/// inside the loop then try to eliminate the cast opeation.
void OptimizeShadowIV(Loop *L);
- /// OptimizeSMax - Rewrite the loop's terminating condition
- /// if it uses an smax computation.
- ICmpInst *OptimizeSMax(Loop *L, ICmpInst *Cond,
- IVStrideUse* &CondUse);
+ /// OptimizeMax - Rewrite the loop's terminating condition
+ /// if it uses a max computation.
+ ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
+ IVStrideUse* &CondUse);
bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
const SCEVHandle *&CondStride);
return Cond;
}
-/// OptimizeSMax - Rewrite the loop's terminating condition if it uses
-/// an smax computation.
+/// OptimizeMax - Rewrite the loop's terminating condition if it uses
+/// a max computation.
///
/// This is a narrow solution to a specific, but acute, problem. For loops
/// like this:
/// p[i] = 0.0;
/// } while (++i < n);
///
-/// where the comparison is signed, the trip count isn't just 'n', because
-/// 'n' could be negative. And unfortunately this can come up even for loops
-/// where the user didn't use a C do-while loop. For example, seemingly
-/// well-behaved top-test loops will commonly be lowered like this:
+/// the trip count isn't just 'n', because 'n' might not be positive. And
+/// unfortunately this can come up even for loops where the user didn't use
+/// a C do-while loop. For example, seemingly well-behaved top-test loops
+/// will commonly be lowered like this:
//
/// if (n > 0) {
/// i = 0;
/// test in such a way that indvars can't find it.
///
/// When indvars can't find the if test in loops like this, it creates a
-/// signed-max expression, which allows it to give the loop a canonical
+/// max expression, which allows it to give the loop a canonical
/// induction variable:
///
/// i = 0;
-/// smax = n < 1 ? 1 : n;
+/// max = n < 1 ? 1 : n;
/// do {
/// p[i] = 0.0;
-/// } while (++i != smax);
+/// } while (++i != max);
///
/// Canonical induction variables are necessary because the loop passes
/// are designed around them. The most obvious example of this is the
/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
/// the instructions for the maximum computation.
///
-ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
- IVStrideUse* &CondUse) {
+ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
+ IVStrideUse* &CondUse) {
// Check that the loop matches the pattern we're looking for.
if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
Cond->getPredicate() != CmpInst::ICMP_NE)
SCEVHandle IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
// Check for a max calculation that matches the pattern.
- const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
- if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
+ if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
+ return Cond;
+ const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
+ if (Max != SE->getSCEV(Sel)) return Cond;
// Two handle a max with more than two operands, this optimization would
// require additional checking and setup.
- if (SMax->getNumOperands() != 2)
+ if (Max->getNumOperands() != 2)
return Cond;
- SCEVHandle SMaxLHS = SMax->getOperand(0);
- SCEVHandle SMaxRHS = SMax->getOperand(1);
- if (!SMaxLHS || SMaxLHS != One) return Cond;
+ SCEVHandle MaxLHS = Max->getOperand(0);
+ SCEVHandle MaxRHS = Max->getOperand(1);
+ if (!MaxLHS || MaxLHS != One) return Cond;
// Check the relevant induction variable for conformance to
// the pattern.
// Check the right operand of the select, and remember it, as it will
// be used in the new comparison instruction.
Value *NewRHS = 0;
- if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
+ if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
NewRHS = Sel->getOperand(1);
- else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
+ else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
NewRHS = Sel->getOperand(2);
if (!NewRHS) return Cond;
+ // Determine the new comparison opcode. It may be signed or unsigned,
+ // and the original comparison may be either equality or inequality.
+ CmpInst::Predicate Pred =
+ isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
+ if (Cond->getPredicate() == CmpInst::ICMP_EQ)
+ Pred = CmpInst::getInversePredicate(Pred);
+
// Ok, everything looks ok to change the condition into an SLT or SGE and
// delete the max calculation.
ICmpInst *NewCond =
- new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
- CmpInst::ICMP_SLT :
- CmpInst::ICMP_SGE,
- Cond->getOperand(0), NewRHS, "scmp", Cond);
+ new ICmpInst(Pred, Cond->getOperand(0), NewRHS, "scmp", Cond);
// Delete the max calculation instructions.
Cond->replaceAllUsesWith(NewCond);
StrideNoReuse.insert(*CondStride);
}
- // If the trip count is computed in terms of an smax (due to ScalarEvolution
+ // If the trip count is computed in terms of a max (due to ScalarEvolution
// being unable to find a sufficient guard, for example), change the loop
- // comparison to use SLT instead of NE.
- Cond = OptimizeSMax(L, Cond, CondUse);
+ // comparison to use SLT or ULT instead of NE.
+ Cond = OptimizeMax(L, Cond, CondUse);
// If possible, change stride and operands of the compare instruction to
// eliminate one stride.
--- /dev/null
+; RUN: llvm-as < %s | llc -march=x86-64 > %t
+; RUN: grep cmov %t | count 2
+; RUN: grep jne %t | count 1
+
+; LSR's OptimizeMax function shouldn't try to eliminate this max, because
+; it has three operands.
+
+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"
+
+define void @foo(double* nocapture %p, i64 %x, i64 %y) nounwind {
+entry:
+ %tmp = icmp eq i64 %y, 0 ; <i1> [#uses=1]
+ %umax = select i1 %tmp, i64 1, i64 %y ; <i64> [#uses=2]
+ %tmp8 = icmp ugt i64 %umax, %x ; <i1> [#uses=1]
+ %umax9 = select i1 %tmp8, i64 %umax, i64 %x ; <i64> [#uses=1]
+ br label %bb4
+
+bb4: ; preds = %bb4, %entry
+ %i.07 = phi i64 [ 0, %entry ], [ %2, %bb4 ] ; <i64> [#uses=2]
+ %scevgep = getelementptr double* %p, i64 %i.07 ; <double*> [#uses=2]
+ %0 = load double* %scevgep, align 8 ; <double> [#uses=1]
+ %1 = fmul double %0, 2.000000e+00 ; <double> [#uses=1]
+ store double %1, double* %scevgep, align 8
+ %2 = add i64 %i.07, 1 ; <i64> [#uses=2]
+ %exitcond = icmp eq i64 %2, %umax9 ; <i1> [#uses=1]
+ br i1 %exitcond, label %return, label %bb4
+
+return: ; preds = %bb4
+ ret void
+}
projects / oota-llvm.git / commitdiff