diff options
author | Nick Lewycky <nicholas@mxc.ca> | 2011-01-23 06:20:19 +0000 |
---|---|---|
committer | Nick Lewycky <nicholas@mxc.ca> | 2011-01-23 06:20:19 +0000 |
commit | 630d85a78cdebf14a6fc2a0750bf48738f26ef08 (patch) | |
tree | 76042102bce3da3bba3365ec6be77ff9ffd0927e | |
parent | 54cdad97eb77caf841ade5827a1d5da6b2d89df3 (diff) | |
download | external_llvm-630d85a78cdebf14a6fc2a0750bf48738f26ef08.zip external_llvm-630d85a78cdebf14a6fc2a0750bf48738f26ef08.tar.gz external_llvm-630d85a78cdebf14a6fc2a0750bf48738f26ef08.tar.bz2 |
Use value ranges to fold ext(trunc) in SCEV when possible.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@124062 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r-- | lib/Analysis/ScalarEvolution.cpp | 34 | ||||
-rw-r--r-- | test/Analysis/ScalarEvolution/fold.ll | 34 |
2 files changed, 68 insertions, 0 deletions
diff --git a/lib/Analysis/ScalarEvolution.cpp b/lib/Analysis/ScalarEvolution.cpp index 63602ff..0f2afb2 100644 --- a/lib/Analysis/ScalarEvolution.cpp +++ b/lib/Analysis/ScalarEvolution.cpp @@ -898,6 +898,23 @@ const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op, void *IP = 0; if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; + // zext(trunc(x)) --> zext(x) or x or trunc(x) + if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) { + // It's possible the bits taken off by the truncate were all zero bits. If + // so, we should be able to simplify this further. + const SCEV *X = ST->getOperand(); + ConstantRange CR = getUnsignedRange(X); + unsigned OrigBits = CR.getBitWidth(); + unsigned TruncBits = getTypeSizeInBits(ST->getType()); + unsigned NewBits = getTypeSizeInBits(Ty); + if (CR.truncate(TruncBits).zeroExtend(NewBits).contains( + CR.zextOrTrunc(NewBits))) { + if (NewBits > OrigBits) return getZeroExtendExpr(X, Ty); + if (NewBits < OrigBits) return getTruncateExpr(X, Ty); + return X; + } + } + // If the input value is a chrec scev, and we can prove that the value // did not overflow the old, smaller, value, we can zero extend all of the // operands (often constants). This allows analysis of something like @@ -1039,6 +1056,23 @@ const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op, if (isKnownNonNegative(Op)) return getZeroExtendExpr(Op, Ty); + // sext(trunc(x)) --> sext(x) or x or trunc(x) + if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) { + // It's possible the bits taken off by the truncate were all sign bits. If + // so, we should be able to simplify this further. + const SCEV *X = ST->getOperand(); + ConstantRange CR = getSignedRange(X); + unsigned OrigBits = CR.getBitWidth(); + unsigned TruncBits = getTypeSizeInBits(ST->getType()); + unsigned NewBits = getTypeSizeInBits(Ty); + if (CR.truncate(TruncBits).signExtend(NewBits).contains( + CR.sextOrTrunc(NewBits))) { + if (NewBits > OrigBits) return getSignExtendExpr(X, Ty); + if (NewBits < OrigBits) return getTruncateExpr(X, Ty); + return X; + } + } + // If the input value is a chrec scev, and we can prove that the value // did not overflow the old, smaller, value, we can sign extend all of the // operands (often constants). This allows analysis of something like diff --git a/test/Analysis/ScalarEvolution/fold.ll b/test/Analysis/ScalarEvolution/fold.ll index 3c1b6ae..90012a3 100644 --- a/test/Analysis/ScalarEvolution/fold.ll +++ b/test/Analysis/ScalarEvolution/fold.ll @@ -22,3 +22,37 @@ define i8 @test3(i8 %x) { ; CHECK: (3 * %x) ret i8 %C } + +define void @test4(i32 %x) { +entry: + %0 = icmp sge i32 %x, 0 + br i1 %0, label %loop, label %exit +loop: + %A = phi i32 [0, %entry], [%I, %loop] + %rand = icmp sgt i32 %A, 10 + %Z = select i1 %rand, i32 %A, i32 10 + %B = trunc i32 %Z to i16 + %C = sext i16 %B to i30 +; CHECK: %C = +; CHECK-NEXT: (trunc i32 (10 smax {0,+,1}<%loop>) to i30) + %D = sext i16 %B to i32 +; CHECK: %D = +; CHECK-NEXT: (10 smax {0,+,1}<%loop>) + %E = sext i16 %B to i34 +; CHECK: %E = +; CHECK-NEXT: (zext i32 (10 smax {0,+,1}<%loop>) to i34) + %F = zext i16 %B to i30 +; CHECK: %F = +; CHECK-NEXT: (trunc i32 (10 smax {0,+,1}<%loop>) to i30 + %G = zext i16 %B to i32 +; CHECK: %G = +; CHECK-NEXT: (10 smax {0,+,1}<%loop>) + %H = zext i16 %B to i34 +; CHECK: %H = +; CHECK-NEXT: (zext i32 (10 smax {0,+,1}<%loop>) to i34) + %I = add i32 %A, 1 + %1 = icmp ne i32 %A, 20 + br i1 %1, label %loop, label %exit +exit: + ret void +}
\ No newline at end of file |