fix a serious pessimization that Tron on IRC pointed out where we would

"boolify" pointers, generating really awful code because getting the pointer
value requires a load itself.  Before:

_foo:
	movb	$1, _X.b
	ret
_get:
	xorl	%ecx, %ecx
	movb	_X.b, %al
	testb	%al, %al
	movl	$_Y, %eax
	cmove	%ecx, %eax
	ret

With the xform disabled:

_foo:
	movl	$_Y, _X
	ret
_get:
	movl	_X, %eax
	ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@66351 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/IPO/GlobalOpt.cpp b/lib/Transforms/IPO/GlobalOpt.cpp
index b93eb6b6d40ef24292c692350c15c2c949442ee1..1d0993a76164f2cbc10a3afe4de70261ed5e0b7f 100644 (file)
--- a/lib/Transforms/IPO/GlobalOpt.cpp
+++ b/lib/Transforms/IPO/GlobalOpt.cpp
@@ -1531,10 +1531,12 @@ static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
   const Type *GVElType = GV->getType()->getElementType();
   
   // If GVElType is already i1, it is already shrunk.  If the type of the GV is
-  // an FP value or vector, don't do this optimization because a select between
-  // them is very expensive and unlikely to lead to later simplification.
+  // an FP value, pointer or vector, don't do this optimization because a select
+  // between them is very expensive and unlikely to lead to later
+  // simplification.  In these cases, we typically end up with "cond ? v1 : v2"
+  // where v1 and v2 both require constant pool loads, a big loss.
   if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
-      isa<VectorType>(GVElType))
+      isa<PointerType>(GVElType) || isa<VectorType>(GVElType))
     return false;
   
   // Walk the use list of the global seeing if all the uses are load or store.

diff --git a/test/Transforms/GlobalOpt/2009-03-07-PromotePtrToBool.ll b/test/Transforms/GlobalOpt/2009-03-07-PromotePtrToBool.ll
new file mode 100644 (file)
index 0000000..8dba5b1
--- /dev/null
+++ b/test/Transforms/GlobalOpt/2009-03-07-PromotePtrToBool.ll
@@ -0,0 +1,17 @@
+; RUN: llvm-as < %s | opt -globalopt | llvm-dis | grep {@X = internal global i32}
+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-f80:128:128"
+target triple = "i386-apple-darwin7"
+@X = internal global i32* null         ; <i32**> [#uses=2]
+@Y = internal global i32 0             ; <i32*> [#uses=1]
+
+define void @foo() nounwind {
+entry:
+       store i32* @Y, i32** @X, align 4
+       ret void
+}
+
+define i32* @get() nounwind {
+entry:
+       %0 = load i32** @X, align 4             ; <i32*> [#uses=1]
+       ret i32* %0
+}