1; Test high-part i64->i128 multiplications. 2; 3; RUN: llc < %s -mtriple=s390x-linux-gnu -asm-verbose=0 | FileCheck %s 4 5declare i64 @foo() 6 7; Check zero-extended multiplication in which only the high part is used. 8define i64 @f1(i64 %dummy, i64 %a, i64 %b) { 9; CHECK-LABEL: f1: 10; CHECK-NOT: {{%r[234]}} 11; CHECK: mlgr %r2, %r4 12; CHECK: br %r14 13 %ax = zext i64 %a to i128 14 %bx = zext i64 %b to i128 15 %mulx = mul i128 %ax, %bx 16 %highx = lshr i128 %mulx, 64 17 %high = trunc i128 %highx to i64 18 ret i64 %high 19} 20 21; Check sign-extended multiplication in which only the high part is used. 22; This needs a rather convoluted sequence. 23define i64 @f2(i64 %dummy, i64 %a, i64 %b) { 24; CHECK-LABEL: f2: 25; CHECK-DAG: srag [[RES1:%r[0-5]]], %r3, 63 26; CHECK-DAG: srag [[RES2:%r[0-5]]], %r4, 63 27; CHECK-DAG: ngr [[RES1]], %r4 28; CHECK-DAG: ngr [[RES2]], %r3 29; CHECK-DAG: agr [[RES2]], [[RES1]] 30; CHECK-DAG: mlgr %r2, %r4 31; CHECK: sgr %r2, [[RES2]] 32; CHECK: br %r14 33 %ax = sext i64 %a to i128 34 %bx = sext i64 %b to i128 35 %mulx = mul i128 %ax, %bx 36 %highx = lshr i128 %mulx, 64 37 %high = trunc i128 %highx to i64 38 ret i64 %high 39} 40 41; Check zero-extended multiplication in which only part of the high half 42; is used. 43define i64 @f3(i64 %dummy, i64 %a, i64 %b) { 44; CHECK-LABEL: f3: 45; CHECK-NOT: {{%r[234]}} 46; CHECK: mlgr %r2, %r4 47; CHECK: srlg %r2, %r2, 3 48; CHECK: br %r14 49 %ax = zext i64 %a to i128 50 %bx = zext i64 %b to i128 51 %mulx = mul i128 %ax, %bx 52 %highx = lshr i128 %mulx, 67 53 %high = trunc i128 %highx to i64 54 ret i64 %high 55} 56 57; Check zero-extended multiplication in which the result is split into 58; high and low halves. 59define i64 @f4(i64 %dummy, i64 %a, i64 %b) { 60; CHECK-LABEL: f4: 61; CHECK-NOT: {{%r[234]}} 62; CHECK: mlgr %r2, %r4 63; CHECK: ogr %r3, %r2 64; CHECK: lgr %r2, %r3 65; CHECK: br %r14 66 %ax = zext i64 %a to i128 67 %bx = zext i64 %b to i128 68 %mulx = mul i128 %ax, %bx 69 %highx = lshr i128 %mulx, 64 70 %high = trunc i128 %highx to i64 71 %low = trunc i128 %mulx to i64 72 %or = or i64 %high, %low 73 ret i64 %or 74} 75 76; Check division by a constant, which should use multiplication instead. 77define i64 @f5(i64 %dummy, i64 %a) { 78; CHECK-LABEL: f5: 79; CHECK: mlgr %r2, 80; CHECK: srlg %r2, %r2, 81; CHECK: br %r14 82 %res = udiv i64 %a, 1234 83 ret i64 %res 84} 85 86; Check MLG with no displacement. 87define i64 @f6(i64 %dummy, i64 %a, ptr %src) { 88; CHECK-LABEL: f6: 89; CHECK-NOT: {{%r[234]}} 90; CHECK: mlg %r2, 0(%r4) 91; CHECK: br %r14 92 %b = load i64, ptr %src 93 %ax = zext i64 %a to i128 94 %bx = zext i64 %b to i128 95 %mulx = mul i128 %ax, %bx 96 %highx = lshr i128 %mulx, 64 97 %high = trunc i128 %highx to i64 98 ret i64 %high 99} 100 101; Check the high end of the aligned MLG range. 102define i64 @f7(i64 %dummy, i64 %a, ptr %src) { 103; CHECK-LABEL: f7: 104; CHECK: mlg %r2, 524280(%r4) 105; CHECK: br %r14 106 %ptr = getelementptr i64, ptr %src, i64 65535 107 %b = load i64, ptr %ptr 108 %ax = zext i64 %a to i128 109 %bx = zext i64 %b to i128 110 %mulx = mul i128 %ax, %bx 111 %highx = lshr i128 %mulx, 64 112 %high = trunc i128 %highx to i64 113 ret i64 %high 114} 115 116; Check the next doubleword up, which requires separate address logic. 117; Other sequences besides this one would be OK. 118define i64 @f8(i64 %dummy, i64 %a, ptr %src) { 119; CHECK-LABEL: f8: 120; CHECK: agfi %r4, 524288 121; CHECK: mlg %r2, 0(%r4) 122; CHECK: br %r14 123 %ptr = getelementptr i64, ptr %src, i64 65536 124 %b = load i64, ptr %ptr 125 %ax = zext i64 %a to i128 126 %bx = zext i64 %b to i128 127 %mulx = mul i128 %ax, %bx 128 %highx = lshr i128 %mulx, 64 129 %high = trunc i128 %highx to i64 130 ret i64 %high 131} 132 133; Check the high end of the negative aligned MLG range. 134define i64 @f9(i64 %dummy, i64 %a, ptr %src) { 135; CHECK-LABEL: f9: 136; CHECK: mlg %r2, -8(%r4) 137; CHECK: br %r14 138 %ptr = getelementptr i64, ptr %src, i64 -1 139 %b = load i64, ptr %ptr 140 %ax = zext i64 %a to i128 141 %bx = zext i64 %b to i128 142 %mulx = mul i128 %ax, %bx 143 %highx = lshr i128 %mulx, 64 144 %high = trunc i128 %highx to i64 145 ret i64 %high 146} 147 148; Check the low end of the MLG range. 149define i64 @f10(i64 %dummy, i64 %a, ptr %src) { 150; CHECK-LABEL: f10: 151; CHECK: mlg %r2, -524288(%r4) 152; CHECK: br %r14 153 %ptr = getelementptr i64, ptr %src, i64 -65536 154 %b = load i64, ptr %ptr 155 %ax = zext i64 %a to i128 156 %bx = zext i64 %b to i128 157 %mulx = mul i128 %ax, %bx 158 %highx = lshr i128 %mulx, 64 159 %high = trunc i128 %highx to i64 160 ret i64 %high 161} 162 163; Check the next doubleword down, which needs separate address logic. 164; Other sequences besides this one would be OK. 165define i64 @f11(ptr %dest, i64 %a, ptr %src) { 166; CHECK-LABEL: f11: 167; CHECK: agfi %r4, -524296 168; CHECK: mlg %r2, 0(%r4) 169; CHECK: br %r14 170 %ptr = getelementptr i64, ptr %src, i64 -65537 171 %b = load i64, ptr %ptr 172 %ax = zext i64 %a to i128 173 %bx = zext i64 %b to i128 174 %mulx = mul i128 %ax, %bx 175 %highx = lshr i128 %mulx, 64 176 %high = trunc i128 %highx to i64 177 ret i64 %high 178} 179 180; Check that MLG allows an index. 181define i64 @f12(ptr %dest, i64 %a, i64 %src, i64 %index) { 182; CHECK-LABEL: f12: 183; CHECK: mlg %r2, 524287(%r5,%r4) 184; CHECK: br %r14 185 %add1 = add i64 %src, %index 186 %add2 = add i64 %add1, 524287 187 %ptr = inttoptr i64 %add2 to ptr 188 %b = load i64, ptr %ptr 189 %ax = zext i64 %a to i128 190 %bx = zext i64 %b to i128 191 %mulx = mul i128 %ax, %bx 192 %highx = lshr i128 %mulx, 64 193 %high = trunc i128 %highx to i64 194 ret i64 %high 195} 196 197; Check that multiplications of spilled values can use MLG rather than MLGR. 198define i64 @f13(ptr %ptr0) { 199; CHECK-LABEL: f13: 200; CHECK: brasl %r14, foo@PLT 201; CHECK: mlg {{%r[0-9]+}}, 160(%r15) 202; CHECK: br %r14 203 %ptr1 = getelementptr i64, ptr %ptr0, i64 2 204 %ptr2 = getelementptr i64, ptr %ptr0, i64 4 205 %ptr3 = getelementptr i64, ptr %ptr0, i64 6 206 %ptr4 = getelementptr i64, ptr %ptr0, i64 8 207 %ptr5 = getelementptr i64, ptr %ptr0, i64 10 208 %ptr6 = getelementptr i64, ptr %ptr0, i64 12 209 %ptr7 = getelementptr i64, ptr %ptr0, i64 14 210 %ptr8 = getelementptr i64, ptr %ptr0, i64 16 211 %ptr9 = getelementptr i64, ptr %ptr0, i64 18 212 213 %val0 = load i64, ptr %ptr0 214 %val1 = load i64, ptr %ptr1 215 %val2 = load i64, ptr %ptr2 216 %val3 = load i64, ptr %ptr3 217 %val4 = load i64, ptr %ptr4 218 %val5 = load i64, ptr %ptr5 219 %val6 = load i64, ptr %ptr6 220 %val7 = load i64, ptr %ptr7 221 %val8 = load i64, ptr %ptr8 222 %val9 = load i64, ptr %ptr9 223 224 %ret = call i64 @foo() 225 226 %retx = zext i64 %ret to i128 227 %val0x = zext i64 %val0 to i128 228 %mul0d = mul i128 %retx, %val0x 229 %mul0x = lshr i128 %mul0d, 64 230 231 %val1x = zext i64 %val1 to i128 232 %mul1d = mul i128 %mul0x, %val1x 233 %mul1x = lshr i128 %mul1d, 64 234 235 %val2x = zext i64 %val2 to i128 236 %mul2d = mul i128 %mul1x, %val2x 237 %mul2x = lshr i128 %mul2d, 64 238 239 %val3x = zext i64 %val3 to i128 240 %mul3d = mul i128 %mul2x, %val3x 241 %mul3x = lshr i128 %mul3d, 64 242 243 %val4x = zext i64 %val4 to i128 244 %mul4d = mul i128 %mul3x, %val4x 245 %mul4x = lshr i128 %mul4d, 64 246 247 %val5x = zext i64 %val5 to i128 248 %mul5d = mul i128 %mul4x, %val5x 249 %mul5x = lshr i128 %mul5d, 64 250 251 %val6x = zext i64 %val6 to i128 252 %mul6d = mul i128 %mul5x, %val6x 253 %mul6x = lshr i128 %mul6d, 64 254 255 %val7x = zext i64 %val7 to i128 256 %mul7d = mul i128 %mul6x, %val7x 257 %mul7x = lshr i128 %mul7d, 64 258 259 %val8x = zext i64 %val8 to i128 260 %mul8d = mul i128 %mul7x, %val8x 261 %mul8x = lshr i128 %mul8d, 64 262 263 %val9x = zext i64 %val9 to i128 264 %mul9d = mul i128 %mul8x, %val9x 265 %mul9x = lshr i128 %mul9d, 64 266 267 %mul9 = trunc i128 %mul9x to i64 268 ret i64 %mul9 269} 270