Don't promote volatile loads/stores. This is needed (for example) to handle setjmp...

This fixes PR1520.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41461 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Utils/PromoteMemoryToRegister.cpp b/lib/Transforms/Utils/PromoteMemoryToRegister.cpp
index a6fb8707d4155c5a6472f3cb216ebf8c92cf8ebf..3348971138864568fd29825eea4283462a1d6576 100644 (file)
--- a/lib/Transforms/Utils/PromoteMemoryToRegister.cpp
+++ b/lib/Transforms/Utils/PromoteMemoryToRegister.cpp
@@ -63,14 +63,17 @@ bool llvm::isAllocaPromotable(const AllocaInst *AI) {
   // FIXME: If the memory unit is of pointer or integer type, we can permit
   // assignments to subsections of the memory unit.
 
-  // Only allow direct loads and stores...
+  // Only allow direct and non-volatile loads and stores...
   for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
        UI != UE; ++UI)     // Loop over all of the uses of the alloca
-    if (isa<LoadInst>(*UI)) {
-      // noop
+    if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
+      if (LI->isVolatile())
+        return false;
     } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
       if (SI->getOperand(0) == AI)
         return false;   // Don't allow a store OF the AI, only INTO the AI.
+      if (SI->isVolatile())
+        return false;
     } else {
       return false;   // Not a load or store.
     }

diff --git a/test/Transforms/Mem2Reg/2007-08-27-VolatileLoadsStores.ll b/test/Transforms/Mem2Reg/2007-08-27-VolatileLoadsStores.ll
new file mode 100644 (file)
index 0000000..579ed41
--- /dev/null
+++ b/test/Transforms/Mem2Reg/2007-08-27-VolatileLoadsStores.ll
@@ -0,0 +1,47 @@
+; RUN: llvm-as < %s | opt -std-compile-opts | llvm-dis | grep volatile | count 3
+; PR1520
+; Don't promote volatile loads/stores. This is really needed to handle setjmp/lonjmp properly.
+
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
+target triple = "i686-pc-linux-gnu"
+       %struct.__jmp_buf_tag = type { [6 x i32], i32, %struct.__sigset_t }
+       %struct.__sigset_t = type { [32 x i32] }
+@j = external global [1 x %struct.__jmp_buf_tag]               ; <[1 x %struct.__jmp_buf_tag]*> [#uses=1]
+
+define i32 @f() {
+entry:
+       %retval = alloca i32, align 4           ; <i32*> [#uses=2]
+       %v = alloca i32, align 4                ; <i32*> [#uses=3]
+       %tmp = alloca i32, align 4              ; <i32*> [#uses=3]
+       %"alloca point" = bitcast i32 0 to i32          ; <i32> [#uses=0]
+       volatile store i32 0, i32* %v, align 4
+       %tmp1 = call i32 @_setjmp( %struct.__jmp_buf_tag* getelementptr ([1 x %struct.__jmp_buf_tag]* @j, i32 0, i32 0) )               ; <i32> [#uses=1]
+       %tmp2 = icmp ne i32 %tmp1, 0            ; <i1> [#uses=1]
+       %tmp23 = zext i1 %tmp2 to i8            ; <i8> [#uses=1]
+       %toBool = icmp ne i8 %tmp23, 0          ; <i1> [#uses=1]
+       br i1 %toBool, label %bb, label %bb5
+
+bb:            ; preds = %entry
+       %tmp4 = volatile load i32* %v, align 4          ; <i32> [#uses=1]
+       store i32 %tmp4, i32* %tmp, align 4
+       br label %bb6
+
+bb5:           ; preds = %entry
+       volatile store i32 1, i32* %v, align 4
+       call void @g( )
+       store i32 0, i32* %tmp, align 4
+       br label %bb6
+
+bb6:           ; preds = %bb5, %bb
+       %tmp7 = load i32* %tmp, align 4         ; <i32> [#uses=1]
+       store i32 %tmp7, i32* %retval, align 4
+       br label %return
+
+return:                ; preds = %bb6
+       %retval8 = load i32* %retval            ; <i32> [#uses=1]
+       ret i32 %retval8
+}
+
+declare i32 @_setjmp(%struct.__jmp_buf_tag*)
+
+declare void @g()