304 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -mtriple=thumbv7-unknown-linux-gnueabihf -arm-parallel-dsp -dce %s -S -o - | FileCheck %s3 4define i32 @first_mul_invalid(ptr nocapture readonly %in, ptr nocapture readonly %b) {5; CHECK-LABEL: @first_mul_invalid(6; CHECK-NEXT: entry:7; CHECK-NEXT: [[TMP0:%.*]] = load i16, ptr [[IN:%.*]], align 28; CHECK-NEXT: [[CONV:%.*]] = sext i16 [[TMP0]] to i329; CHECK-NEXT: [[TMP1:%.*]] = load i16, ptr [[B:%.*]], align 210; CHECK-NEXT: [[CONV2:%.*]] = sext i16 [[TMP1]] to i3211; CHECK-NEXT: [[CALL:%.*]] = tail call i32 @bar(i32 [[CONV]], i32 [[CONV2]])12; CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds i16, ptr [[IN]], i32 -113; CHECK-NEXT: [[TMP2:%.*]] = load i16, ptr [[ARRAYIDX3]], align 214; CHECK-NEXT: [[CONV4:%.*]] = sext i16 [[TMP2]] to i3215; CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, ptr [[B]], i32 116; CHECK-NEXT: [[TMP3:%.*]] = load i16, ptr [[ARRAYIDX5]], align 217; CHECK-NEXT: [[CONV6:%.*]] = sext i16 [[TMP3]] to i3218; CHECK-NEXT: [[MUL:%.*]] = mul nsw i32 [[CONV6]], [[CONV4]]19; CHECK-NEXT: [[ADD:%.*]] = add i32 [[MUL]], [[CALL]]20; CHECK-NEXT: [[ARRAYIDX13:%.*]] = getelementptr inbounds i16, ptr [[IN]], i32 -321; CHECK-NEXT: [[TMP5:%.*]] = load i32, ptr [[ARRAYIDX13]], align 222; CHECK-NEXT: [[ARRAYIDX9:%.*]] = getelementptr inbounds i16, ptr [[B]], i32 223; CHECK-NEXT: [[TMP7:%.*]] = load i32, ptr [[ARRAYIDX9]], align 224; CHECK-NEXT: [[TMP8:%.*]] = call i32 @llvm.arm.smladx(i32 [[TMP5]], i32 [[TMP7]], i32 [[ADD]])25; CHECK-NEXT: [[ARRAYIDX25:%.*]] = getelementptr inbounds i16, ptr [[IN]], i32 -526; CHECK-NEXT: [[TMP10:%.*]] = load i32, ptr [[ARRAYIDX25]], align 227; CHECK-NEXT: [[ARRAYIDX21:%.*]] = getelementptr inbounds i16, ptr [[B]], i32 428; CHECK-NEXT: [[TMP12:%.*]] = load i32, ptr [[ARRAYIDX21]], align 229; CHECK-NEXT: [[TMP13:%.*]] = call i32 @llvm.arm.smladx(i32 [[TMP10]], i32 [[TMP12]], i32 [[TMP8]])30; CHECK-NEXT: ret i32 [[TMP13]]31;32entry:33 %0 = load i16, ptr %in, align 234 %conv = sext i16 %0 to i3235 %1 = load i16, ptr %b, align 236 %conv2 = sext i16 %1 to i3237 %call = tail call i32 @bar(i32 %conv, i32 %conv2)38 %arrayidx3 = getelementptr inbounds i16, ptr %in, i32 -139 %2 = load i16, ptr %arrayidx3, align 240 %conv4 = sext i16 %2 to i3241 %arrayidx5 = getelementptr inbounds i16, ptr %b, i32 142 %3 = load i16, ptr %arrayidx5, align 243 %conv6 = sext i16 %3 to i3244 %mul = mul nsw i32 %conv6, %conv445 %add = add i32 %mul, %call46 %arrayidx7 = getelementptr inbounds i16, ptr %in, i32 -247 %4 = load i16, ptr %arrayidx7, align 248 %conv8 = sext i16 %4 to i3249 %arrayidx9 = getelementptr inbounds i16, ptr %b, i32 250 %5 = load i16, ptr %arrayidx9, align 251 %conv10 = sext i16 %5 to i3252 %mul11 = mul nsw i32 %conv10, %conv853 %add12 = add i32 %add, %mul1154 %arrayidx13 = getelementptr inbounds i16, ptr %in, i32 -355 %6 = load i16, ptr %arrayidx13, align 256 %conv14 = sext i16 %6 to i3257 %arrayidx15 = getelementptr inbounds i16, ptr %b, i32 358 %7 = load i16, ptr %arrayidx15, align 259 %conv16 = sext i16 %7 to i3260 %mul17 = mul nsw i32 %conv16, %conv1461 %add18 = add i32 %add12, %mul1762 %arrayidx19 = getelementptr inbounds i16, ptr %in, i32 -463 %8 = load i16, ptr %arrayidx19, align 264 %conv20 = sext i16 %8 to i3265 %arrayidx21 = getelementptr inbounds i16, ptr %b, i32 466 %9 = load i16, ptr %arrayidx21, align 267 %conv22 = sext i16 %9 to i3268 %mul23 = mul nsw i32 %conv22, %conv2069 %add24 = add i32 %add18, %mul2370 %arrayidx25 = getelementptr inbounds i16, ptr %in, i32 -571 %10 = load i16, ptr %arrayidx25, align 272 %conv26 = sext i16 %10 to i3273 %arrayidx27 = getelementptr inbounds i16, ptr %b, i32 574 %11 = load i16, ptr %arrayidx27, align 275 %conv28 = sext i16 %11 to i3276 %mul29 = mul nsw i32 %conv28, %conv2677 %add30 = add i32 %add24, %mul2978 ret i32 %add3079}80 81define i32 @with_no_acc_input(ptr nocapture readonly %in, ptr nocapture readonly %b) {82; CHECK-LABEL: @with_no_acc_input(83; CHECK-NEXT: entry:84; CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds i16, ptr [[IN:%.*]], i32 -185; CHECK-NEXT: [[LD_2:%.*]] = load i16, ptr [[ARRAYIDX3]], align 286; CHECK-NEXT: [[CONV4:%.*]] = sext i16 [[LD_2]] to i3287; CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, ptr [[B:%.*]], i32 188; CHECK-NEXT: [[LD_3:%.*]] = load i16, ptr [[ARRAYIDX5]], align 289; CHECK-NEXT: [[CONV6:%.*]] = sext i16 [[LD_3]] to i3290; CHECK-NEXT: [[MUL:%.*]] = mul nsw i32 [[CONV6]], [[CONV4]]91; CHECK-NEXT: [[ARRAYIDX13:%.*]] = getelementptr inbounds i16, ptr [[IN]], i32 -392; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[ARRAYIDX13]], align 293; CHECK-NEXT: [[ARRAYIDX9:%.*]] = getelementptr inbounds i16, ptr [[B]], i32 294; CHECK-NEXT: [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX9]], align 295; CHECK-NEXT: [[TMP4:%.*]] = call i32 @llvm.arm.smladx(i32 [[TMP1]], i32 [[TMP3]], i32 [[MUL]])96; CHECK-NEXT: [[ARRAYIDX25:%.*]] = getelementptr inbounds i16, ptr [[IN]], i32 -597; CHECK-NEXT: [[TMP6:%.*]] = load i32, ptr [[ARRAYIDX25]], align 298; CHECK-NEXT: [[ARRAYIDX21:%.*]] = getelementptr inbounds i16, ptr [[B]], i32 499; CHECK-NEXT: [[TMP8:%.*]] = load i32, ptr [[ARRAYIDX21]], align 2100; CHECK-NEXT: [[TMP9:%.*]] = call i32 @llvm.arm.smladx(i32 [[TMP6]], i32 [[TMP8]], i32 [[TMP4]])101; CHECK-NEXT: ret i32 [[TMP9]]102;103entry:104 %arrayidx3 = getelementptr inbounds i16, ptr %in, i32 -1105 %ld.2 = load i16, ptr %arrayidx3, align 2106 %conv4 = sext i16 %ld.2 to i32107 %arrayidx5 = getelementptr inbounds i16, ptr %b, i32 1108 %ld.3 = load i16, ptr %arrayidx5, align 2109 %conv6 = sext i16 %ld.3 to i32110 %mul = mul nsw i32 %conv6, %conv4111 %arrayidx7 = getelementptr inbounds i16, ptr %in, i32 -2112 %ld.4 = load i16, ptr %arrayidx7, align 2113 %conv8 = sext i16 %ld.4 to i32114 %arrayidx9 = getelementptr inbounds i16, ptr %b, i32 2115 %ld.5 = load i16, ptr %arrayidx9, align 2116 %conv10 = sext i16 %ld.5 to i32117 %mul11 = mul nsw i32 %conv10, %conv8118 %add12 = add i32 %mul, %mul11119 %arrayidx13 = getelementptr inbounds i16, ptr %in, i32 -3120 %ld.6 = load i16, ptr %arrayidx13, align 2121 %conv14 = sext i16 %ld.6 to i32122 %arrayidx15 = getelementptr inbounds i16, ptr %b, i32 3123 %ld.7 = load i16, ptr %arrayidx15, align 2124 %conv16 = sext i16 %ld.7 to i32125 %mul17 = mul nsw i32 %conv16, %conv14126 %add18 = add i32 %add12, %mul17127 %arrayidx19 = getelementptr inbounds i16, ptr %in, i32 -4128 %ld.8 = load i16, ptr %arrayidx19, align 2129 %conv20 = sext i16 %ld.8 to i32130 %arrayidx21 = getelementptr inbounds i16, ptr %b, i32 4131 %ld.9 = load i16, ptr %arrayidx21, align 2132 %conv22 = sext i16 %ld.9 to i32133 %mul23 = mul nsw i32 %conv22, %conv20134 %add24 = add i32 %add18, %mul23135 %arrayidx25 = getelementptr inbounds i16, ptr %in, i32 -5136 %ld.10 = load i16, ptr %arrayidx25, align 2137 %conv26 = sext i16 %ld.10 to i32138 %arrayidx27 = getelementptr inbounds i16, ptr %b, i32 5139 %ld.11 = load i16, ptr %arrayidx27, align 2140 %conv28 = sext i16 %ld.11 to i32141 %mul29 = mul nsw i32 %conv28, %conv26142 %add30 = add i32 %add24, %mul29143 ret i32 %add30144}145 146define i64 @with_64bit_acc(ptr nocapture readonly %in, ptr nocapture readonly %b) {147; CHECK-LABEL: @with_64bit_acc(148; CHECK-NEXT: entry:149; CHECK-NEXT: [[TMP0:%.*]] = load i16, ptr [[IN:%.*]], align 2150; CHECK-NEXT: [[CONV:%.*]] = sext i16 [[TMP0]] to i32151; CHECK-NEXT: [[TMP1:%.*]] = load i16, ptr [[B:%.*]], align 2152; CHECK-NEXT: [[CONV2:%.*]] = sext i16 [[TMP1]] to i32153; CHECK-NEXT: [[CALL:%.*]] = tail call i32 @bar(i32 [[CONV]], i32 [[CONV2]])154; CHECK-NEXT: [[SEXT_0:%.*]] = sext i32 [[CALL]] to i64155; CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds i16, ptr [[IN]], i32 -1156; CHECK-NEXT: [[TMP2:%.*]] = load i16, ptr [[ARRAYIDX3]], align 2157; CHECK-NEXT: [[CONV4:%.*]] = sext i16 [[TMP2]] to i32158; CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, ptr [[B]], i32 1159; CHECK-NEXT: [[TMP3:%.*]] = load i16, ptr [[ARRAYIDX5]], align 2160; CHECK-NEXT: [[CONV6:%.*]] = sext i16 [[TMP3]] to i32161; CHECK-NEXT: [[MUL:%.*]] = mul nsw i32 [[CONV6]], [[CONV4]]162; CHECK-NEXT: [[SEXT_1:%.*]] = sext i32 [[MUL]] to i64163; CHECK-NEXT: [[ADD:%.*]] = add i64 [[SEXT_0]], [[SEXT_1]]164; CHECK-NEXT: [[ARRAYIDX13:%.*]] = getelementptr inbounds i16, ptr [[IN]], i32 -3165; CHECK-NEXT: [[TMP5:%.*]] = load i32, ptr [[ARRAYIDX13]], align 2166; CHECK-NEXT: [[ARRAYIDX9:%.*]] = getelementptr inbounds i16, ptr [[B]], i32 2167; CHECK-NEXT: [[TMP7:%.*]] = load i32, ptr [[ARRAYIDX9]], align 2168; CHECK-NEXT: [[TMP8:%.*]] = call i64 @llvm.arm.smlaldx(i32 [[TMP5]], i32 [[TMP7]], i64 [[ADD]])169; CHECK-NEXT: [[ARRAYIDX25:%.*]] = getelementptr inbounds i16, ptr [[IN]], i32 -5170; CHECK-NEXT: [[TMP10:%.*]] = load i32, ptr [[ARRAYIDX25]], align 2171; CHECK-NEXT: [[ARRAYIDX21:%.*]] = getelementptr inbounds i16, ptr [[B]], i32 4172; CHECK-NEXT: [[TMP12:%.*]] = load i32, ptr [[ARRAYIDX21]], align 2173; CHECK-NEXT: [[TMP13:%.*]] = call i64 @llvm.arm.smlaldx(i32 [[TMP10]], i32 [[TMP12]], i64 [[TMP8]])174; CHECK-NEXT: ret i64 [[TMP13]]175;176entry:177 %0 = load i16, ptr %in, align 2178 %conv = sext i16 %0 to i32179 %1 = load i16, ptr %b, align 2180 %conv2 = sext i16 %1 to i32181 %call = tail call i32 @bar(i32 %conv, i32 %conv2)182 %sext.0 = sext i32 %call to i64183 %arrayidx3 = getelementptr inbounds i16, ptr %in, i32 -1184 %2 = load i16, ptr %arrayidx3, align 2185 %conv4 = sext i16 %2 to i32186 %arrayidx5 = getelementptr inbounds i16, ptr %b, i32 1187 %3 = load i16, ptr %arrayidx5, align 2188 %conv6 = sext i16 %3 to i32189 %mul = mul nsw i32 %conv6, %conv4190 %sext.1 = sext i32 %mul to i64191 %add = add i64 %sext.0, %sext.1192 %arrayidx7 = getelementptr inbounds i16, ptr %in, i32 -2193 %4 = load i16, ptr %arrayidx7, align 2194 %conv8 = sext i16 %4 to i32195 %arrayidx9 = getelementptr inbounds i16, ptr %b, i32 2196 %5 = load i16, ptr %arrayidx9, align 2197 %conv10 = sext i16 %5 to i32198 %mul11 = mul nsw i32 %conv10, %conv8199 %sext.2 = sext i32 %mul11 to i64200 %add12 = add i64 %add, %sext.2201 %arrayidx13 = getelementptr inbounds i16, ptr %in, i32 -3202 %6 = load i16, ptr %arrayidx13, align 2203 %conv14 = sext i16 %6 to i32204 %arrayidx15 = getelementptr inbounds i16, ptr %b, i32 3205 %7 = load i16, ptr %arrayidx15, align 2206 %conv16 = sext i16 %7 to i32207 %mul17 = mul nsw i32 %conv16, %conv14208 %sext.3 = sext i32 %mul17 to i64209 %add18 = add i64 %add12, %sext.3210 %arrayidx19 = getelementptr inbounds i16, ptr %in, i32 -4211 %8 = load i16, ptr %arrayidx19, align 2212 %conv20 = sext i16 %8 to i32213 %arrayidx21 = getelementptr inbounds i16, ptr %b, i32 4214 %9 = load i16, ptr %arrayidx21, align 2215 %conv22 = sext i16 %9 to i32216 %mul23 = mul nsw i32 %conv22, %conv20217 %sext.4 = sext i32 %mul23 to i64218 %add24 = add i64 %add18, %sext.4219 %arrayidx25 = getelementptr inbounds i16, ptr %in, i32 -5220 %10 = load i16, ptr %arrayidx25, align 2221 %conv26 = sext i16 %10 to i32222 %arrayidx27 = getelementptr inbounds i16, ptr %b, i32 5223 %11 = load i16, ptr %arrayidx27, align 2224 %conv28 = sext i16 %11 to i32225 %mul29 = mul nsw i32 %conv28, %conv26226 %sext.5 = sext i32 %mul29 to i64227 %add30 = add i64 %add24, %sext.5228 ret i64 %add30229}230 231define i64 @with_64bit_add_acc(ptr nocapture readonly %px.10756.unr, ptr nocapture readonly %py.8757.unr, i32 %acc) {232; CHECK-LABEL: @with_64bit_add_acc(233; CHECK-NEXT: entry:234; CHECK-NEXT: [[SUM_3758_UNR:%.*]] = sext i32 [[ACC:%.*]] to i64235; CHECK-NEXT: br label [[BB_1:%.*]]236; CHECK: bb.1:237; CHECK-NEXT: [[INCDEC_PTR184_EPIL:%.*]] = getelementptr inbounds i16, ptr [[PX_10756_UNR:%.*]], i32 1238; CHECK-NEXT: [[TMP216:%.*]] = load i16, ptr [[PX_10756_UNR]], align 2239; CHECK-NEXT: [[CONV185_EPIL:%.*]] = sext i16 [[TMP216]] to i32240; CHECK-NEXT: [[INCDEC_PTR186_EPIL:%.*]] = getelementptr inbounds i16, ptr [[PY_8757_UNR:%.*]], i32 -1241; CHECK-NEXT: [[TMP217:%.*]] = load i16, ptr [[PY_8757_UNR]], align 2242; CHECK-NEXT: [[CONV187_EPIL:%.*]] = sext i16 [[TMP217]] to i32243; CHECK-NEXT: [[MUL_EPIL:%.*]] = mul nsw i32 [[CONV187_EPIL]], [[CONV185_EPIL]]244; CHECK-NEXT: [[CONV188_EPIL:%.*]] = sext i32 [[MUL_EPIL]] to i64245; CHECK-NEXT: [[ADD189_EPIL:%.*]] = add nsw i64 [[SUM_3758_UNR]], [[CONV188_EPIL]]246; CHECK-NEXT: [[INCDEC_PTR190_EPIL:%.*]] = getelementptr inbounds i16, ptr [[PX_10756_UNR]], i32 2247; CHECK-NEXT: [[TMP218:%.*]] = load i16, ptr [[INCDEC_PTR184_EPIL]], align 2248; CHECK-NEXT: [[CONV191_EPIL:%.*]] = sext i16 [[TMP218]] to i32249; CHECK-NEXT: [[TMP219:%.*]] = load i16, ptr [[INCDEC_PTR186_EPIL]], align 2250; CHECK-NEXT: [[CONV193_EPIL:%.*]] = sext i16 [[TMP219]] to i32251; CHECK-NEXT: [[MUL194_EPIL:%.*]] = mul nsw i32 [[CONV193_EPIL]], [[CONV191_EPIL]]252; CHECK-NEXT: [[TMP0:%.*]] = sext i32 [[MUL194_EPIL]] to i64253; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[TMP0]], [[ADD189_EPIL]]254; CHECK-NEXT: [[TMP3:%.*]] = load i32, ptr [[INCDEC_PTR190_EPIL]], align 2255; CHECK-NEXT: [[INCDEC_PTR199_EPIL:%.*]] = getelementptr inbounds i16, ptr [[PY_8757_UNR]], i32 -3256; CHECK-NEXT: [[TMP5:%.*]] = load i32, ptr [[INCDEC_PTR199_EPIL]], align 2257; CHECK-NEXT: [[TMP6:%.*]] = call i64 @llvm.arm.smlaldx(i32 [[TMP5]], i32 [[TMP3]], i64 [[TMP1]])258; CHECK-NEXT: ret i64 [[TMP6]]259;260entry:261 %sum.3758.unr = sext i32 %acc to i64262 br label %bb.1263 264bb.1:265 %incdec.ptr184.epil = getelementptr inbounds i16, ptr %px.10756.unr, i32 1266 %tmp216 = load i16, ptr %px.10756.unr, align 2267 %conv185.epil = sext i16 %tmp216 to i32268 %incdec.ptr186.epil = getelementptr inbounds i16, ptr %py.8757.unr, i32 -1269 %tmp217 = load i16, ptr %py.8757.unr, align 2270 %conv187.epil = sext i16 %tmp217 to i32271 %mul.epil = mul nsw i32 %conv187.epil, %conv185.epil272 %conv188.epil = sext i32 %mul.epil to i64273 %add189.epil = add nsw i64 %sum.3758.unr, %conv188.epil274 %incdec.ptr190.epil = getelementptr inbounds i16, ptr %px.10756.unr, i32 2275 %tmp218 = load i16, ptr %incdec.ptr184.epil, align 2276 %conv191.epil = sext i16 %tmp218 to i32277 %incdec.ptr192.epil = getelementptr inbounds i16, ptr %py.8757.unr, i32 -2278 %tmp219 = load i16, ptr %incdec.ptr186.epil, align 2279 %conv193.epil = sext i16 %tmp219 to i32280 %mul194.epil = mul nsw i32 %conv193.epil, %conv191.epil281 %conv195.epil = sext i32 %mul194.epil to i64282 %add196.epil = add nsw i64 %add189.epil, %conv195.epil283 %incdec.ptr197.epil = getelementptr inbounds i16, ptr %px.10756.unr, i32 3284 %tmp220 = load i16, ptr %incdec.ptr190.epil, align 2285 %conv198.epil = sext i16 %tmp220 to i32286 %incdec.ptr199.epil = getelementptr inbounds i16, ptr %py.8757.unr, i32 -3287 %tmp221 = load i16, ptr %incdec.ptr192.epil, align 2288 %conv200.epil = sext i16 %tmp221 to i32289 %mul201.epil = mul nsw i32 %conv200.epil, %conv198.epil290 %conv202.epil = sext i32 %mul201.epil to i64291 %add203.epil = add nsw i64 %add196.epil, %conv202.epil292 %tmp222 = load i16, ptr %incdec.ptr197.epil, align 2293 %conv205.epil = sext i16 %tmp222 to i32294 %tmp223 = load i16, ptr %incdec.ptr199.epil, align 2295 %conv207.epil = sext i16 %tmp223 to i32296 %mul208.epil = mul nsw i32 %conv207.epil, %conv205.epil297 %conv209.epil = sext i32 %mul208.epil to i64298 %add210.epil = add nsw i64 %add203.epil, %conv209.epil299 ret i64 %add210.epil300}301 302declare dso_local i32 @bar(i32, i32) local_unnamed_addr303 304