Fix the order that SCEVExpander considers add operands in so that

it doesn't miss an opportunity to form a GEP, regardless of the
relative loop depths of the operands. This fixes rdar://8197217.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@108475 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Dan Gohman 2010-07-15 23:38:13 +00:00
parent e9bf7e692e
commit bb5d92741b
2 changed files with 54 additions and 2 deletions

View File

@ -647,6 +647,11 @@ public:
bool operator()(std::pair<const Loop *, const SCEV *> LHS,
std::pair<const Loop *, const SCEV *> RHS) const {
// Keep pointer operands sorted at the end.
if (LHS.second->getType()->isPointerTy() !=
RHS.second->getType()->isPointerTy())
return LHS.second->getType()->isPointerTy();
// Compare loops with PickMostRelevantLoop.
if (LHS.first != RHS.first)
return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
@ -699,8 +704,15 @@ Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
// The running sum expression is a pointer. Try to form a getelementptr
// at this level with that as the base.
SmallVector<const SCEV *, 4> NewOps;
for (; I != E && I->first == CurLoop; ++I)
NewOps.push_back(I->second);
for (; I != E && I->first == CurLoop; ++I) {
// If the operand is SCEVUnknown and not instructions, peek through
// it, to enable more of it to be folded into the GEP.
const SCEV *X = I->second;
if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
if (!isa<Instruction>(U->getValue()))
X = SE.getSCEV(U->getValue());
NewOps.push_back(X);
}
Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
} else if (const PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
// The running sum is an integer, and there's a pointer at this level.

View File

@ -0,0 +1,40 @@
; RUN: opt -indvars -S | not grep uglygep
; rdar://8197217
; Indvars should be able to emit a clean GEP here, not an uglygep.
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
target triple = "x86_64-apple-darwin11.0"
@numf2s = external global i32 ; <i32*> [#uses=1]
@numf1s = external global i32 ; <i32*> [#uses=1]
@tds = external global double** ; <double***> [#uses=1]
define void @init_td(i32 %tmp7) nounwind {
entry:
br label %bb4
bb4: ; preds = %bb3, %entry
%i.0 = phi i32 [ 0, %entry ], [ %tmp9, %bb3 ] ; <i32> [#uses=3]
br label %bb
bb: ; preds = %bb4
br label %bb2
bb2: ; preds = %bb1, %bb
%j.0 = phi i32 [ 0, %bb ], [ %tmp6, %bb1 ] ; <i32> [#uses=3]
%tmp8 = icmp slt i32 %j.0, %tmp7 ; <i1> [#uses=1]
br i1 %tmp8, label %bb1, label %bb3
bb1: ; preds = %bb2
%tmp = load double*** @tds, align 8 ; <double**> [#uses=1]
%tmp1 = sext i32 %i.0 to i64 ; <i64> [#uses=1]
%tmp2 = getelementptr inbounds double** %tmp, i64 %tmp1 ; <double**> [#uses=1]
%tmp3 = load double** %tmp2, align 1 ; <double*> [#uses=1]
%tmp6 = add nsw i32 %j.0, 1 ; <i32> [#uses=1]
br label %bb2
bb3: ; preds = %bb2
%tmp9 = add nsw i32 %i.0, 1 ; <i32> [#uses=1]
br label %bb4
}