1; Test strict multiplication of two f32s, producing an f32 result. 2; 3; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 \ 4; RUN: | FileCheck -check-prefix=CHECK -check-prefix=CHECK-SCALAR %s 5; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 | FileCheck %s 6 7declare float @foo() 8declare float @llvm.experimental.constrained.fmul.f32(float, float, metadata, metadata) 9 10; Check register multiplication. 11define float @f1(float %f1, float %f2) #0 { 12; CHECK-LABEL: f1: 13; CHECK: meebr %f0, %f2 14; CHECK: br %r14 15 %res = call float @llvm.experimental.constrained.fmul.f32( 16 float %f1, float %f2, 17 metadata !"round.dynamic", 18 metadata !"fpexcept.strict") #0 19 ret float %res 20} 21 22; Check the low end of the MEEB range. 23define float @f2(float %f1, float *%ptr) #0 { 24; CHECK-LABEL: f2: 25; CHECK: meeb %f0, 0(%r2) 26; CHECK: br %r14 27 %f2 = load float, float *%ptr 28 %res = call float @llvm.experimental.constrained.fmul.f32( 29 float %f1, float %f2, 30 metadata !"round.dynamic", 31 metadata !"fpexcept.strict") #0 32 ret float %res 33} 34 35; Check the high end of the aligned MEEB range. 36define float @f3(float %f1, float *%base) #0 { 37; CHECK-LABEL: f3: 38; CHECK: meeb %f0, 4092(%r2) 39; CHECK: br %r14 40 %ptr = getelementptr float, float *%base, i64 1023 41 %f2 = load float, float *%ptr 42 %res = call float @llvm.experimental.constrained.fmul.f32( 43 float %f1, float %f2, 44 metadata !"round.dynamic", 45 metadata !"fpexcept.strict") #0 46 ret float %res 47} 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, float *%base) #0 { 52; CHECK-LABEL: f4: 53; CHECK: aghi %r2, 4096 54; CHECK: meeb %f0, 0(%r2) 55; CHECK: br %r14 56 %ptr = getelementptr float, float *%base, i64 1024 57 %f2 = load float, float *%ptr 58 %res = call float @llvm.experimental.constrained.fmul.f32( 59 float %f1, float %f2, 60 metadata !"round.dynamic", 61 metadata !"fpexcept.strict") #0 62 ret float %res 63} 64 65; Check negative displacements, which also need separate address logic. 66define float @f5(float %f1, float *%base) #0 { 67; CHECK-LABEL: f5: 68; CHECK: aghi %r2, -4 69; CHECK: meeb %f0, 0(%r2) 70; CHECK: br %r14 71 %ptr = getelementptr float, float *%base, i64 -1 72 %f2 = load float, float *%ptr 73 %res = call float @llvm.experimental.constrained.fmul.f32( 74 float %f1, float %f2, 75 metadata !"round.dynamic", 76 metadata !"fpexcept.strict") #0 77 ret float %res 78} 79 80; Check that MEEB allows indices. 81define float @f6(float %f1, float *%base, i64 %index) #0 { 82; CHECK-LABEL: f6: 83; CHECK: sllg %r1, %r3, 2 84; CHECK: meeb %f0, 400(%r1,%r2) 85; CHECK: br %r14 86 %ptr1 = getelementptr float, float *%base, i64 %index 87 %ptr2 = getelementptr float, float *%ptr1, i64 100 88 %f2 = load float, float *%ptr2 89 %res = call float @llvm.experimental.constrained.fmul.f32( 90 float %f1, float %f2, 91 metadata !"round.dynamic", 92 metadata !"fpexcept.strict") #0 93 ret float %res 94} 95 96; Check that multiplications of spilled values can use MEEB rather than MEEBR. 97define float @f7(float *%ptr0) #0 { 98; CHECK-LABEL: f7: 99; CHECK: brasl %r14, foo@PLT 100; CHECK-SCALAR: meeb %f0, 16{{[04]}}(%r15) 101; CHECK: br %r14 102 %ptr1 = getelementptr float, float *%ptr0, i64 2 103 %ptr2 = getelementptr float, float *%ptr0, i64 4 104 %ptr3 = getelementptr float, float *%ptr0, i64 6 105 %ptr4 = getelementptr float, float *%ptr0, i64 8 106 %ptr5 = getelementptr float, float *%ptr0, i64 10 107 %ptr6 = getelementptr float, float *%ptr0, i64 12 108 %ptr7 = getelementptr float, float *%ptr0, i64 14 109 %ptr8 = getelementptr float, float *%ptr0, i64 16 110 %ptr9 = getelementptr float, float *%ptr0, i64 18 111 %ptr10 = getelementptr float, float *%ptr0, i64 20 112 113 %val0 = load float, float *%ptr0 114 %val1 = load float, float *%ptr1 115 %val2 = load float, float *%ptr2 116 %val3 = load float, float *%ptr3 117 %val4 = load float, float *%ptr4 118 %val5 = load float, float *%ptr5 119 %val6 = load float, float *%ptr6 120 %val7 = load float, float *%ptr7 121 %val8 = load float, float *%ptr8 122 %val9 = load float, float *%ptr9 123 %val10 = load float, float *%ptr10 124 125 %ret = call float @foo() #0 126 127 %mul0 = call float @llvm.experimental.constrained.fmul.f32( 128 float %ret, float %val0, 129 metadata !"round.dynamic", 130 metadata !"fpexcept.strict") #0 131 %mul1 = call float @llvm.experimental.constrained.fmul.f32( 132 float %mul0, float %val1, 133 metadata !"round.dynamic", 134 metadata !"fpexcept.strict") #0 135 %mul2 = call float @llvm.experimental.constrained.fmul.f32( 136 float %mul1, float %val2, 137 metadata !"round.dynamic", 138 metadata !"fpexcept.strict") #0 139 %mul3 = call float @llvm.experimental.constrained.fmul.f32( 140 float %mul2, float %val3, 141 metadata !"round.dynamic", 142 metadata !"fpexcept.strict") #0 143 %mul4 = call float @llvm.experimental.constrained.fmul.f32( 144 float %mul3, float %val4, 145 metadata !"round.dynamic", 146 metadata !"fpexcept.strict") #0 147 %mul5 = call float @llvm.experimental.constrained.fmul.f32( 148 float %mul4, float %val5, 149 metadata !"round.dynamic", 150 metadata !"fpexcept.strict") #0 151 %mul6 = call float @llvm.experimental.constrained.fmul.f32( 152 float %mul5, float %val6, 153 metadata !"round.dynamic", 154 metadata !"fpexcept.strict") #0 155 %mul7 = call float @llvm.experimental.constrained.fmul.f32( 156 float %mul6, float %val7, 157 metadata !"round.dynamic", 158 metadata !"fpexcept.strict") #0 159 %mul8 = call float @llvm.experimental.constrained.fmul.f32( 160 float %mul7, float %val8, 161 metadata !"round.dynamic", 162 metadata !"fpexcept.strict") #0 163 %mul9 = call float @llvm.experimental.constrained.fmul.f32( 164 float %mul8, float %val9, 165 metadata !"round.dynamic", 166 metadata !"fpexcept.strict") #0 167 %mul10 = call float @llvm.experimental.constrained.fmul.f32( 168 float %mul9, float %val10, 169 metadata !"round.dynamic", 170 metadata !"fpexcept.strict") #0 171 172 ret float %mul10 173} 174 175attributes #0 = { strictfp } 176