1; Test 32-bit square root. 2; 3; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s 4 5declare float @llvm.sqrt.f32(float) 6declare float @sqrtf(float) 7 8; Check register square root. 9define float @f1(float %val) { 10; CHECK-LABEL: f1: 11; CHECK: sqebr %f0, %f0 12; CHECK: br %r14 13 %res = call float @llvm.sqrt.f32(float %val) 14 ret float %res 15} 16 17; Check the low end of the SQEB range. 18define float @f2(float *%ptr) { 19; CHECK-LABEL: f2: 20; CHECK: sqeb %f0, 0(%r2) 21; CHECK: br %r14 22 %val = load float , float *%ptr 23 %res = call float @llvm.sqrt.f32(float %val) 24 ret float %res 25} 26 27; Check the high end of the aligned SQEB range. 28define float @f3(float *%base) { 29; CHECK-LABEL: f3: 30; CHECK: sqeb %f0, 4092(%r2) 31; CHECK: br %r14 32 %ptr = getelementptr float, float *%base, i64 1023 33 %val = load float , float *%ptr 34 %res = call float @llvm.sqrt.f32(float %val) 35 ret float %res 36} 37 38; Check the next word up, which needs separate address logic. 39; Other sequences besides this one would be OK. 40define float @f4(float *%base) { 41; CHECK-LABEL: f4: 42; CHECK: aghi %r2, 4096 43; CHECK: sqeb %f0, 0(%r2) 44; CHECK: br %r14 45 %ptr = getelementptr float, float *%base, i64 1024 46 %val = load float , float *%ptr 47 %res = call float @llvm.sqrt.f32(float %val) 48 ret float %res 49} 50 51; Check negative displacements, which also need separate address logic. 52define float @f5(float *%base) { 53; CHECK-LABEL: f5: 54; CHECK: aghi %r2, -4 55; CHECK: sqeb %f0, 0(%r2) 56; CHECK: br %r14 57 %ptr = getelementptr float, float *%base, i64 -1 58 %val = load float , float *%ptr 59 %res = call float @llvm.sqrt.f32(float %val) 60 ret float %res 61} 62 63; Check that SQEB allows indices. 64define float @f6(float *%base, i64 %index) { 65; CHECK-LABEL: f6: 66; CHECK: sllg %r1, %r3, 2 67; CHECK: sqeb %f0, 400(%r1,%r2) 68; CHECK: br %r14 69 %ptr1 = getelementptr float, float *%base, i64 %index 70 %ptr2 = getelementptr float, float *%ptr1, i64 100 71 %val = load float , float *%ptr2 72 %res = call float @llvm.sqrt.f32(float %val) 73 ret float %res 74} 75 76; Test a case where we spill the source of at least one SQEBR. We want 77; to use SQEB if possible. 78define void @f7(float *%ptr) { 79; CHECK-LABEL: f7: 80; CHECK: sqeb {{%f[0-9]+}}, 16{{[04]}}(%r15) 81; CHECK: br %r14 82 %val0 = load volatile float , float *%ptr 83 %val1 = load volatile float , float *%ptr 84 %val2 = load volatile float , float *%ptr 85 %val3 = load volatile float , float *%ptr 86 %val4 = load volatile float , float *%ptr 87 %val5 = load volatile float , float *%ptr 88 %val6 = load volatile float , float *%ptr 89 %val7 = load volatile float , float *%ptr 90 %val8 = load volatile float , float *%ptr 91 %val9 = load volatile float , float *%ptr 92 %val10 = load volatile float , float *%ptr 93 %val11 = load volatile float , float *%ptr 94 %val12 = load volatile float , float *%ptr 95 %val13 = load volatile float , float *%ptr 96 %val14 = load volatile float , float *%ptr 97 %val15 = load volatile float , float *%ptr 98 %val16 = load volatile float , float *%ptr 99 100 %sqrt0 = call float @llvm.sqrt.f32(float %val0) 101 %sqrt1 = call float @llvm.sqrt.f32(float %val1) 102 %sqrt2 = call float @llvm.sqrt.f32(float %val2) 103 %sqrt3 = call float @llvm.sqrt.f32(float %val3) 104 %sqrt4 = call float @llvm.sqrt.f32(float %val4) 105 %sqrt5 = call float @llvm.sqrt.f32(float %val5) 106 %sqrt6 = call float @llvm.sqrt.f32(float %val6) 107 %sqrt7 = call float @llvm.sqrt.f32(float %val7) 108 %sqrt8 = call float @llvm.sqrt.f32(float %val8) 109 %sqrt9 = call float @llvm.sqrt.f32(float %val9) 110 %sqrt10 = call float @llvm.sqrt.f32(float %val10) 111 %sqrt11 = call float @llvm.sqrt.f32(float %val11) 112 %sqrt12 = call float @llvm.sqrt.f32(float %val12) 113 %sqrt13 = call float @llvm.sqrt.f32(float %val13) 114 %sqrt14 = call float @llvm.sqrt.f32(float %val14) 115 %sqrt15 = call float @llvm.sqrt.f32(float %val15) 116 %sqrt16 = call float @llvm.sqrt.f32(float %val16) 117 118 store volatile float %val0, float *%ptr 119 store volatile float %val1, float *%ptr 120 store volatile float %val2, float *%ptr 121 store volatile float %val3, float *%ptr 122 store volatile float %val4, float *%ptr 123 store volatile float %val5, float *%ptr 124 store volatile float %val6, float *%ptr 125 store volatile float %val7, float *%ptr 126 store volatile float %val8, float *%ptr 127 store volatile float %val9, float *%ptr 128 store volatile float %val10, float *%ptr 129 store volatile float %val11, float *%ptr 130 store volatile float %val12, float *%ptr 131 store volatile float %val13, float *%ptr 132 store volatile float %val14, float *%ptr 133 store volatile float %val15, float *%ptr 134 store volatile float %val16, float *%ptr 135 136 store volatile float %sqrt0, float *%ptr 137 store volatile float %sqrt1, float *%ptr 138 store volatile float %sqrt2, float *%ptr 139 store volatile float %sqrt3, float *%ptr 140 store volatile float %sqrt4, float *%ptr 141 store volatile float %sqrt5, float *%ptr 142 store volatile float %sqrt6, float *%ptr 143 store volatile float %sqrt7, float *%ptr 144 store volatile float %sqrt8, float *%ptr 145 store volatile float %sqrt9, float *%ptr 146 store volatile float %sqrt10, float *%ptr 147 store volatile float %sqrt11, float *%ptr 148 store volatile float %sqrt12, float *%ptr 149 store volatile float %sqrt13, float *%ptr 150 store volatile float %sqrt14, float *%ptr 151 store volatile float %sqrt15, float *%ptr 152 store volatile float %sqrt16, float *%ptr 153 154 ret void 155} 156 157; Check that a call to the normal sqrtf function is lowered. 158define float @f8(float %dummy, float %val) { 159; CHECK-LABEL: f8: 160; CHECK: sqebr %f0, %f2 161; CHECK: cebr %f0, %f0 162; CHECK: bnor %r14 163; CHECK: ler %f0, %f2 164; CHECK: jg sqrtf@PLT 165 %res = tail call float @sqrtf(float %val) 166 ret float %res 167} 168