Fix comparisons of alloca alignment in inliner merging

Duncan pointed out a mistake in my fix in r186425 when only one of the allocas
being compared had the target-default alignment. This is essentially his
suggested solution. Thanks!

llvm-svn: 186510

diff --git a/llvm/lib/Transforms/IPO/Inliner.cpp b/llvm/lib/Transforms/IPO/Inliner.cpp
index f72121d..d75d6ca 100644 (file)
--- a/llvm/lib/Transforms/IPO/Inliner.cpp
+++ b/llvm/lib/Transforms/IPO/Inliner.cpp
@@ -216,9 +216,18 @@ static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
       
       AI->replaceAllUsesWith(AvailableAlloca);
 
-      if (Align1 > Align2 || (!Align1 && TD &&
-          TD->getABITypeAlignment(AI->getAllocatedType()) > Align2))
-        AvailableAlloca->setAlignment(Align1);
+      if (Align1 != Align2) {
+        if (!Align1 || !Align2) {
+          assert(TD && "DataLayout required to compare default alignments");
+          unsigned TypeAlign = TD->getABITypeAlignment(AI->getAllocatedType());
+
+          Align1 = Align1 ? Align1 : TypeAlign;
+          Align2 = Align2 ? Align2 : TypeAlign;
+        }
+
+        if (Align1 > Align2)
+          AvailableAlloca->setAlignment(AI->getAlignment());
+      }
 
       AI->eraseFromParent();
       MergedAwayAlloca = true;

diff --git a/llvm/test/Transforms/Inline/alloca-merge-align.ll b/llvm/test/Transforms/Inline/alloca-merge-align.ll
index de4707d..d789c79 100644 (file)
--- a/llvm/test/Transforms/Inline/alloca-merge-align.ll
+++ b/llvm/test/Transforms/Inline/alloca-merge-align.ll
@@ -35,6 +35,21 @@ entry:
   ret void
 }
 
+define void @foo1(%struct.s* byval nocapture readonly %a) {
+entry:
+  %x = alloca [2 x i32], align 1
+  %a1 = getelementptr inbounds %struct.s* %a, i64 0, i32 0
+  %0 = load i32* %a1, align 4, !tbaa !0
+  %arrayidx = getelementptr inbounds [2 x i32]* %x, i64 0, i64 0
+  store i32 %0, i32* %arrayidx, align 4, !tbaa !0
+  %b = getelementptr inbounds %struct.s* %a, i64 0, i32 1
+  %1 = load i32* %b, align 4, !tbaa !0
+  %arrayidx2 = getelementptr inbounds [2 x i32]* %x, i64 0, i64 1
+  store i32 %1, i32* %arrayidx2, align 4, !tbaa !0
+  call void @bar(i32* %arrayidx) #2
+  ret void
+}
+
 declare void @bar(i32*) #1
 
 define void @goo(%struct.s* byval nocapture readonly %a) {
@@ -88,6 +103,24 @@ entry:
   ret i32 0
 }
 
+; CHECK-LABEL: @test1
+; CHECK: {{alloca \[2 x i32\]$}}
+; CHECK-NOT: alloca [2 x i32]
+; CHECK: ret i32 0
+
+define signext i32 @test1() {
+entry:
+  %a = alloca i64, align 8
+  %tmpcast = bitcast i64* %a to %struct.s*
+  store i64 0, i64* %a, align 8
+  %a1 = bitcast i64* %a to i32*
+  store i32 1, i32* %a1, align 8, !tbaa !0
+  call void @foo0(%struct.s* byval %tmpcast)
+  store i32 2, i32* %a1, align 8, !tbaa !0
+  call void @foo1(%struct.s* byval %tmpcast)
+  ret i32 0
+}
+
 !0 = metadata !{metadata !"int", metadata !1}
 !1 = metadata !{metadata !"omnipotent char", metadata !2}
 !2 = metadata !{metadata !"Simple C/C++ TBAA"}