1; Test 64-bit square root. 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=z13 | FileCheck %s 6 7declare double @llvm.sqrt.f64(double %f) 8declare double @sqrt(double) 9 10; Check register square root. 11define double @f1(double %val) { 12; CHECK-LABEL: f1: 13; CHECK: sqdbr %f0, %f0 14; CHECK: br %r14 15 %res = call double @llvm.sqrt.f64(double %val) 16 ret double %res 17} 18 19; Check the low end of the SQDB range. 20define double @f2(double *%ptr) { 21; CHECK-LABEL: f2: 22; CHECK: sqdb %f0, 0(%r2) 23; CHECK: br %r14 24 %val = load double , double *%ptr 25 %res = call double @llvm.sqrt.f64(double %val) 26 ret double %res 27} 28 29; Check the high end of the aligned SQDB range. 30define double @f3(double *%base) { 31; CHECK-LABEL: f3: 32; CHECK: sqdb %f0, 4088(%r2) 33; CHECK: br %r14 34 %ptr = getelementptr double, double *%base, i64 511 35 %val = load double , double *%ptr 36 %res = call double @llvm.sqrt.f64(double %val) 37 ret double %res 38} 39 40; Check the next doubleword up, which needs separate address logic. 41; Other sequences besides this one would be OK. 42define double @f4(double *%base) { 43; CHECK-LABEL: f4: 44; CHECK: aghi %r2, 4096 45; CHECK: sqdb %f0, 0(%r2) 46; CHECK: br %r14 47 %ptr = getelementptr double, double *%base, i64 512 48 %val = load double , double *%ptr 49 %res = call double @llvm.sqrt.f64(double %val) 50 ret double %res 51} 52 53; Check negative displacements, which also need separate address logic. 54define double @f5(double *%base) { 55; CHECK-LABEL: f5: 56; CHECK: aghi %r2, -8 57; CHECK: sqdb %f0, 0(%r2) 58; CHECK: br %r14 59 %ptr = getelementptr double, double *%base, i64 -1 60 %val = load double , double *%ptr 61 %res = call double @llvm.sqrt.f64(double %val) 62 ret double %res 63} 64 65; Check that SQDB allows indices. 66define double @f6(double *%base, i64 %index) { 67; CHECK-LABEL: f6: 68; CHECK: sllg %r1, %r3, 3 69; CHECK: sqdb %f0, 800(%r1,%r2) 70; CHECK: br %r14 71 %ptr1 = getelementptr double, double *%base, i64 %index 72 %ptr2 = getelementptr double, double *%ptr1, i64 100 73 %val = load double , double *%ptr2 74 %res = call double @llvm.sqrt.f64(double %val) 75 ret double %res 76} 77 78; Test a case where we spill the source of at least one SQDBR. We want 79; to use SQDB if possible. 80define void @f7(double *%ptr) { 81; CHECK-LABEL: f7: 82; CHECK-SCALAR: sqdb {{%f[0-9]+}}, 160(%r15) 83; CHECK: br %r14 84 %val0 = load volatile double , double *%ptr 85 %val1 = load volatile double , double *%ptr 86 %val2 = load volatile double , double *%ptr 87 %val3 = load volatile double , double *%ptr 88 %val4 = load volatile double , double *%ptr 89 %val5 = load volatile double , double *%ptr 90 %val6 = load volatile double , double *%ptr 91 %val7 = load volatile double , double *%ptr 92 %val8 = load volatile double , double *%ptr 93 %val9 = load volatile double , double *%ptr 94 %val10 = load volatile double , double *%ptr 95 %val11 = load volatile double , double *%ptr 96 %val12 = load volatile double , double *%ptr 97 %val13 = load volatile double , double *%ptr 98 %val14 = load volatile double , double *%ptr 99 %val15 = load volatile double , double *%ptr 100 %val16 = load volatile double , double *%ptr 101 102 %sqrt0 = call double @llvm.sqrt.f64(double %val0) 103 %sqrt1 = call double @llvm.sqrt.f64(double %val1) 104 %sqrt2 = call double @llvm.sqrt.f64(double %val2) 105 %sqrt3 = call double @llvm.sqrt.f64(double %val3) 106 %sqrt4 = call double @llvm.sqrt.f64(double %val4) 107 %sqrt5 = call double @llvm.sqrt.f64(double %val5) 108 %sqrt6 = call double @llvm.sqrt.f64(double %val6) 109 %sqrt7 = call double @llvm.sqrt.f64(double %val7) 110 %sqrt8 = call double @llvm.sqrt.f64(double %val8) 111 %sqrt9 = call double @llvm.sqrt.f64(double %val9) 112 %sqrt10 = call double @llvm.sqrt.f64(double %val10) 113 %sqrt11 = call double @llvm.sqrt.f64(double %val11) 114 %sqrt12 = call double @llvm.sqrt.f64(double %val12) 115 %sqrt13 = call double @llvm.sqrt.f64(double %val13) 116 %sqrt14 = call double @llvm.sqrt.f64(double %val14) 117 %sqrt15 = call double @llvm.sqrt.f64(double %val15) 118 %sqrt16 = call double @llvm.sqrt.f64(double %val16) 119 120 store volatile double %val0, double *%ptr 121 store volatile double %val1, double *%ptr 122 store volatile double %val2, double *%ptr 123 store volatile double %val3, double *%ptr 124 store volatile double %val4, double *%ptr 125 store volatile double %val5, double *%ptr 126 store volatile double %val6, double *%ptr 127 store volatile double %val7, double *%ptr 128 store volatile double %val8, double *%ptr 129 store volatile double %val9, double *%ptr 130 store volatile double %val10, double *%ptr 131 store volatile double %val11, double *%ptr 132 store volatile double %val12, double *%ptr 133 store volatile double %val13, double *%ptr 134 store volatile double %val14, double *%ptr 135 store volatile double %val15, double *%ptr 136 store volatile double %val16, double *%ptr 137 138 store volatile double %sqrt0, double *%ptr 139 store volatile double %sqrt1, double *%ptr 140 store volatile double %sqrt2, double *%ptr 141 store volatile double %sqrt3, double *%ptr 142 store volatile double %sqrt4, double *%ptr 143 store volatile double %sqrt5, double *%ptr 144 store volatile double %sqrt6, double *%ptr 145 store volatile double %sqrt7, double *%ptr 146 store volatile double %sqrt8, double *%ptr 147 store volatile double %sqrt9, double *%ptr 148 store volatile double %sqrt10, double *%ptr 149 store volatile double %sqrt11, double *%ptr 150 store volatile double %sqrt12, double *%ptr 151 store volatile double %sqrt13, double *%ptr 152 store volatile double %sqrt14, double *%ptr 153 store volatile double %sqrt15, double *%ptr 154 store volatile double %sqrt16, double *%ptr 155 156 ret void 157} 158 159; Check that a call to the normal sqrt function is lowered. 160define double @f8(double %dummy, double %val) { 161; CHECK-LABEL: f8: 162; CHECK: sqdbr %f0, %f2 163; CHECK: cdbr %f0, %f0 164; CHECK: bnor %r14 165; CHECK: ldr %f0, %f2 166; CHECK: jg sqrt@PLT 167 %res = tail call double @sqrt(double %val) 168 ret double %res 169} 170