From 1b9706b8c64114217e49981e4f3d26ccb148fabe Mon Sep 17 00:00:00 2001 From: Juergen Ributzka Date: Mon, 10 Nov 2014 21:05:31 +0000 Subject: [PATCH] [AArch64][FastISel] Fix kill flags for integer extends. In the case we optimize an integer extend away and replace it directly with the source register, we also have to clear all kill flags at all its uses. This is necessary, because the orignal IR instruction might be trivially dead, but we replaced it with a nop at MI level. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@221628 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/AArch64/AArch64FastISel.cpp | 8 ++++++++ test/CodeGen/AArch64/fast-isel-int-ext4.ll | 20 ++++++++++++++++++++ 2 files changed, 28 insertions(+) create mode 100644 test/CodeGen/AArch64/fast-isel-int-ext4.ll diff --git a/lib/Target/AArch64/AArch64FastISel.cpp b/lib/Target/AArch64/AArch64FastISel.cpp index 01b2dee2612..8ec0ce4c926 100644 --- a/lib/Target/AArch64/AArch64FastISel.cpp +++ b/lib/Target/AArch64/AArch64FastISel.cpp @@ -4197,6 +4197,14 @@ bool AArch64FastISel::selectIntExt(const Instruction *I) { .addImm(AArch64::sub_32); SrcReg = ResultReg; } + // Conservatively clear all kill flags from all uses, because we are + // replacing a sign-/zero-extend instruction at IR level with a nop at MI + // level. The result of the instruction at IR level might have been + // trivially dead, which is now not longer true. + unsigned UseReg = lookUpRegForValue(I); + if (UseReg) + MRI.clearKillFlags(UseReg); + updateValueMap(I, SrcReg); return true; } diff --git a/test/CodeGen/AArch64/fast-isel-int-ext4.ll b/test/CodeGen/AArch64/fast-isel-int-ext4.ll new file mode 100644 index 00000000000..f25bb98af75 --- /dev/null +++ b/test/CodeGen/AArch64/fast-isel-int-ext4.ll @@ -0,0 +1,20 @@ +; RUN: llc -mtriple=aarch64-apple-darwin -fast-isel -fast-isel-abort -verify-machineinstrs < %s | FileCheck %s + +define i32 @kill_flag(i16 signext %a) { +; CHECK-LABEL: kill_flag +entry: + %0 = sext i16 %a to i32 + br label %bb1 + +bb1: + %1 = icmp slt i32 undef, %0 + br i1 %1, label %loop, label %exit + +loop: + %2 = sext i16 %a to i32 + %3 = icmp slt i32 undef, %2 + br i1 %3, label %bb1, label %exit + +exit: + ret i32 0 +} -- 2.34.1