1; RUN: opt -S -codegenprepare < %s | FileCheck %s --check-prefix=SLOW 2; RUN: opt -S -codegenprepare -mattr=+bmi < %s | FileCheck %s --check-prefix=FAST_TZ 3; RUN: opt -S -codegenprepare -mattr=+lzcnt < %s | FileCheck %s --check-prefix=FAST_LZ 4 5target triple = "x86_64-unknown-unknown" 6target datalayout = "e-n32:64" 7 8; If the intrinsic is cheap, nothing should change. 9; If the intrinsic is expensive, check if the input is zero to avoid the call. 10; This is undoing speculation that may have been created by SimplifyCFG + InstCombine. 11 12define i64 @cttz(i64 %A) { 13entry: 14 %z = call i64 @llvm.cttz.i64(i64 %A, i1 false) 15 ret i64 %z 16 17; SLOW-LABEL: @cttz( 18; SLOW: entry: 19; SLOW: %cmpz = icmp eq i64 %A, 0 20; SLOW: br i1 %cmpz, label %cond.end, label %cond.false 21; SLOW: cond.false: 22; SLOW: %z = call i64 @llvm.cttz.i64(i64 %A, i1 true) 23; SLOW: br label %cond.end 24; SLOW: cond.end: 25; SLOW: %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ] 26; SLOW: ret i64 %ctz 27 28; FAST_TZ-LABEL: @cttz( 29; FAST_TZ: %z = call i64 @llvm.cttz.i64(i64 %A, i1 false) 30; FAST_TZ: ret i64 %z 31} 32 33define i64 @ctlz(i64 %A) { 34entry: 35 %z = call i64 @llvm.ctlz.i64(i64 %A, i1 false) 36 ret i64 %z 37 38; SLOW-LABEL: @ctlz( 39; SLOW: entry: 40; SLOW: %cmpz = icmp eq i64 %A, 0 41; SLOW: br i1 %cmpz, label %cond.end, label %cond.false 42; SLOW: cond.false: 43; SLOW: %z = call i64 @llvm.ctlz.i64(i64 %A, i1 true) 44; SLOW: br label %cond.end 45; SLOW: cond.end: 46; SLOW: %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ] 47; SLOW: ret i64 %ctz 48 49; FAST_LZ-LABEL: @ctlz( 50; FAST_LZ: %z = call i64 @llvm.ctlz.i64(i64 %A, i1 false) 51; FAST_LZ: ret i64 %z 52} 53 54declare i64 @llvm.cttz.i64(i64, i1) 55declare i64 @llvm.ctlz.i64(i64, i1) 56 57