From 8b2dfae47acde261683f3b27dad58dcf74843315 Mon Sep 17 00:00:00 2001 From: Philip Reames Date: Tue, 6 Oct 2015 20:20:45 +0000 Subject: [PATCH] Extend known bits to understand @llvm.bswap This is a cleaned up patch from the one written by John Regehr based on the findings of the Souper superoptimizer. When writing tests, I was surprised to find that instsimplify apparently doesn't know how to collapse bit test sequences based purely on known bits. This required me to split my tests across both instsimplify and instcombine. Differential Revision: http://reviews.llvm.org/D13250 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@249453 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Analysis/ValueTracking.cpp | 6 +++ .../InstCombine/bswap-known-bits.ll | 47 +++++++++++++++++++ test/Transforms/InstSimplify/bswap.ll | 41 ++++++++++++++++ 3 files changed, 94 insertions(+) create mode 100644 test/Transforms/InstCombine/bswap-known-bits.ll create mode 100644 test/Transforms/InstSimplify/bswap.ll diff --git a/lib/Analysis/ValueTracking.cpp b/lib/Analysis/ValueTracking.cpp index f0b04b7133d..d3723037dda 100644 --- a/lib/Analysis/ValueTracking.cpp +++ b/lib/Analysis/ValueTracking.cpp @@ -1358,6 +1358,12 @@ static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero, if (IntrinsicInst *II = dyn_cast(I)) { switch (II->getIntrinsicID()) { default: break; + case Intrinsic::bswap: + computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, + Depth + 1, Q); + KnownZero |= KnownZero2.byteSwap(); + KnownOne |= KnownOne2.byteSwap(); + break; case Intrinsic::ctlz: case Intrinsic::cttz: { unsigned LowBits = Log2_32(BitWidth)+1; diff --git a/test/Transforms/InstCombine/bswap-known-bits.ll b/test/Transforms/InstCombine/bswap-known-bits.ll new file mode 100644 index 00000000000..1f3285af65c --- /dev/null +++ b/test/Transforms/InstCombine/bswap-known-bits.ll @@ -0,0 +1,47 @@ +; RUN: opt < %s -S -instcombine | FileCheck %s +; Note: This is testing functionality in computeKnownBits. I'd have rather +; used instsimplify, but the bit test folding is apparently only in instcombine. + +declare i16 @llvm.bswap.i16(i16) +declare i32 @llvm.bswap.i32(i32) + +define i1 @test1(i16 %arg) { +; CHECK-LABEL: @test1 +; CHECK: ret i1 true + %a = or i16 %arg, 511 + %b = call i16 @llvm.bswap.i16(i16 %a) + %and = and i16 %b, 256 + %res = icmp eq i16 %and, 256 + ret i1 %res +} + +define i1 @test2(i16 %arg) { +; CHECK-LABEL: @test2 +; CHECK: ret i1 true + %a = or i16 %arg, 1 + %b = call i16 @llvm.bswap.i16(i16 %a) + %and = and i16 %b, 256 + %res = icmp eq i16 %and, 256 + ret i1 %res +} + + +define i1 @test3(i16 %arg) { +; CHECK-LABEL: @test3 +; CHECK: ret i1 true + %a = or i16 %arg, 256 + %b = call i16 @llvm.bswap.i16(i16 %a) + %and = and i16 %b, 1 + %res = icmp eq i16 %and, 1 + ret i1 %res +} + +define i1 @test4(i32 %arg) { +; CHECK-LABEL: @test4 +; CHECK: ret i1 true + %a = or i32 %arg, 2147483647 ; i32_MAX + %b = call i32 @llvm.bswap.i32(i32 %a) + %and = and i32 %b, 127 + %res = icmp eq i32 %and, 127 + ret i1 %res +} diff --git a/test/Transforms/InstSimplify/bswap.ll b/test/Transforms/InstSimplify/bswap.ll new file mode 100644 index 00000000000..7bc3af9e307 --- /dev/null +++ b/test/Transforms/InstSimplify/bswap.ll @@ -0,0 +1,41 @@ +; RUN: opt < %s -S -instsimplify | FileCheck %s + +declare i16 @llvm.bswap.i16(i16) + +define i1 @test1(i16 %arg) { +; CHECK-LABEL: @test1 +; CHECK: ret i1 false + %a = or i16 %arg, 1 + %b = call i16 @llvm.bswap.i16(i16 %a) + %res = icmp eq i16 %b, 0 + ret i1 %res +} + +define i1 @test2(i16 %arg) { +; CHECK-LABEL: @test2 +; CHECK: ret i1 false + %a = or i16 %arg, 1024 + %b = call i16 @llvm.bswap.i16(i16 %a) + %res = icmp eq i16 %b, 0 + ret i1 %res +} + +define i1 @test3(i16 %arg) { +; CHECK-LABEL: @test3 +; CHECK: ret i1 false + %a = and i16 %arg, 1 + %b = call i16 @llvm.bswap.i16(i16 %a) + %and = and i16 %b, 1 + %res = icmp eq i16 %and, 1 + ret i1 %res +} + +define i1 @test4(i16 %arg) { +; CHECK-LABEL: @test4 +; CHECK: ret i1 false + %a = and i16 %arg, 511 + %b = call i16 @llvm.bswap.i16(i16 %a) + %and = and i16 %b, 256 + %res = icmp eq i16 %and, 1 + ret i1 %res +} -- 2.34.1