105 lines · plain
1; RUN: llc < %s -mcpu=cortex-a57 -mattr=+neon -fp-contract=fast -regalloc=pbqp -pbqp-coalescing | FileCheck %s --check-prefix CHECK --check-prefix CHECK-EVEN2; RUN: llc < %s -mcpu=cortex-a57 -mattr=+neon -fp-contract=fast -regalloc=pbqp -pbqp-coalescing | FileCheck %s --check-prefix CHECK --check-prefix CHECK-ODD3;4; Test PBQP is able to fulfill the accumulator chaining constraint.5target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"6target triple = "aarch64"7 8; CHECK-LABEL: fir9; CHECK-EVEN: fmadd {{d[0-9]*[02468]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[02468]}}10; CHECK-EVEN: fmadd {{d[0-9]*[02468]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[02468]}}11; CHECK-EVEN: fmadd {{d[0-9]*[02468]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[02468]}}12; CHECK-EVEN: fmadd {{d[0-9]*[02468]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[02468]}}13; CHECK-EVEN: fmadd {{d[0-9]*[02468]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[02468]}}14; CHECK-EVEN: fmadd {{d[0-9]*[02468]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[02468]}}15; CHECK-EVEN: fmadd {{d[0-9]*[02468]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[02468]}}16; CHECK-ODD: fmadd {{d[0-9]*[13579]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[13579]}}17; CHECK-ODD: fmadd {{d[0-9]*[13579]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[13579]}}18; CHECK-ODD: fmadd {{d[0-9]*[13579]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[13579]}}19; CHECK-ODD: fmadd {{d[0-9]*[13579]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[13579]}}20; CHECK-ODD: fmadd {{d[0-9]*[13579]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[13579]}}21; CHECK-ODD: fmadd {{d[0-9]*[13579]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[13579]}}22; CHECK-ODD: fmadd {{d[0-9]*[13579]}}, {{d[0-9]*}}, {{d[0-9]*}}, {{d[0-9]*[13579]}}23define void @fir(ptr nocapture %rx, ptr nocapture %ry, ptr nocapture %c, ptr nocapture %x, ptr nocapture %y) {24entry:25 %0 = load double, ptr %c, align 826 %1 = load double, ptr %x, align 827 %mul = fmul fast double %1, %028 %2 = load double, ptr %y, align 829 %mul7 = fmul fast double %2, %030 %arrayidx.1 = getelementptr inbounds double, ptr %c, i64 131 %3 = load double, ptr %arrayidx.1, align 832 %arrayidx2.1 = getelementptr inbounds double, ptr %x, i64 133 %4 = load double, ptr %arrayidx2.1, align 834 %mul.1 = fmul fast double %4, %335 %add.1 = fadd fast double %mul.1, %mul36 %arrayidx6.1 = getelementptr inbounds double, ptr %y, i64 137 %5 = load double, ptr %arrayidx6.1, align 838 %mul7.1 = fmul fast double %5, %339 %add8.1 = fadd fast double %mul7.1, %mul740 %arrayidx.2 = getelementptr inbounds double, ptr %c, i64 241 %6 = load double, ptr %arrayidx.2, align 842 %arrayidx2.2 = getelementptr inbounds double, ptr %x, i64 243 %7 = load double, ptr %arrayidx2.2, align 844 %mul.2 = fmul fast double %7, %645 %add.2 = fadd fast double %mul.2, %add.146 %arrayidx6.2 = getelementptr inbounds double, ptr %y, i64 247 %8 = load double, ptr %arrayidx6.2, align 848 %mul7.2 = fmul fast double %8, %649 %add8.2 = fadd fast double %mul7.2, %add8.150 %arrayidx.3 = getelementptr inbounds double, ptr %c, i64 351 %9 = load double, ptr %arrayidx.3, align 852 %arrayidx2.3 = getelementptr inbounds double, ptr %x, i64 353 %10 = load double, ptr %arrayidx2.3, align 854 %mul.3 = fmul fast double %10, %955 %add.3 = fadd fast double %mul.3, %add.256 %arrayidx6.3 = getelementptr inbounds double, ptr %y, i64 357 %11 = load double, ptr %arrayidx6.3, align 858 %mul7.3 = fmul fast double %11, %959 %add8.3 = fadd fast double %mul7.3, %add8.260 %arrayidx.4 = getelementptr inbounds double, ptr %c, i64 461 %12 = load double, ptr %arrayidx.4, align 862 %arrayidx2.4 = getelementptr inbounds double, ptr %x, i64 463 %13 = load double, ptr %arrayidx2.4, align 864 %mul.4 = fmul fast double %13, %1265 %add.4 = fadd fast double %mul.4, %add.366 %arrayidx6.4 = getelementptr inbounds double, ptr %y, i64 467 %14 = load double, ptr %arrayidx6.4, align 868 %mul7.4 = fmul fast double %14, %1269 %add8.4 = fadd fast double %mul7.4, %add8.370 %arrayidx.5 = getelementptr inbounds double, ptr %c, i64 571 %15 = load double, ptr %arrayidx.5, align 872 %arrayidx2.5 = getelementptr inbounds double, ptr %x, i64 573 %16 = load double, ptr %arrayidx2.5, align 874 %mul.5 = fmul fast double %16, %1575 %add.5 = fadd fast double %mul.5, %add.476 %arrayidx6.5 = getelementptr inbounds double, ptr %y, i64 577 %17 = load double, ptr %arrayidx6.5, align 878 %mul7.5 = fmul fast double %17, %1579 %add8.5 = fadd fast double %mul7.5, %add8.480 %arrayidx.6 = getelementptr inbounds double, ptr %c, i64 681 %18 = load double, ptr %arrayidx.6, align 882 %arrayidx2.6 = getelementptr inbounds double, ptr %x, i64 683 %19 = load double, ptr %arrayidx2.6, align 884 %mul.6 = fmul fast double %19, %1885 %add.6 = fadd fast double %mul.6, %add.586 %arrayidx6.6 = getelementptr inbounds double, ptr %y, i64 687 %20 = load double, ptr %arrayidx6.6, align 888 %mul7.6 = fmul fast double %20, %1889 %add8.6 = fadd fast double %mul7.6, %add8.590 %arrayidx.7 = getelementptr inbounds double, ptr %c, i64 791 %21 = load double, ptr %arrayidx.7, align 892 %arrayidx2.7 = getelementptr inbounds double, ptr %x, i64 793 %22 = load double, ptr %arrayidx2.7, align 894 %mul.7 = fmul fast double %22, %2195 %add.7 = fadd fast double %mul.7, %add.696 %arrayidx6.7 = getelementptr inbounds double, ptr %y, i64 797 %23 = load double, ptr %arrayidx6.7, align 898 %mul7.7 = fmul fast double %23, %2199 %add8.7 = fadd fast double %mul7.7, %add8.6100 store double %add.7, ptr %rx, align 8101 store double %add8.7, ptr %ry, align 8102 ret void103}104 105