250 lines · plain
1; Test 64-bit floating-point signaling comparison. The tests assume a z102; implementation of select, using conditional branches rather than LOCGR.3;4; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 \5; RUN: | FileCheck -check-prefix=CHECK -check-prefix=CHECK-SCALAR %s6; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z13 -verify-machineinstrs\7; RUN: | FileCheck -check-prefix=CHECK -check-prefix=CHECK-VECTOR %s8 9declare double @foo()10 11; Check comparison with registers.12define i64 @f1(i64 %a, i64 %b, double %f1, double %f2) #0 {13; CHECK-LABEL: f1:14; CHECK: kdbr %f0, %f215; CHECK-SCALAR-NEXT: ber %r1416; CHECK-SCALAR: lgr %r2, %r317; CHECK-VECTOR-NEXT: locgrne %r2, %r318; CHECK: br %r1419 %cond = call i1 @llvm.experimental.constrained.fcmps.f64(20 double %f1, double %f2,21 metadata !"oeq",22 metadata !"fpexcept.strict") #023 %res = select i1 %cond, i64 %a, i64 %b24 ret i64 %res25}26 27; Check the low end of the KDB range.28define i64 @f2(i64 %a, i64 %b, double %f1, ptr %ptr) #0 {29; CHECK-LABEL: f2:30; CHECK: kdb %f0, 0(%r4)31; CHECK-SCALAR-NEXT: ber %r1432; CHECK-SCALAR: lgr %r2, %r333; CHECK-VECTOR-NEXT: locgrne %r2, %r334; CHECK: br %r1435 %f2 = load double, ptr %ptr36 %cond = call i1 @llvm.experimental.constrained.fcmps.f64(37 double %f1, double %f2,38 metadata !"oeq",39 metadata !"fpexcept.strict") #040 %res = select i1 %cond, i64 %a, i64 %b41 ret i64 %res42}43 44; Check the high end of the aligned KDB range.45define i64 @f3(i64 %a, i64 %b, double %f1, ptr %base) #0 {46; CHECK-LABEL: f3:47; CHECK: kdb %f0, 4088(%r4)48; CHECK-SCALAR-NEXT: ber %r1449; CHECK-SCALAR: lgr %r2, %r350; CHECK-VECTOR-NEXT: locgrne %r2, %r351; CHECK: br %r1452 %ptr = getelementptr double, ptr %base, i64 51153 %f2 = load double, ptr %ptr54 %cond = call i1 @llvm.experimental.constrained.fcmps.f64(55 double %f1, double %f2,56 metadata !"oeq",57 metadata !"fpexcept.strict") #058 %res = select i1 %cond, i64 %a, i64 %b59 ret i64 %res60}61 62; Check the next doubleword up, which needs separate address logic.63; Other sequences besides this one would be OK.64define i64 @f4(i64 %a, i64 %b, double %f1, ptr %base) #0 {65; CHECK-LABEL: f4:66; CHECK: aghi %r4, 409667; CHECK: kdb %f0, 0(%r4)68; CHECK-SCALAR-NEXT: ber %r1469; CHECK-SCALAR: lgr %r2, %r370; CHECK-VECTOR-NEXT: locgrne %r2, %r371; CHECK: br %r1472 %ptr = getelementptr double, ptr %base, i64 51273 %f2 = load double, ptr %ptr74 %cond = call i1 @llvm.experimental.constrained.fcmps.f64(75 double %f1, double %f2,76 metadata !"oeq",77 metadata !"fpexcept.strict") #078 %res = select i1 %cond, i64 %a, i64 %b79 ret i64 %res80}81 82; Check negative displacements, which also need separate address logic.83define i64 @f5(i64 %a, i64 %b, double %f1, ptr %base) #0 {84; CHECK-LABEL: f5:85; CHECK: aghi %r4, -886; CHECK: kdb %f0, 0(%r4)87; CHECK-SCALAR-NEXT: ber %r1488; CHECK-SCALAR: lgr %r2, %r389; CHECK-VECTOR-NEXT: locgrne %r2, %r390; CHECK: br %r1491 %ptr = getelementptr double, ptr %base, i64 -192 %f2 = load double, ptr %ptr93 %cond = call i1 @llvm.experimental.constrained.fcmps.f64(94 double %f1, double %f2,95 metadata !"oeq",96 metadata !"fpexcept.strict") #097 %res = select i1 %cond, i64 %a, i64 %b98 ret i64 %res99}100 101; Check that KDB allows indices.102define i64 @f6(i64 %a, i64 %b, double %f1, ptr %base, i64 %index) #0 {103; CHECK-LABEL: f6:104; CHECK: sllg %r1, %r5, 3105; CHECK: kdb %f0, 800(%r1,%r4)106; CHECK-SCALAR-NEXT: ber %r14107; CHECK-SCALAR: lgr %r2, %r3108; CHECK-VECTOR-NEXT: locgrne %r2, %r3109; CHECK: br %r14110 %ptr1 = getelementptr double, ptr %base, i64 %index111 %ptr2 = getelementptr double, ptr %ptr1, i64 100112 %f2 = load double, ptr %ptr2113 %cond = call i1 @llvm.experimental.constrained.fcmps.f64(114 double %f1, double %f2,115 metadata !"oeq",116 metadata !"fpexcept.strict") #0117 %res = select i1 %cond, i64 %a, i64 %b118 ret i64 %res119}120 121; Check that comparisons of spilled values can use KDB rather than KDBR.122define double @f7(ptr %ptr0) #0 {123; CHECK-LABEL: f7:124; CHECK: brasl %r14, foo@PLT125; CHECK-SCALAR: kdb {{%f[0-9]+}}, 160(%r15)126; CHECK: br %r14127 %ptr1 = getelementptr double, ptr %ptr0, i64 2128 %ptr2 = getelementptr double, ptr %ptr0, i64 4129 %ptr3 = getelementptr double, ptr %ptr0, i64 6130 %ptr4 = getelementptr double, ptr %ptr0, i64 8131 %ptr5 = getelementptr double, ptr %ptr0, i64 10132 %ptr6 = getelementptr double, ptr %ptr0, i64 12133 %ptr7 = getelementptr double, ptr %ptr0, i64 14134 %ptr8 = getelementptr double, ptr %ptr0, i64 16135 %ptr9 = getelementptr double, ptr %ptr0, i64 18136 %ptr10 = getelementptr double, ptr %ptr0, i64 20137 138 %val0 = load double, ptr %ptr0139 %val1 = load double, ptr %ptr1140 %val2 = load double, ptr %ptr2141 %val3 = load double, ptr %ptr3142 %val4 = load double, ptr %ptr4143 %val5 = load double, ptr %ptr5144 %val6 = load double, ptr %ptr6145 %val7 = load double, ptr %ptr7146 %val8 = load double, ptr %ptr8147 %val9 = load double, ptr %ptr9148 %val10 = load double, ptr %ptr10149 150 %ret = call double @foo() #0151 152 %cmp0 = call i1 @llvm.experimental.constrained.fcmps.f64(153 double %ret, double %val0,154 metadata !"olt",155 metadata !"fpexcept.strict") #0156 %cmp1 = call i1 @llvm.experimental.constrained.fcmps.f64(157 double %ret, double %val1,158 metadata !"olt",159 metadata !"fpexcept.strict") #0160 %cmp2 = call i1 @llvm.experimental.constrained.fcmps.f64(161 double %ret, double %val2,162 metadata !"olt",163 metadata !"fpexcept.strict") #0164 %cmp3 = call i1 @llvm.experimental.constrained.fcmps.f64(165 double %ret, double %val3,166 metadata !"olt",167 metadata !"fpexcept.strict") #0168 %cmp4 = call i1 @llvm.experimental.constrained.fcmps.f64(169 double %ret, double %val4,170 metadata !"olt",171 metadata !"fpexcept.strict") #0172 %cmp5 = call i1 @llvm.experimental.constrained.fcmps.f64(173 double %ret, double %val5,174 metadata !"olt",175 metadata !"fpexcept.strict") #0176 %cmp6 = call i1 @llvm.experimental.constrained.fcmps.f64(177 double %ret, double %val6,178 metadata !"olt",179 metadata !"fpexcept.strict") #0180 %cmp7 = call i1 @llvm.experimental.constrained.fcmps.f64(181 double %ret, double %val7,182 metadata !"olt",183 metadata !"fpexcept.strict") #0184 %cmp8 = call i1 @llvm.experimental.constrained.fcmps.f64(185 double %ret, double %val8,186 metadata !"olt",187 metadata !"fpexcept.strict") #0188 %cmp9 = call i1 @llvm.experimental.constrained.fcmps.f64(189 double %ret, double %val9,190 metadata !"olt",191 metadata !"fpexcept.strict") #0192 %cmp10 = call i1 @llvm.experimental.constrained.fcmps.f64(193 double %ret, double %val10,194 metadata !"olt",195 metadata !"fpexcept.strict") #0196 197 %sel0 = select i1 %cmp0, double %ret, double 0.0198 %sel1 = select i1 %cmp1, double %sel0, double 1.0199 %sel2 = select i1 %cmp2, double %sel1, double 2.0200 %sel3 = select i1 %cmp3, double %sel2, double 3.0201 %sel4 = select i1 %cmp4, double %sel3, double 4.0202 %sel5 = select i1 %cmp5, double %sel4, double 5.0203 %sel6 = select i1 %cmp6, double %sel5, double 6.0204 %sel7 = select i1 %cmp7, double %sel6, double 7.0205 %sel8 = select i1 %cmp8, double %sel7, double 8.0206 %sel9 = select i1 %cmp9, double %sel8, double 9.0207 %sel10 = select i1 %cmp10, double %sel9, double 10.0208 209 ret double %sel10210}211 212; Check comparison with zero - cannot use LOAD AND TEST.213define i64 @f8(i64 %a, i64 %b, double %f) #0 {214; CHECK-LABEL: f8:215; CHECK: lzdr [[REG:%f[0-9]+]]216; CHECK-NEXT: kdbr %f0, [[REG]]217; CHECK-SCALAR-NEXT: ber %r14218; CHECK-SCALAR: lgr %r2, %r3219; CHECK-VECTOR-NEXT: locgrne %r2, %r3220; CHECK: br %r14221 %cond = call i1 @llvm.experimental.constrained.fcmps.f64(222 double %f, double 0.0,223 metadata !"oeq",224 metadata !"fpexcept.strict") #0225 %res = select i1 %cond, i64 %a, i64 %b226 ret i64 %res227}228 229; Check the comparison can be reversed if that allows KDB to be used,230define i64 @f9(i64 %a, i64 %b, double %f2, ptr %ptr) #0 {231; CHECK-LABEL: f9:232; CHECK: kdb %f0, 0(%r4)233; CHECK-SCALAR-NEXT: blr %r14234; CHECK-SCALAR: lgr %r2, %r3235; CHECK-VECTOR-NEXT: locgrnl %r2, %r3236; CHECK: br %r14237 %f1 = load double, ptr %ptr238 %cond = call i1 @llvm.experimental.constrained.fcmps.f64(239 double %f1, double %f2,240 metadata !"ogt",241 metadata !"fpexcept.strict") #0242 %res = select i1 %cond, i64 %a, i64 %b243 ret i64 %res244}245 246attributes #0 = { strictfp }247 248declare i1 @llvm.experimental.constrained.fcmps.f64(double, double, metadata, metadata)249 250