From ad7321f58a485ae80fe3da1f400aa695e250b1a8 Mon Sep 17 00:00:00 2001 From: Dan Gohman Date: Mon, 15 Sep 2008 21:22:06 +0000 Subject: [PATCH] Teach LSR to optimize away SMAX operations for tripcounts in common cases. See the comment above OptimizeSMax for the full story, and the testcase for an example. This cancels out a pessimization commonly attributed to indvars, and will allow us to lift some of the artificial throttles in indvars, rather than add new ones. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56230 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Transforms/Scalar/LoopStrengthReduce.cpp | 127 ++++++++++ test/CodeGen/X86/optimize-smax.ll | 236 +++++++++++++++++++ 2 files changed, 363 insertions(+) create mode 100644 test/CodeGen/X86/optimize-smax.ll diff --git a/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/lib/Transforms/Scalar/LoopStrengthReduce.cpp index 145aa92a60e..07438e30c71 100644 --- a/lib/Transforms/Scalar/LoopStrengthReduce.cpp +++ b/lib/Transforms/Scalar/LoopStrengthReduce.cpp @@ -184,6 +184,11 @@ private: /// 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); + bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse, const SCEVHandle *&CondStride); bool RequiresTypeConversion(const Type *Ty, const Type *NewTy); @@ -1695,6 +1700,123 @@ ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond, return Cond; } +/// OptimizeSMax - Rewrite the loop's terminating condition if it uses +/// an smax computation. +/// +/// This is a narrow solution to a specific, but acute, problem. For loops +/// like this: +/// +/// i = 0; +/// do { +/// 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: +// +/// if (n > 0) { +/// i = 0; +/// do { +/// p[i] = 0.0; +/// } while (++i < n); +/// } +/// +/// and then it's possible for subsequent optimization to obscure the if +/// 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 +/// induction variable: +/// +/// i = 0; +/// smax = n < 1 ? 1 : n; +/// do { +/// p[i] = 0.0; +/// } while (++i != smax); +/// +/// Canonical induction variables are necessary because the loop passes +/// are designed around them. The most obvious example of this is the +/// LoopInfo analysis, which doesn't remember trip count values. It +/// expects to be able to rediscover the trip count each time it is +/// needed, and it does this using a simple analyis that only succeeds if +/// the loop has a canonical induction variable. +/// +/// However, when it comes time to generate code, the maximum operation +/// can be quite costly, especially if it's inside of an outer loop. +/// +/// This function solves this problem by detecting this type of loop and +/// 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) { + // Check that the loop matches the pattern we're looking for. + if (Cond->getPredicate() != CmpInst::ICMP_EQ && + Cond->getPredicate() != CmpInst::ICMP_NE) + return Cond; + + SelectInst *Sel = dyn_cast(Cond->getOperand(1)); + if (!Sel || !Sel->hasOneUse()) return Cond; + + SCEVHandle IterationCount = SE->getIterationCount(L); + if (isa(IterationCount)) + return Cond; + SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType()); + + // Adjust for an annoying getIterationCount quirk. + IterationCount = SE->getAddExpr(IterationCount, One); + + // Check for a max calculation that matches the pattern. + SCEVSMaxExpr *SMax = dyn_cast(IterationCount); + if (!SMax || SMax != SE->getSCEV(Sel)) return Cond; + + SCEVHandle SMaxLHS = SMax->getOperand(0); + SCEVHandle SMaxRHS = SMax->getOperand(1); + if (!SMaxLHS || SMaxLHS != One) return Cond; + + // Check the relevant induction variable for conformance to + // the pattern. + SCEVHandle IV = SE->getSCEV(Cond->getOperand(0)); + SCEVAddRecExpr *AR = dyn_cast(IV); + if (!AR || !AR->isAffine() || + AR->getStart() != One || + AR->getStepRecurrence(*SE) != One) + return Cond; + + // 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) + NewRHS = Sel->getOperand(1); + else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS) + NewRHS = Sel->getOperand(2); + if (!NewRHS) return Cond; + + // 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); + + // Delete the max calculation instructions. + Cond->replaceAllUsesWith(NewCond); + Cond->eraseFromParent(); + SE->deleteValueFromRecords(Cond); + Instruction *Cmp = cast(Sel->getOperand(0)); + Sel->eraseFromParent(); + SE->deleteValueFromRecords(Sel); + if (Cmp->use_empty()) { + Cmp->eraseFromParent(); + SE->deleteValueFromRecords(Cmp); + } + CondUse->User = NewCond; + return NewCond; +} + /// OptimizeShadowIV - If IV is used in a int-to-float cast /// inside the loop then try to eliminate the cast opeation. void LoopStrengthReduce::OptimizeShadowIV(Loop *L) { @@ -1836,6 +1958,11 @@ void LoopStrengthReduce::OptimizeIndvars(Loop *L) { if (!FindIVUserForCond(Cond, CondUse, CondStride)) return; // setcc doesn't use the IV. + // If the trip count is computed in terms of an smax (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); + // If possible, change stride and operands of the compare instruction to // eliminate one stride. Cond = ChangeCompareStride(L, Cond, CondUse, CondStride); diff --git a/test/CodeGen/X86/optimize-smax.ll b/test/CodeGen/X86/optimize-smax.ll new file mode 100644 index 00000000000..0c3be31e293 --- /dev/null +++ b/test/CodeGen/X86/optimize-smax.ll @@ -0,0 +1,236 @@ +; RUN: llvm-as < %s | llc -march=x86 | not grep cmov + +; LSR should be able to eliminate the smax computations by +; making the loops use slt comparisons instead of ne comparisons. + +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-darwin9" + +define void @foo(i8* %r, i32 %s, i32 %w, i32 %x, i8* %j, i32 %d) nounwind { +entry: + %0 = mul i32 %x, %w ; [#uses=2] + %1 = mul i32 %x, %w ; [#uses=1] + %2 = sdiv i32 %1, 4 ; [#uses=1] + %.sum2 = add i32 %2, %0 ; [#uses=2] + %cond = icmp eq i32 %d, 1 ; [#uses=1] + br i1 %cond, label %bb29, label %bb10.preheader + +bb10.preheader: ; preds = %entry + %3 = icmp sgt i32 %x, 0 ; [#uses=1] + br i1 %3, label %bb.nph9, label %bb18.loopexit + +bb.nph7: ; preds = %bb7.preheader + %4 = mul i32 %y.08, %w ; [#uses=1] + %5 = mul i32 %y.08, %s ; [#uses=1] + %6 = add i32 %5, 1 ; [#uses=1] + %tmp8 = icmp sgt i32 1, %w ; [#uses=1] + %smax9 = select i1 %tmp8, i32 1, i32 %w ; [#uses=1] + br label %bb6 + +bb6: ; preds = %bb7, %bb.nph7 + %x.06 = phi i32 [ 0, %bb.nph7 ], [ %indvar.next7, %bb7 ] ; [#uses=3] + %7 = add i32 %x.06, %4 ; [#uses=1] + %8 = shl i32 %x.06, 1 ; [#uses=1] + %9 = add i32 %6, %8 ; [#uses=1] + %10 = getelementptr i8* %r, i32 %9 ; [#uses=1] + %11 = load i8* %10, align 1 ; [#uses=1] + %12 = getelementptr i8* %j, i32 %7 ; [#uses=1] + store i8 %11, i8* %12, align 1 + br label %bb7 + +bb7: ; preds = %bb6 + %indvar.next7 = add i32 %x.06, 1 ; [#uses=2] + %exitcond10 = icmp ne i32 %indvar.next7, %smax9 ; [#uses=1] + br i1 %exitcond10, label %bb6, label %bb7.bb9_crit_edge + +bb7.bb9_crit_edge: ; preds = %bb7 + br label %bb9 + +bb9: ; preds = %bb7.preheader, %bb7.bb9_crit_edge + br label %bb10 + +bb10: ; preds = %bb9 + %indvar.next11 = add i32 %y.08, 1 ; [#uses=2] + %exitcond12 = icmp ne i32 %indvar.next11, %x ; [#uses=1] + br i1 %exitcond12, label %bb7.preheader, label %bb10.bb18.loopexit_crit_edge + +bb10.bb18.loopexit_crit_edge: ; preds = %bb10 + br label %bb10.bb18.loopexit_crit_edge.split + +bb10.bb18.loopexit_crit_edge.split: ; preds = %bb.nph9, %bb10.bb18.loopexit_crit_edge + br label %bb18.loopexit + +bb.nph9: ; preds = %bb10.preheader + %13 = icmp sgt i32 %w, 0 ; [#uses=1] + br i1 %13, label %bb.nph9.split, label %bb10.bb18.loopexit_crit_edge.split + +bb.nph9.split: ; preds = %bb.nph9 + br label %bb7.preheader + +bb7.preheader: ; preds = %bb.nph9.split, %bb10 + %y.08 = phi i32 [ 0, %bb.nph9.split ], [ %indvar.next11, %bb10 ] ; [#uses=3] + br i1 true, label %bb.nph7, label %bb9 + +bb.nph5: ; preds = %bb18.loopexit + %14 = sdiv i32 %w, 2 ; [#uses=1] + %15 = icmp slt i32 %w, 2 ; [#uses=1] + %16 = sdiv i32 %x, 2 ; [#uses=2] + br i1 %15, label %bb18.bb20_crit_edge.split, label %bb.nph5.split + +bb.nph5.split: ; preds = %bb.nph5 + %tmp2 = icmp sgt i32 1, %16 ; [#uses=1] + %smax3 = select i1 %tmp2, i32 1, i32 %16 ; [#uses=1] + br label %bb13 + +bb13: ; preds = %bb18, %bb.nph5.split + %y.14 = phi i32 [ 0, %bb.nph5.split ], [ %indvar.next1, %bb18 ] ; [#uses=4] + %17 = mul i32 %14, %y.14 ; [#uses=2] + %18 = shl i32 %y.14, 1 ; [#uses=1] + %19 = srem i32 %y.14, 2 ; [#uses=1] + %20 = add i32 %19, %18 ; [#uses=1] + %21 = mul i32 %20, %s ; [#uses=2] + br i1 true, label %bb.nph3, label %bb17 + +bb.nph3: ; preds = %bb13 + %22 = add i32 %17, %0 ; [#uses=1] + %23 = add i32 %17, %.sum2 ; [#uses=1] + %24 = sdiv i32 %w, 2 ; [#uses=2] + %tmp = icmp sgt i32 1, %24 ; [#uses=1] + %smax = select i1 %tmp, i32 1, i32 %24 ; [#uses=1] + br label %bb14 + +bb14: ; preds = %bb15, %bb.nph3 + %x.12 = phi i32 [ 0, %bb.nph3 ], [ %indvar.next, %bb15 ] ; [#uses=5] + %25 = shl i32 %x.12, 2 ; [#uses=1] + %26 = add i32 %25, %21 ; [#uses=1] + %27 = getelementptr i8* %r, i32 %26 ; [#uses=1] + %28 = load i8* %27, align 1 ; [#uses=1] + %.sum = add i32 %22, %x.12 ; [#uses=1] + %29 = getelementptr i8* %j, i32 %.sum ; [#uses=1] + store i8 %28, i8* %29, align 1 + %30 = shl i32 %x.12, 2 ; [#uses=1] + %31 = or i32 %30, 2 ; [#uses=1] + %32 = add i32 %31, %21 ; [#uses=1] + %33 = getelementptr i8* %r, i32 %32 ; [#uses=1] + %34 = load i8* %33, align 1 ; [#uses=1] + %.sum6 = add i32 %23, %x.12 ; [#uses=1] + %35 = getelementptr i8* %j, i32 %.sum6 ; [#uses=1] + store i8 %34, i8* %35, align 1 + br label %bb15 + +bb15: ; preds = %bb14 + %indvar.next = add i32 %x.12, 1 ; [#uses=2] + %exitcond = icmp ne i32 %indvar.next, %smax ; [#uses=1] + br i1 %exitcond, label %bb14, label %bb15.bb17_crit_edge + +bb15.bb17_crit_edge: ; preds = %bb15 + br label %bb17 + +bb17: ; preds = %bb15.bb17_crit_edge, %bb13 + br label %bb18 + +bb18.loopexit: ; preds = %bb10.bb18.loopexit_crit_edge.split, %bb10.preheader + %36 = icmp slt i32 %x, 2 ; [#uses=1] + br i1 %36, label %bb20, label %bb.nph5 + +bb18: ; preds = %bb17 + %indvar.next1 = add i32 %y.14, 1 ; [#uses=2] + %exitcond4 = icmp ne i32 %indvar.next1, %smax3 ; [#uses=1] + br i1 %exitcond4, label %bb13, label %bb18.bb20_crit_edge + +bb18.bb20_crit_edge: ; preds = %bb18 + br label %bb18.bb20_crit_edge.split + +bb18.bb20_crit_edge.split: ; preds = %bb18.bb20_crit_edge, %bb.nph5 + br label %bb20 + +bb20: ; preds = %bb18.bb20_crit_edge.split, %bb18.loopexit + switch i32 %d, label %return [ + i32 3, label %bb22 + i32 1, label %bb29 + ] + +bb22: ; preds = %bb20 + %37 = mul i32 %x, %w ; [#uses=1] + %38 = sdiv i32 %37, 4 ; [#uses=1] + %.sum3 = add i32 %38, %.sum2 ; [#uses=2] + %39 = add i32 %x, 15 ; [#uses=1] + %40 = and i32 %39, -16 ; [#uses=1] + %41 = add i32 %w, 15 ; [#uses=1] + %42 = and i32 %41, -16 ; [#uses=1] + %43 = mul i32 %40, %s ; [#uses=1] + %44 = icmp sgt i32 %x, 0 ; [#uses=1] + br i1 %44, label %bb.nph, label %bb26 + +bb.nph: ; preds = %bb22 + br label %bb23 + +bb23: ; preds = %bb24, %bb.nph + %y.21 = phi i32 [ 0, %bb.nph ], [ %indvar.next5, %bb24 ] ; [#uses=3] + %45 = mul i32 %y.21, %42 ; [#uses=1] + %.sum1 = add i32 %45, %43 ; [#uses=1] + %46 = getelementptr i8* %r, i32 %.sum1 ; [#uses=1] + %47 = mul i32 %y.21, %w ; [#uses=1] + %.sum5 = add i32 %47, %.sum3 ; [#uses=1] + %48 = getelementptr i8* %j, i32 %.sum5 ; [#uses=1] + tail call void @llvm.memcpy.i32(i8* %48, i8* %46, i32 %w, i32 1) + br label %bb24 + +bb24: ; preds = %bb23 + %indvar.next5 = add i32 %y.21, 1 ; [#uses=2] + %exitcond6 = icmp ne i32 %indvar.next5, %x ; [#uses=1] + br i1 %exitcond6, label %bb23, label %bb24.bb26_crit_edge + +bb24.bb26_crit_edge: ; preds = %bb24 + br label %bb26 + +bb26: ; preds = %bb24.bb26_crit_edge, %bb22 + %49 = mul i32 %x, %w ; [#uses=1] + %.sum4 = add i32 %.sum3, %49 ; [#uses=1] + %50 = getelementptr i8* %j, i32 %.sum4 ; [#uses=1] + %51 = mul i32 %x, %w ; [#uses=1] + %52 = sdiv i32 %51, 2 ; [#uses=1] + tail call void @llvm.memset.i32(i8* %50, i8 -128, i32 %52, i32 1) + ret void + +bb29: ; preds = %bb20, %entry + %53 = add i32 %w, 15 ; [#uses=1] + %54 = and i32 %53, -16 ; [#uses=1] + %55 = icmp sgt i32 %x, 0 ; [#uses=1] + br i1 %55, label %bb.nph11, label %bb33 + +bb.nph11: ; preds = %bb29 + br label %bb30 + +bb30: ; preds = %bb31, %bb.nph11 + %y.310 = phi i32 [ 0, %bb.nph11 ], [ %indvar.next13, %bb31 ] ; [#uses=3] + %56 = mul i32 %y.310, %54 ; [#uses=1] + %57 = getelementptr i8* %r, i32 %56 ; [#uses=1] + %58 = mul i32 %y.310, %w ; [#uses=1] + %59 = getelementptr i8* %j, i32 %58 ; [#uses=1] + tail call void @llvm.memcpy.i32(i8* %59, i8* %57, i32 %w, i32 1) + br label %bb31 + +bb31: ; preds = %bb30 + %indvar.next13 = add i32 %y.310, 1 ; [#uses=2] + %exitcond14 = icmp ne i32 %indvar.next13, %x ; [#uses=1] + br i1 %exitcond14, label %bb30, label %bb31.bb33_crit_edge + +bb31.bb33_crit_edge: ; preds = %bb31 + br label %bb33 + +bb33: ; preds = %bb31.bb33_crit_edge, %bb29 + %60 = mul i32 %x, %w ; [#uses=1] + %61 = getelementptr i8* %j, i32 %60 ; [#uses=1] + %62 = mul i32 %x, %w ; [#uses=1] + %63 = sdiv i32 %62, 2 ; [#uses=1] + tail call void @llvm.memset.i32(i8* %61, i8 -128, i32 %63, i32 1) + ret void + +return: ; preds = %bb20 + ret void +} + +declare void @llvm.memcpy.i32(i8*, i8*, i32, i32) nounwind + +declare void @llvm.memset.i32(i8*, i8, i32, i32) nounwind