160 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 22; Test high-part i128->i256 multiplications on z17.3;4; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z17 | FileCheck %s5 6; Multiply high signed.7define i128 @f1(i128 %a, i128 %b) {8; CHECK-LABEL: f1:9; CHECK: # %bb.0:10; CHECK-NEXT: vl %v0, 0(%r4), 311; CHECK-NEXT: vl %v1, 0(%r3), 312; CHECK-NEXT: vmhq %v0, %v1, %v013; CHECK-NEXT: vst %v0, 0(%r2), 314; CHECK-NEXT: br %r1415 %exta = sext i128 %a to i25616 %extb = sext i128 %b to i25617 %extres = mul i256 %exta, %extb18 %shiftres = lshr i256 %extres, 12819 %res = trunc i256 %shiftres to i12820 ret i128 %res21}22 23; Multiply high unsigned.24define i128 @f2(i128 %a, i128 %b) {25; CHECK-LABEL: f2:26; CHECK: # %bb.0:27; CHECK-NEXT: vl %v0, 0(%r4), 328; CHECK-NEXT: vl %v1, 0(%r3), 329; CHECK-NEXT: vmlhq %v0, %v1, %v030; CHECK-NEXT: vst %v0, 0(%r2), 331; CHECK-NEXT: br %r1432 %exta = zext i128 %a to i25633 %extb = zext i128 %b to i25634 %extres = mul i256 %exta, %extb35 %shiftres = lshr i256 %extres, 12836 %res = trunc i256 %shiftres to i12837 ret i128 %res38}39 40; Multiply-and-add high signed.41define i128 @f3(i128 %a, i128 %b, i128 %add) {42; CHECK-LABEL: f3:43; CHECK: # %bb.0:44; CHECK-NEXT: vl %v0, 0(%r5), 345; CHECK-NEXT: vl %v1, 0(%r4), 346; CHECK-NEXT: vl %v2, 0(%r3), 347; CHECK-NEXT: vmahq %v0, %v2, %v1, %v048; CHECK-NEXT: vst %v0, 0(%r2), 349; CHECK-NEXT: br %r1450 %exta = sext i128 %a to i25651 %extb = sext i128 %b to i25652 %extadd = sext i128 %add to i25653 %extmul = mul i256 %exta, %extb54 %extres = add i256 %extmul, %extadd55 %shiftres = lshr i256 %extres, 12856 %res = trunc i256 %shiftres to i12857 ret i128 %res58}59 60; Multiply-and-add high unsigned.61define i128 @f4(i128 %a, i128 %b, i128 %add) {62; CHECK-LABEL: f4:63; CHECK: # %bb.0:64; CHECK-NEXT: vl %v0, 0(%r5), 365; CHECK-NEXT: vl %v1, 0(%r4), 366; CHECK-NEXT: vl %v2, 0(%r3), 367; CHECK-NEXT: vmalhq %v0, %v2, %v1, %v068; CHECK-NEXT: vst %v0, 0(%r2), 369; CHECK-NEXT: br %r1470 %exta = zext i128 %a to i25671 %extb = zext i128 %b to i25672 %extadd = zext i128 %add to i25673 %extmul = mul i256 %exta, %extb74 %extres = add i256 %extmul, %extadd75 %shiftres = lshr i256 %extres, 12876 %res = trunc i256 %shiftres to i12877 ret i128 %res78}79 80; Multiply-and-add high signed with immediate operand to multiply.81define i128 @f5(i128 %a, i128 %add) {82; CHECK-LABEL: f5:83; CHECK: # %bb.0:84; CHECK-NEXT: larl %r1, .LCPI4_085; CHECK-NEXT: vl %v0, 0(%r4), 386; CHECK-NEXT: vl %v1, 0(%r3), 387; CHECK-NEXT: vl %v2, 0(%r1), 388; CHECK-NEXT: vmahq %v0, %v1, %v2, %v089; CHECK-NEXT: vst %v0, 0(%r2), 390; CHECK-NEXT: br %r1491 %exta = sext i128 %a to i25692 %extadd = sext i128 %add to i25693 %extmul = mul i256 %exta, 1234594 %extres = add i256 %extmul, %extadd95 %shiftres = lshr i256 %extres, 12896 %res = trunc i256 %shiftres to i12897 ret i128 %res98}99 100; Multiply-and-add high unsigned with immediate operand to multiply.101define i128 @f6(i128 %a, i128 %add) {102; CHECK-LABEL: f6:103; CHECK: # %bb.0:104; CHECK-NEXT: larl %r1, .LCPI5_0105; CHECK-NEXT: vl %v0, 0(%r4), 3106; CHECK-NEXT: vl %v1, 0(%r3), 3107; CHECK-NEXT: vl %v2, 0(%r1), 3108; CHECK-NEXT: vmalhq %v0, %v1, %v2, %v0109; CHECK-NEXT: vst %v0, 0(%r2), 3110; CHECK-NEXT: br %r14111 %exta = zext i128 %a to i256112 %extadd = zext i128 %add to i256113 %extmul = mul i256 %exta, 12345114 %extres = add i256 %extmul, %extadd115 %shiftres = lshr i256 %extres, 128116 %res = trunc i256 %shiftres to i128117 ret i128 %res118}119 120; Multiply-and-add high signed with immediate operand to addition.121define i128 @f7(i128 %a, i128 %b) {122; CHECK-LABEL: f7:123; CHECK: # %bb.0:124; CHECK-NEXT: larl %r1, .LCPI6_0125; CHECK-NEXT: vl %v0, 0(%r4), 3126; CHECK-NEXT: vl %v1, 0(%r3), 3127; CHECK-NEXT: vl %v2, 0(%r1), 3128; CHECK-NEXT: vmahq %v0, %v1, %v0, %v2129; CHECK-NEXT: vst %v0, 0(%r2), 3130; CHECK-NEXT: br %r14131 %exta = sext i128 %a to i256132 %extb = sext i128 %b to i256133 %extmul = mul i256 %exta, %extb134 %extres = add i256 %extmul, 12345135 %shiftres = lshr i256 %extres, 128136 %res = trunc i256 %shiftres to i128137 ret i128 %res138}139 140; Multiply-and-add high unsigned with immediate operand to addition.141define i128 @f8(i128 %a, i128 %b) {142; CHECK-LABEL: f8:143; CHECK: # %bb.0:144; CHECK-NEXT: larl %r1, .LCPI7_0145; CHECK-NEXT: vl %v0, 0(%r4), 3146; CHECK-NEXT: vl %v1, 0(%r3), 3147; CHECK-NEXT: vl %v2, 0(%r1), 3148; CHECK-NEXT: vmalhq %v0, %v1, %v0, %v2149; CHECK-NEXT: vst %v0, 0(%r2), 3150; CHECK-NEXT: br %r14151 %exta = zext i128 %a to i256152 %extb = zext i128 %b to i256153 %extmul = mul i256 %exta, %extb154 %extres = add i256 %extmul, 12345155 %shiftres = lshr i256 %extres, 128156 %res = trunc i256 %shiftres to i128157 ret i128 %res158}159 160