1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-23 19:23:23 +01:00

Sneaky, sneaky: move the -1 to the outside of the SMax. Reinstate the

optimization of SGE/SLE with unit stride, now that it works properly.

llvm-svn: 60881
This commit is contained in:
Nick Lewycky 2008-12-11 17:40:14 +00:00
parent 74411f221b
commit 312d95be37
3 changed files with 42 additions and 16 deletions

View File

@ -2924,8 +2924,12 @@ bool ScalarEvolutionsImpl::potentialInfiniteLoop(SCEV *Stride, SCEV *RHS,
if (!R)
return true;
if (isSigned)
if (isSigned) {
if (SC->getValue()->isOne())
return R->getValue()->isMaxValue(true);
return true; // XXX: because we don't have an sdiv scev.
}
// If negative, it wraps around every iteration, but we don't care about that.
APInt S = SC->getValue()->getValue().abs();
@ -2971,14 +2975,6 @@ HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L,
// run (m-n)/s times.
SCEVHandle End = RHS;
if (!executesAtLeastOnce(L, isSigned, trueWhenEqual,
SE.getMinusSCEV(Start, One), RHS)) {
// If not, we get the value of the LHS in the first iteration in which
// the above condition doesn't hold. This equals to max(m,n).
End = isSigned ? SE.getSMaxExpr(RHS, Start)
: SE.getUMaxExpr(RHS, Start);
}
// If the expression is less-than-or-equal to, we need to extend the
// loop by one iteration.
//
@ -2987,12 +2983,16 @@ HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L,
// division would equal one, but the loop runs twice putting the
// induction variable at 12.
if (!trueWhenEqual)
// (Stride - 1) is correct only because we know it's unsigned.
// What we really want is to decrease the magnitude of Stride by one.
Start = SE.getMinusSCEV(Start, SE.getMinusSCEV(Stride, One));
else
Start = SE.getMinusSCEV(Start, Stride);
if (trueWhenEqual)
End = SE.getAddExpr(End, One);
if (!executesAtLeastOnce(L, isSigned, trueWhenEqual,
SE.getMinusSCEV(Start, One), RHS)) {
// If not, we get the value of the LHS in the first iteration in which
// the above condition doesn't hold. This equals to max(m,n).
End = isSigned ? SE.getSMaxExpr(End, Start)
: SE.getUMaxExpr(End, Start);
}
// Finally, we subtract these two values to get the number of times the
// backedge is executed: max(m,n)-n.

View File

@ -1,5 +1,4 @@
; RUN: llvm-as < %s | opt -analyze -scalar-evolution | grep {255 iterations}
; XFAIL: *
define i32 @foo(i32 %x, i32 %y, i32* %lam, i32* %alp) nounwind {
bb1.thread:

View File

@ -0,0 +1,27 @@
; RUN: llvm-as < %s | opt -analyze -scalar-evolution | grep {0 smax}
define i32 @f(i32 %c.idx.val) {
bb2:
%k.018 = add i32 %c.idx.val, -1 ; <i32> [#uses=2]
%a14 = icmp slt i32 %k.018, 0 ; <i1> [#uses=1]
br i1 %a14, label %bb19, label %bb16.preheader
bb16.preheader:
%k.019 = phi i32 [ %k.0, %bb18 ], [ %k.018, %bb2 ] ; <i32> [#uses=5]
%x = phi i32 [ 0, %bb2 ], [ %x.1, %bb18]
br label %bb18
bb18: ; preds = %bb18.loopexit
%x.1 = add i32 %x, 1
%k.0 = add i32 %k.019, -1 ; <i32> [#uses=2]
%a107 = icmp slt i32 %k.0, 0 ; <i1> [#uses=1]
br i1 %a107, label %bb18.bb19_crit_edge, label %bb16.preheader
bb18.bb19_crit_edge:
ret i32 %x
bb19:
ret i32 0
}