From 1723f6398ea8aa8d602a478f47695bf3b0374d35 Mon Sep 17 00:00:00 2001
From: John McCall <rjmccall@apple.com>
Date: Thu, 7 Mar 2013 21:36:54 +0000
Subject: [PATCH] Evaluate compound literals directly into the result aggregate
 when that aggregate isn't potentially aliased.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@176654 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/CodeGen/CGExprAgg.cpp       |  6 ++----
 test/CodeGen/compound-literal.c | 34 +++++++++++++++++++++++++++++++++
 2 files changed, 36 insertions(+), 4 deletions(-)

diff --git a/lib/CodeGen/CGExprAgg.cpp b/lib/CodeGen/CGExprAgg.cpp
index 8c64e8a6e7..f9f2733401 100644
--- a/lib/CodeGen/CGExprAgg.cpp
+++ b/lib/CodeGen/CGExprAgg.cpp
@@ -531,12 +531,10 @@ void AggExprEmitter::VisitOpaqueValueExpr(OpaqueValueExpr *e) {
 
 void
 AggExprEmitter::VisitCompoundLiteralExpr(CompoundLiteralExpr *E) {
-  if (E->getType().isPODType(CGF.getContext())) {
+  if (Dest.isPotentiallyAliased() &&
+      E->getType().isPODType(CGF.getContext())) {
     // For a POD type, just emit a load of the lvalue + a copy, because our
     // compound literal might alias the destination.
-    // FIXME: This is a band-aid; the real problem appears to be in our handling
-    // of assignments, where we store directly into the LHS without checking
-    // whether anything in the RHS aliases.
     EmitAggLoadOfLValue(E);
     return;
   }
diff --git a/test/CodeGen/compound-literal.c b/test/CodeGen/compound-literal.c
index a8eec615ad..e4bf962e12 100644
--- a/test/CodeGen/compound-literal.c
+++ b/test/CodeGen/compound-literal.c
@@ -32,3 +32,37 @@ void f() {
   s = (S){s.y,s.x};
   // CHECK-NEXT: ret void
 }
+
+// CHECK: define i48 @g(
+struct G { short x, y, z; };
+struct G g(int x, int y, int z) {
+  // CHECK:      [[RESULT:%.*]] = alloca [[G:%.*]], align 2
+  // CHECK-NEXT: [[X:%.*]] = alloca i32, align 4
+  // CHECK-NEXT: [[Y:%.*]] = alloca i32, align 4
+  // CHECK-NEXT: [[Z:%.*]] = alloca i32, align 4
+  // CHECK-NEXT: [[COERCE_TEMP:%.*]] = alloca i48
+  // CHECK-NEXT: store i32
+  // CHECK-NEXT: store i32
+  // CHECK-NEXT: store i32
+
+  // Evaluate the compound literal directly in the result value slot.
+  // CHECK-NEXT: [[T0:%.*]] = getelementptr inbounds [[G]]* [[RESULT]], i32 0, i32 0
+  // CHECK-NEXT: [[T1:%.*]] = load i32* [[X]], align 4
+  // CHECK-NEXT: [[T2:%.*]] = trunc i32 [[T1]] to i16
+  // CHECK-NEXT: store i16 [[T2]], i16* [[T0]], align 2
+  // CHECK-NEXT: [[T0:%.*]] = getelementptr inbounds [[G]]* [[RESULT]], i32 0, i32 1
+  // CHECK-NEXT: [[T1:%.*]] = load i32* [[Y]], align 4
+  // CHECK-NEXT: [[T2:%.*]] = trunc i32 [[T1]] to i16
+  // CHECK-NEXT: store i16 [[T2]], i16* [[T0]], align 2
+  // CHECK-NEXT: [[T0:%.*]] = getelementptr inbounds [[G]]* [[RESULT]], i32 0, i32 2
+  // CHECK-NEXT: [[T1:%.*]] = load i32* [[Z]], align 4
+  // CHECK-NEXT: [[T2:%.*]] = trunc i32 [[T1]] to i16
+  // CHECK-NEXT: store i16 [[T2]], i16* [[T0]], align 2
+  return (struct G) { x, y, z };
+
+  // CHECK-NEXT: [[T0:%.*]] = bitcast i48* [[COERCE_TEMP]] to i8*
+  // CHECK-NEXT: [[T1:%.*]] = bitcast [[G]]* [[RESULT]] to i8*
+  // CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[T0]], i8* [[T1]], i64 6
+  // CHECK-NEXT: [[T0:%.*]] = load i48* [[COERCE_TEMP]]
+  // CHECK-NEXT: ret i48 [[T0]]
+}
-- 
2.40.0