Fix a miscompile introduced in r220178.

The original code had an implicit assumption that if the test for
allocas or globals was reached, the two pointers were not equal. With my
changes to make the pointer analysis more powerful here, I also had to
guard against circumstances where the results weren't useful. That in
turn violated the assumption and gave rise to a circumstance in which we
could have a store with both the queried pointer and stored pointer
rooted at *the same* alloca. Clearly, we cannot ignore such a store.
There are other things we might do in this code to better handle the
case of both pointers ending up at the same alloca or global, but it
seems best to at least make the test explicit in what it intends to
check.

I've added tests for both the alloca and global case here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@220190 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Analysis/Loads.cpp b/lib/Analysis/Loads.cpp
index 65a30f2816fbfcbb711dd366cdd374f494d73668..bb0d60e263e82621e0cf9a039642c2fe49a9dda8 100644 (file)
--- a/lib/Analysis/Loads.cpp
+++ b/lib/Analysis/Loads.cpp
@@ -220,11 +220,12 @@ Value *llvm::FindAvailableLoadedValue(Value *Ptr, BasicBlock *ScanBB,
         return SI->getOperand(0);
       }
 
-      // If Ptr is an alloca and this is a store to a different alloca, ignore
-      // the store.  This is a trivial form of alias analysis that is important
-      // for reg2mem'd code.
+      // If both StrippedPtr and StorePtr reach all the way to an alloca or
+      // global and they are different, ignore the store. This is a trivial form
+      // of alias analysis that is important for reg2mem'd code.
       if ((isa<AllocaInst>(StrippedPtr) || isa<GlobalVariable>(StrippedPtr)) &&
-          (isa<AllocaInst>(StorePtr) || isa<GlobalVariable>(StorePtr)))
+          (isa<AllocaInst>(StorePtr) || isa<GlobalVariable>(StorePtr)) &&
+          StrippedPtr != StorePtr)
         continue;
 
       // If we have alias analysis and it says the store won't modify the loaded

diff --git a/test/Transforms/InstCombine/load.ll b/test/Transforms/InstCombine/load.ll
index 20d40e2ccfd6b03ed04b54bd7a1861310ea02e7e..b4b7558a1a05af8962ea8a14ef105cc5849823d2 100644 (file)
--- a/test/Transforms/InstCombine/load.ll
+++ b/test/Transforms/InstCombine/load.ll
@@ -119,3 +119,34 @@ define <16 x i8> @test13(<2 x i64> %x) {
   %tmp = load <16 x i8>* bitcast ([4 x i32]* @GLOBAL to <16 x i8>*)
   ret <16 x i8> %tmp
 }
+
+define i8 @test14(i8 %x, i32 %y) {
+; This test must not have the store of %x forwarded to the load -- there is an
+; intervening store if %y. However, the intervening store occurs with a different
+; type and size and to a different pointer value. This is ensuring that none of
+; those confuse the analysis into thinking that the second store does not alias
+; the first.
+; CHECK-LABEL: @test14(
+; CHECK:         %[[R:.*]] = load i8*
+; CHECK-NEXT:    ret i8 %[[R]]
+  %a = alloca i32
+  %a.i8 = bitcast i32* %a to i8*
+  store i8 %x, i8* %a.i8
+  store i32 %y, i32* %a
+  %r = load i8* %a.i8
+  ret i8 %r
+}
+
+@test15_global = external global i32
+
+define i8 @test15(i8 %x, i32 %y) {
+; Same test as @test14 essentially, but using a global instead of an alloca.
+; CHECK-LABEL: @test15(
+; CHECK:         %[[R:.*]] = load i8*
+; CHECK-NEXT:    ret i8 %[[R]]
+  %g.i8 = bitcast i32* @test15_global to i8*
+  store i8 %x, i8* %g.i8
+  store i32 %y, i32* @test15_global
+  %r = load i8* %g.i8
+  ret i8 %r
+}