176 lines · plain
1; Test 32-bit floating-point strict addition.2;3; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 \4; RUN: | FileCheck -check-prefix=CHECK -check-prefix=CHECK-SCALAR %s5; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 | FileCheck %s6 7declare float @foo()8declare float @llvm.experimental.constrained.fadd.f32(float, float, metadata, metadata)9 10; Check register addition.11define float @f1(float %f1, float %f2) #0 {12; CHECK-LABEL: f1:13; CHECK: aebr %f0, %f214; CHECK: br %r1415 %res = call float @llvm.experimental.constrained.fadd.f32(16 float %f1, float %f2,17 metadata !"round.dynamic",18 metadata !"fpexcept.strict") #019 ret float %res20}21 22; Check the low end of the AEB range.23define float @f2(float %f1, ptr %ptr) #0 {24; CHECK-LABEL: f2:25; CHECK: aeb %f0, 0(%r2)26; CHECK: br %r1427 %f2 = load float, ptr %ptr28 %res = call float @llvm.experimental.constrained.fadd.f32(29 float %f1, float %f2,30 metadata !"round.dynamic",31 metadata !"fpexcept.strict") #032 ret float %res33}34 35; Check the high end of the aligned AEB range.36define float @f3(float %f1, ptr %base) #0 {37; CHECK-LABEL: f3:38; CHECK: aeb %f0, 4092(%r2)39; CHECK: br %r1440 %ptr = getelementptr float, ptr %base, i64 102341 %f2 = load float, ptr %ptr42 %res = call float @llvm.experimental.constrained.fadd.f32(43 float %f1, float %f2,44 metadata !"round.dynamic",45 metadata !"fpexcept.strict") #046 ret float %res47}48 49; Check the next word up, which needs separate address logic.50; Other sequences besides this one would be OK.51define float @f4(float %f1, ptr %base) #0 {52; CHECK-LABEL: f4:53; CHECK: aghi %r2, 409654; CHECK: aeb %f0, 0(%r2)55; CHECK: br %r1456 %ptr = getelementptr float, ptr %base, i64 102457 %f2 = load float, ptr %ptr58 %res = call float @llvm.experimental.constrained.fadd.f32(59 float %f1, float %f2,60 metadata !"round.dynamic",61 metadata !"fpexcept.strict") #062 ret float %res63}64 65; Check negative displacements, which also need separate address logic.66define float @f5(float %f1, ptr %base) #0 {67; CHECK-LABEL: f5:68; CHECK: aghi %r2, -469; CHECK: aeb %f0, 0(%r2)70; CHECK: br %r1471 %ptr = getelementptr float, ptr %base, i64 -172 %f2 = load float, ptr %ptr73 %res = call float @llvm.experimental.constrained.fadd.f32(74 float %f1, float %f2,75 metadata !"round.dynamic",76 metadata !"fpexcept.strict") #077 ret float %res78}79 80; Check that AEB allows indices.81define float @f6(float %f1, ptr %base, i64 %index) #0 {82; CHECK-LABEL: f6:83; CHECK: sllg %r1, %r3, 284; CHECK: aeb %f0, 400(%r1,%r2)85; CHECK: br %r1486 %ptr1 = getelementptr float, ptr %base, i64 %index87 %ptr2 = getelementptr float, ptr %ptr1, i64 10088 %f2 = load float, ptr %ptr289 %res = call float @llvm.experimental.constrained.fadd.f32(90 float %f1, float %f2,91 metadata !"round.dynamic",92 metadata !"fpexcept.strict") #093 ret float %res94}95 96; Check that additions of spilled values can use AEB rather than AEBR.97define float @f7(ptr %ptr0) #0 {98; CHECK-LABEL: f7:99; CHECK: brasl %r14, foo@PLT100; CHECK-SCALAR: aeb %f0, 16{{[04]}}(%r15)101; CHECK: br %r14102 %ptr1 = getelementptr float, ptr %ptr0, i64 2103 %ptr2 = getelementptr float, ptr %ptr0, i64 4104 %ptr3 = getelementptr float, ptr %ptr0, i64 6105 %ptr4 = getelementptr float, ptr %ptr0, i64 8106 %ptr5 = getelementptr float, ptr %ptr0, i64 10107 %ptr6 = getelementptr float, ptr %ptr0, i64 12108 %ptr7 = getelementptr float, ptr %ptr0, i64 14109 %ptr8 = getelementptr float, ptr %ptr0, i64 16110 %ptr9 = getelementptr float, ptr %ptr0, i64 18111 %ptr10 = getelementptr float, ptr %ptr0, i64 20112 113 %val0 = load float, ptr %ptr0114 %val1 = load float, ptr %ptr1115 %val2 = load float, ptr %ptr2116 %val3 = load float, ptr %ptr3117 %val4 = load float, ptr %ptr4118 %val5 = load float, ptr %ptr5119 %val6 = load float, ptr %ptr6120 %val7 = load float, ptr %ptr7121 %val8 = load float, ptr %ptr8122 %val9 = load float, ptr %ptr9123 %val10 = load float, ptr %ptr10124 125 %ret = call float @foo() #0126 127 %add0 = call float @llvm.experimental.constrained.fadd.f32(128 float %ret, float %val0,129 metadata !"round.dynamic",130 metadata !"fpexcept.strict") #0131 %add1 = call float @llvm.experimental.constrained.fadd.f32(132 float %add0, float %val1,133 metadata !"round.dynamic",134 metadata !"fpexcept.strict") #0135 %add2 = call float @llvm.experimental.constrained.fadd.f32(136 float %add1, float %val2,137 metadata !"round.dynamic",138 metadata !"fpexcept.strict") #0139 %add3 = call float @llvm.experimental.constrained.fadd.f32(140 float %add2, float %val3,141 metadata !"round.dynamic",142 metadata !"fpexcept.strict") #0143 %add4 = call float @llvm.experimental.constrained.fadd.f32(144 float %add3, float %val4,145 metadata !"round.dynamic",146 metadata !"fpexcept.strict") #0147 %add5 = call float @llvm.experimental.constrained.fadd.f32(148 float %add4, float %val5,149 metadata !"round.dynamic",150 metadata !"fpexcept.strict") #0151 %add6 = call float @llvm.experimental.constrained.fadd.f32(152 float %add5, float %val6,153 metadata !"round.dynamic",154 metadata !"fpexcept.strict") #0155 %add7 = call float @llvm.experimental.constrained.fadd.f32(156 float %add6, float %val7,157 metadata !"round.dynamic",158 metadata !"fpexcept.strict") #0159 %add8 = call float @llvm.experimental.constrained.fadd.f32(160 float %add7, float %val8,161 metadata !"round.dynamic",162 metadata !"fpexcept.strict") #0163 %add9 = call float @llvm.experimental.constrained.fadd.f32(164 float %add8, float %val9,165 metadata !"round.dynamic",166 metadata !"fpexcept.strict") #0167 %add10 = call float @llvm.experimental.constrained.fadd.f32(168 float %add9, float %val10,169 metadata !"round.dynamic",170 metadata !"fpexcept.strict") #0171 172 ret float %add10173}174 175attributes #0 = { strictfp }176