1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3 2; Test high-part i64->i128 multiplications on z13. 3; 4; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z13 | FileCheck %s 5 6; Check zero-extended multiplication in which only the high part is used. 7define i64 @f1(i64 %dummy, i64 %a, i64 %b) { 8; CHECK-LABEL: f1: 9; CHECK: # %bb.0: 10; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 11; CHECK-NEXT: mlgr %r2, %r4 12; CHECK-NEXT: # kill: def $r2d killed $r2d killed $r2q 13; CHECK-NEXT: br %r14 14 %ax = zext i64 %a to i128 15 %bx = zext i64 %b to i128 16 %mulx = mul i128 %ax, %bx 17 %highx = lshr i128 %mulx, 64 18 %high = trunc i128 %highx to i64 19 ret i64 %high 20} 21 22; Check sign-extended multiplication in which only the high part is used. 23; This needs a rather convoluted sequence. 24define i64 @f2(i64 %dummy, i64 %a, i64 %b) { 25; CHECK-LABEL: f2: 26; CHECK: # %bb.0: 27; CHECK-NEXT: srag %r1, %r4, 63 28; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 29; CHECK-NEXT: srag %r0, %r3, 63 30; CHECK-NEXT: ngr %r1, %r3 31; CHECK-NEXT: mlgr %r2, %r4 32; CHECK-NEXT: ngr %r0, %r4 33; CHECK-NEXT: agr %r0, %r1 34; CHECK-NEXT: sgr %r2, %r0 35; CHECK-NEXT: br %r14 36 %ax = sext i64 %a to i128 37 %bx = sext i64 %b to i128 38 %mulx = mul i128 %ax, %bx 39 %highx = lshr i128 %mulx, 64 40 %high = trunc i128 %highx to i64 41 ret i64 %high 42} 43 44; Check zero-extended multiplication in which only part of the high half 45; is used. 46define i64 @f3(i64 %dummy, i64 %a, i64 %b) { 47; CHECK-LABEL: f3: 48; CHECK: # %bb.0: 49; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 50; CHECK-NEXT: mlgr %r2, %r4 51; CHECK-NEXT: srlg %r2, %r2, 3 52; CHECK-NEXT: br %r14 53 %ax = zext i64 %a to i128 54 %bx = zext i64 %b to i128 55 %mulx = mul i128 %ax, %bx 56 %highx = lshr i128 %mulx, 67 57 %high = trunc i128 %highx to i64 58 ret i64 %high 59} 60 61; Check zero-extended multiplication in which the result is split into 62; high and low halves. 63define i64 @f4(i64 %dummy, i64 %a, i64 %b) { 64; CHECK-LABEL: f4: 65; CHECK: # %bb.0: 66; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 67; CHECK-NEXT: mlgr %r2, %r4 68; CHECK-NEXT: ogr %r2, %r3 69; CHECK-NEXT: br %r14 70 %ax = zext i64 %a to i128 71 %bx = zext i64 %b to i128 72 %mulx = mul i128 %ax, %bx 73 %highx = lshr i128 %mulx, 64 74 %high = trunc i128 %highx to i64 75 %low = trunc i128 %mulx to i64 76 %or = or i64 %high, %low 77 ret i64 %or 78} 79 80; Check division by a constant, which should use multiplication instead. 81define i64 @f5(i64 %dummy, i64 %a) { 82; CHECK-LABEL: f5: 83; CHECK: # %bb.0: 84; CHECK-NEXT: llihf %r0, 1782028570 85; CHECK-NEXT: oilf %r0, 598650223 86; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 87; CHECK-NEXT: mlgr %r2, %r0 88; CHECK-NEXT: srlg %r2, %r2, 9 89; CHECK-NEXT: br %r14 90 %res = udiv i64 %a, 1234 91 ret i64 %res 92} 93 94; Check MLG with no displacement. 95define i64 @f6(i64 %dummy, i64 %a, ptr %src) { 96; CHECK-LABEL: f6: 97; CHECK: # %bb.0: 98; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 99; CHECK-NEXT: mlg %r2, 0(%r4) 100; CHECK-NEXT: # kill: def $r2d killed $r2d killed $r2q 101; CHECK-NEXT: br %r14 102 %b = load i64, ptr %src 103 %ax = zext i64 %a to i128 104 %bx = zext i64 %b to i128 105 %mulx = mul i128 %ax, %bx 106 %highx = lshr i128 %mulx, 64 107 %high = trunc i128 %highx to i64 108 ret i64 %high 109} 110 111; Check the high end of the aligned MLG range. 112define i64 @f7(i64 %dummy, i64 %a, ptr %src) { 113; CHECK-LABEL: f7: 114; CHECK: # %bb.0: 115; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 116; CHECK-NEXT: mlg %r2, 524280(%r4) 117; CHECK-NEXT: # kill: def $r2d killed $r2d killed $r2q 118; CHECK-NEXT: br %r14 119 %ptr = getelementptr i64, ptr %src, i64 65535 120 %b = load i64, ptr %ptr 121 %ax = zext i64 %a to i128 122 %bx = zext i64 %b to i128 123 %mulx = mul i128 %ax, %bx 124 %highx = lshr i128 %mulx, 64 125 %high = trunc i128 %highx to i64 126 ret i64 %high 127} 128 129; Check the next doubleword up, which requires separate address logic. 130; Other sequences besides this one would be OK. 131define i64 @f8(i64 %dummy, i64 %a, ptr %src) { 132; CHECK-LABEL: f8: 133; CHECK: # %bb.0: 134; CHECK-NEXT: agfi %r4, 524288 135; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 136; CHECK-NEXT: mlg %r2, 0(%r4) 137; CHECK-NEXT: # kill: def $r2d killed $r2d killed $r2q 138; CHECK-NEXT: br %r14 139 %ptr = getelementptr i64, ptr %src, i64 65536 140 %b = load i64, ptr %ptr 141 %ax = zext i64 %a to i128 142 %bx = zext i64 %b to i128 143 %mulx = mul i128 %ax, %bx 144 %highx = lshr i128 %mulx, 64 145 %high = trunc i128 %highx to i64 146 ret i64 %high 147} 148 149; Check the high end of the negative aligned MLG range. 150define i64 @f9(i64 %dummy, i64 %a, ptr %src) { 151; CHECK-LABEL: f9: 152; CHECK: # %bb.0: 153; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 154; CHECK-NEXT: mlg %r2, -8(%r4) 155; CHECK-NEXT: # kill: def $r2d killed $r2d killed $r2q 156; CHECK-NEXT: br %r14 157 %ptr = getelementptr i64, ptr %src, i64 -1 158 %b = load i64, ptr %ptr 159 %ax = zext i64 %a to i128 160 %bx = zext i64 %b to i128 161 %mulx = mul i128 %ax, %bx 162 %highx = lshr i128 %mulx, 64 163 %high = trunc i128 %highx to i64 164 ret i64 %high 165} 166 167; Check the low end of the MLG range. 168define i64 @f10(i64 %dummy, i64 %a, ptr %src) { 169; CHECK-LABEL: f10: 170; CHECK: # %bb.0: 171; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 172; CHECK-NEXT: mlg %r2, -524288(%r4) 173; CHECK-NEXT: # kill: def $r2d killed $r2d killed $r2q 174; CHECK-NEXT: br %r14 175 %ptr = getelementptr i64, ptr %src, i64 -65536 176 %b = load i64, ptr %ptr 177 %ax = zext i64 %a to i128 178 %bx = zext i64 %b to i128 179 %mulx = mul i128 %ax, %bx 180 %highx = lshr i128 %mulx, 64 181 %high = trunc i128 %highx to i64 182 ret i64 %high 183} 184 185; Check the next doubleword down, which needs separate address logic. 186; Other sequences besides this one would be OK. 187define i64 @f11(ptr %dest, i64 %a, ptr %src) { 188; CHECK-LABEL: f11: 189; CHECK: # %bb.0: 190; CHECK-NEXT: agfi %r4, -524296 191; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 192; CHECK-NEXT: mlg %r2, 0(%r4) 193; CHECK-NEXT: # kill: def $r2d killed $r2d killed $r2q 194; CHECK-NEXT: br %r14 195 %ptr = getelementptr i64, ptr %src, i64 -65537 196 %b = load i64, ptr %ptr 197 %ax = zext i64 %a to i128 198 %bx = zext i64 %b to i128 199 %mulx = mul i128 %ax, %bx 200 %highx = lshr i128 %mulx, 64 201 %high = trunc i128 %highx to i64 202 ret i64 %high 203} 204 205; Check that MLG allows an index. 206define i64 @f12(ptr %dest, i64 %a, i64 %src, i64 %index) { 207; CHECK-LABEL: f12: 208; CHECK: # %bb.0: 209; CHECK-NEXT: # kill: def $r3d killed $r3d def $r2q 210; CHECK-NEXT: mlg %r2, 524287(%r5,%r4) 211; CHECK-NEXT: # kill: def $r2d killed $r2d killed $r2q 212; CHECK-NEXT: br %r14 213 %add1 = add i64 %src, %index 214 %add2 = add i64 %add1, 524287 215 %ptr = inttoptr i64 %add2 to ptr 216 %b = load i64, ptr %ptr 217 %ax = zext i64 %a to i128 218 %bx = zext i64 %b to i128 219 %mulx = mul i128 %ax, %bx 220 %highx = lshr i128 %mulx, 64 221 %high = trunc i128 %highx to i64 222 ret i64 %high 223} 224 225