cpp b/llvm/lib/Transforms/Scalar/InstructionCombining.cpp
index 816a1c626af6..33fd843d63a2 100644
--- a/llvm/lib/Transforms/Scalar/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/Scalar/InstructionCombining.cpp
@@ -8790,6 +8790,11 @@ static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
   Value *Op = LI.getOperand(0);
 
+  // Attempt to improve the alignment.
+  unsigned KnownAlign = GetKnownAlignment(Op, TD);
+  if (KnownAlign > LI.getAlignment())
+    LI.setAlignment(KnownAlign);
+
   // load (cast X) --> cast (load X) iff safe
   if (isa<CastInst>(Op))
     if (Instruction *Res = InstCombineLoadCast(*this, LI))
@@ -8985,6 +8990,11 @@ Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
       }
   }
 
+  // Attempt to improve the alignment.
+  unsigned KnownAlign = GetKnownAlignment(Ptr, TD);
+  if (KnownAlign > SI.getAlignment())
+    SI.setAlignment(KnownAlign);
+
   // Do really simple DSE, to catch cases where there are several consequtive
   // stores to the same location, separated by a few arithmetic operations. This
   // situation often occurs with bitfield accesses.
diff --git a/llvm/test/Transforms/InstCombine/loadstore-alignment.ll b/llvm/test/Transforms/InstCombine/loadstore-alignment.ll
new file mode 100644
index 000000000000..fd8b0e2e17ae
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/loadstore-alignment.ll
@@ -0,0 +1,66 @@
+; RUN: llvm-as < %s | opt -instcombine | llvm-dis | grep {, align 16} | wc -l | grep 12
+
+@x = external global <2 x i64>, align 16
+@xx = external global [13 x <2 x i64>], align 16
+
+define <2 x i64> @static_hem() {
+	%t = getelementptr <2 x i64>* @x, i32 7
+	%tmp1 = load <2 x i64>* %t, align 1
+	ret <2 x i64> %tmp1
+}
+
+define <2 x i64> @hem(i32 %i) {
+	%t = getelementptr <2 x i64>* @x, i32 %i
+	%tmp1 = load <2 x i64>* %t, align 1
+	ret <2 x i64> %tmp1
+}
+
+define <2 x i64> @hem_2d(i32 %i, i32 %j) {
+	%t = getelementptr [13 x <2 x i64>]* @xx, i32 %i, i32 %j
+	%tmp1 = load <2 x i64>* %t, align 1
+	ret <2 x i64> %tmp1
+}
+
+define <2 x i64> @foo() {
+	%tmp1 = load <2 x i64>* @x, align 1
+	ret <2 x i64> %tmp1
+}
+
+define <2 x i64> @bar() {
+	%t = alloca <2 x i64>
+        call void @kip(<2 x i64>* %t);
+	%tmp1 = load <2 x i64>* %t, align 1
+	ret <2 x i64> %tmp1
+}
+
+define void @static_hem_store(<2 x i64> %y) {
+	%t = getelementptr <2 x i64>* @x, i32 7
+	store <2 x i64> %y, <2 x i64>* %t, align 1
+        ret void
+}
+
+define void @hem_store(i32 %i, <2 x i64> %y) {
+	%t = getelementptr <2 x i64>* @x, i32 %i
+	store <2 x i64> %y, <2 x i64>* %t, align 1
+        ret void
+}
+
+define void @hem_2d_store(i32 %i, i32 %j, <2 x i64> %y) {
+	%t = getelementptr [13 x <2 x i64>]* @xx, i32 %i, i32 %j
+	store <2 x i64> %y, <2 x i64>* %t, align 1
+        ret void
+}
+
+define void @foo_store(<2 x i64> %y) {
+	store <2 x i64> %y, <2 x i64>* @x, align 1
+        ret void
+}
+
+define void @bar_store(<2 x i64> %y) {
+	%t = alloca <2 x i64>
+        call void @kip(<2 x i64>* %t);
+	store <2 x i64> %y, <2 x i64>* %t, align 1
+        ret void
+}
+
+declare void @kip(<2 x i64>* %t)