148 lines · plain
1; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 \2; RUN: | FileCheck -check-prefix=CHECK -check-prefix=CHECK-SCALAR %s3; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 \4; RUN: | FileCheck -check-prefix=CHECK -check-prefix=CHECK-VECTOR %s5 6declare float @llvm.experimental.constrained.fma.f32(float %f1, float %f2, float %f3, metadata, metadata)7 8define float @f1(float %f1, float %f2, float %acc) #0 {9; CHECK-LABEL: f1:10; CHECK-SCALAR: msebr %f4, %f0, %f211; CHECK-SCALAR: ler %f0, %f412; CHECK-VECTOR: wfmssb %f0, %f0, %f2, %f413; CHECK: br %r1414 %negacc = fneg float %acc15 %res = call float @llvm.experimental.constrained.fma.f32 (16 float %f1, float %f2, float %negacc,17 metadata !"round.dynamic",18 metadata !"fpexcept.strict") #019 ret float %res20}21 22define float @f2(float %f1, ptr %ptr, float %acc) #0 {23; CHECK-LABEL: f2:24; CHECK: mseb %f2, %f0, 0(%r2)25; CHECK-SCALAR: ler %f0, %f226; CHECK-VECTOR: ldr %f0, %f227; CHECK: br %r1428 %f2 = load float, ptr %ptr29 %negacc = fneg float %acc30 %res = call float @llvm.experimental.constrained.fma.f32 (31 float %f1, float %f2, float %negacc,32 metadata !"round.dynamic",33 metadata !"fpexcept.strict") #034 ret float %res35}36 37define float @f3(float %f1, ptr %base, float %acc) #0 {38; CHECK-LABEL: f3:39; CHECK: mseb %f2, %f0, 4092(%r2)40; CHECK-SCALAR: ler %f0, %f241; CHECK-VECTOR: ldr %f0, %f242; CHECK: br %r1443 %ptr = getelementptr float, ptr %base, i64 102344 %f2 = load float, ptr %ptr45 %negacc = fneg float %acc46 %res = call float @llvm.experimental.constrained.fma.f32 (47 float %f1, float %f2, float %negacc,48 metadata !"round.dynamic",49 metadata !"fpexcept.strict") #050 ret float %res51}52 53define float @f4(float %f1, ptr %base, float %acc) #0 {54; The important thing here is that we don't generate an out-of-range55; displacement. Other sequences besides this one would be OK.56;57; CHECK-LABEL: f4:58; CHECK: aghi %r2, 409659; CHECK: mseb %f2, %f0, 0(%r2)60; CHECK-SCALAR: ler %f0, %f261; CHECK-VECTOR: ldr %f0, %f262; CHECK: br %r1463 %ptr = getelementptr float, ptr %base, i64 102464 %f2 = load float, ptr %ptr65 %negacc = fneg float %acc66 %res = call float @llvm.experimental.constrained.fma.f32 (67 float %f1, float %f2, float %negacc,68 metadata !"round.dynamic",69 metadata !"fpexcept.strict") #070 ret float %res71}72 73define float @f5(float %f1, ptr %base, float %acc) #0 {74; Here too the important thing is that we don't generate an out-of-range75; displacement. Other sequences besides this one would be OK.76;77; CHECK-LABEL: f5:78; CHECK: aghi %r2, -479; CHECK: mseb %f2, %f0, 0(%r2)80; CHECK-SCALAR: ler %f0, %f281; CHECK-VECTOR: ldr %f0, %f282; CHECK: br %r1483 %ptr = getelementptr float, ptr %base, i64 -184 %f2 = load float, ptr %ptr85 %negacc = fneg float %acc86 %res = call float @llvm.experimental.constrained.fma.f32 (87 float %f1, float %f2, float %negacc,88 metadata !"round.dynamic",89 metadata !"fpexcept.strict") #090 ret float %res91}92 93define float @f6(float %f1, ptr %base, i64 %index, float %acc) #0 {94; CHECK-LABEL: f6:95; CHECK: sllg %r1, %r3, 296; CHECK: mseb %f2, %f0, 0(%r1,%r2)97; CHECK-SCALAR: ler %f0, %f298; CHECK-VECTOR: ldr %f0, %f299; CHECK: br %r14100 %ptr = getelementptr float, ptr %base, i64 %index101 %f2 = load float, ptr %ptr102 %negacc = fneg float %acc103 %res = call float @llvm.experimental.constrained.fma.f32 (104 float %f1, float %f2, float %negacc,105 metadata !"round.dynamic",106 metadata !"fpexcept.strict") #0107 ret float %res108}109 110define float @f7(float %f1, ptr %base, i64 %index, float %acc) #0 {111; CHECK-LABEL: f7:112; CHECK: sllg %r1, %r3, 2113; CHECK: mseb %f2, %f0, 4092({{%r1,%r2|%r2,%r1}})114; CHECK-SCALAR: ler %f0, %f2115; CHECK-VECTOR: ldr %f0, %f2116; CHECK: br %r14117 %index2 = add i64 %index, 1023118 %ptr = getelementptr float, ptr %base, i64 %index2119 %f2 = load float, ptr %ptr120 %negacc = fneg float %acc121 %res = call float @llvm.experimental.constrained.fma.f32 (122 float %f1, float %f2, float %negacc,123 metadata !"round.dynamic",124 metadata !"fpexcept.strict") #0125 ret float %res126}127 128define float @f8(float %f1, ptr %base, i64 %index, float %acc) #0 {129; CHECK-LABEL: f8:130; CHECK: sllg %r1, %r3, 2131; CHECK: lay %r1, 4096({{%r1,%r2|%r2,%r1}})132; CHECK: mseb %f2, %f0, 0(%r1)133; CHECK-SCALAR: ler %f0, %f2134; CHECK-VECTOR: ldr %f0, %f2135; CHECK: br %r14136 %index2 = add i64 %index, 1024137 %ptr = getelementptr float, ptr %base, i64 %index2138 %f2 = load float, ptr %ptr139 %negacc = fneg float %acc140 %res = call float @llvm.experimental.constrained.fma.f32 (141 float %f1, float %f2, float %negacc,142 metadata !"round.dynamic",143 metadata !"fpexcept.strict") #0144 ret float %res145}146 147attributes #0 = { strictfp }148