1*9880d681SAndroid Build Coastguard Worker; RUN: llc -march=x86-64 -mcpu=core-avx2 < %s | FileCheck %s 2*9880d681SAndroid Build Coastguard Worker; 3*9880d681SAndroid Build Coastguard Worker; Test multiple peephole-time folds in a single basic block. 4*9880d681SAndroid Build Coastguard Worker; <rdar://problem/16478629> 5*9880d681SAndroid Build Coastguard Worker 6*9880d681SAndroid Build Coastguard Workerdefine <8 x float> @test_peephole_multi_fold(<8 x float>* %p1, <8 x float>* %p2) { 7*9880d681SAndroid Build Coastguard Workerentry: 8*9880d681SAndroid Build Coastguard Worker br label %loopbody 9*9880d681SAndroid Build Coastguard Worker 10*9880d681SAndroid Build Coastguard Workerloopbody: 11*9880d681SAndroid Build Coastguard Worker; CHECK: test_peephole_multi_fold: 12*9880d681SAndroid Build Coastguard Worker; CHECK: vfmadd231ps ({{%rdi|%rcx}}), 13*9880d681SAndroid Build Coastguard Worker; CHECK: vfmadd231ps ({{%rsi|%rdx}}), 14*9880d681SAndroid Build Coastguard Worker %vsum1 = phi <8 x float> [ %vsum1.next, %loopbody ], [ zeroinitializer, %entry ] 15*9880d681SAndroid Build Coastguard Worker %vsum2 = phi <8 x float> [ %vsum2.next, %loopbody ], [ zeroinitializer, %entry ] 16*9880d681SAndroid Build Coastguard Worker %m1 = load <8 x float>, <8 x float>* %p1, align 1 17*9880d681SAndroid Build Coastguard Worker %m2 = load <8 x float>, <8 x float>* %p2, align 1 18*9880d681SAndroid Build Coastguard Worker %vsum1.next = tail call <8 x float> @llvm.x86.fma.vfmadd.ps.256(<8 x float> %m1, <8 x float> zeroinitializer, <8 x float> %vsum1) 19*9880d681SAndroid Build Coastguard Worker %vsum2.next = tail call <8 x float> @llvm.x86.fma.vfmadd.ps.256(<8 x float> %m2, <8 x float> zeroinitializer, <8 x float> %vsum2) 20*9880d681SAndroid Build Coastguard Worker %vsum1.next.1 = extractelement <8 x float> %vsum1.next, i32 0 21*9880d681SAndroid Build Coastguard Worker %c = fcmp oeq float %vsum1.next.1, 0.0 22*9880d681SAndroid Build Coastguard Worker br i1 %c, label %loopbody, label %loopexit 23*9880d681SAndroid Build Coastguard Worker 24*9880d681SAndroid Build Coastguard Workerloopexit: 25*9880d681SAndroid Build Coastguard Worker %r = fadd <8 x float> %vsum1.next, %vsum2.next 26*9880d681SAndroid Build Coastguard Worker ret <8 x float> %r 27*9880d681SAndroid Build Coastguard Worker} 28*9880d681SAndroid Build Coastguard Worker 29*9880d681SAndroid Build Coastguard Workerdeclare <8 x float> @llvm.x86.fma.vfmadd.ps.256(<8 x float>, <8 x float>, <8 x float>) 30