242 lines · plain
1; RUN: opt -mtriple=armv8m.main-none-none-eabi -mattr=+dsp -arm-parallel-dsp %s -S -o - | FileCheck %s2; RUN: opt -mtriple=armv6m-none-none-eabi < %s -arm-parallel-dsp -S | FileCheck %s --check-prefix=CHECK-UNSUPPORTED3; RUN: opt -mtriple=armv8m.main-none-none-eabi -mattr=-dsp < %s -arm-parallel-dsp -S | FileCheck %s --check-prefix=CHECK-UNSUPPORTED4 5define i64 @smlaldx(ptr nocapture readonly %pIn1, ptr nocapture readonly %pIn2, i32 %j, i32 %limit) {6 7; CHECK-LABEL: smlaldx8; CHECK: = phi i32 [ 0, %for.body.preheader.new ],9; CHECK: [[ACC0:%[^ ]+]] = phi i64 [ 0, %for.body.preheader.new ], [ [[ACC2:%[^ ]+]], %for.body ]10; CHECK: [[IN21:%[^ ]+]] = load i32, ptr %pIn2.1, align 211; CHECK: [[IN10:%[^ ]+]] = load i32, ptr %pIn1.0, align 212; CHECK: [[ACC1:%[^ ]+]] = call i64 @llvm.arm.smlaldx(i32 [[IN21]], i32 [[IN10]], i64 [[ACC0]])13; CHECK: [[IN23:%[^ ]+]] = load i32, ptr %pIn2.3, align 214; CHECK: [[IN12:%[^ ]+]] = load i32, ptr %pIn1.2, align 215; CHECK: [[ACC2]] = call i64 @llvm.arm.smlaldx(i32 [[IN23]], i32 [[IN12]], i64 [[ACC1]])16; CHECK-NOT: call i64 @llvm.arm.smlad17; CHECK-UNSUPPORTED-NOT: call i64 @llvm.arm.smlad18 19entry:20 %cmp9 = icmp eq i32 %limit, 021 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader22 23for.body.preheader:24 %0 = add i32 %limit, -125 %xtraiter = and i32 %limit, 326 %1 = icmp ult i32 %0, 327 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new28 29for.body.preheader.new:30 %unroll_iter = sub i32 %limit, %xtraiter31 br label %for.body32 33for.cond.cleanup.loopexit.unr-lcssa:34 %add.lcssa.ph = phi i64 [ undef, %for.body.preheader ], [ %add.3, %for.body ]35 %i.011.unr = phi i32 [ 0, %for.body.preheader ], [ %inc.3, %for.body ]36 %sum.010.unr = phi i64 [ 0, %for.body.preheader ], [ %add.3, %for.body ]37 %lcmp.mod = icmp eq i32 %xtraiter, 038 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil39 40for.body.epil:41 %i.011.epil = phi i32 [ %inc.epil, %for.body.epil ], [ %i.011.unr, %for.cond.cleanup.loopexit.unr-lcssa ]42 %sum.010.epil = phi i64 [ %add.epil, %for.body.epil ], [ %sum.010.unr, %for.cond.cleanup.loopexit.unr-lcssa ]43 %epil.iter = phi i32 [ %epil.iter.sub, %for.body.epil ], [ %xtraiter, %for.cond.cleanup.loopexit.unr-lcssa ]44 %sub.epil = sub i32 %j, %i.011.epil45 %arrayidx.epil = getelementptr inbounds i16, ptr %pIn2, i32 %sub.epil46 %2 = load i16, ptr %arrayidx.epil, align 247 %conv.epil = sext i16 %2 to i3248 %arrayidx1.epil = getelementptr inbounds i16, ptr %pIn1, i32 %i.011.epil49 %3 = load i16, ptr %arrayidx1.epil, align 250 %conv2.epil = sext i16 %3 to i3251 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil52 %sext.mul.epil = sext i32 %mul.epil to i6453 %add.epil = add nsw i64 %sext.mul.epil, %sum.010.epil54 %inc.epil = add nuw i32 %i.011.epil, 155 %epil.iter.sub = add i32 %epil.iter, -156 %epil.iter.cmp = icmp eq i32 %epil.iter.sub, 057 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil58 59for.cond.cleanup:60 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add.lcssa.ph, %for.cond.cleanup.loopexit.unr-lcssa ], [ %add.epil, %for.body.epil ]61 ret i64 %sum.0.lcssa62 63for.body:64 %i.011 = phi i32 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ]65 %sum.010 = phi i64 [ 0, %for.body.preheader.new ], [ %add.3, %for.body ]66 %niter = phi i32 [ %unroll_iter, %for.body.preheader.new ], [ %niter.nsub.3, %for.body ]67 %pIn2Base = phi ptr [ %pIn2, %for.body.preheader.new ], [ %pIn2.4, %for.body ]68 %In2 = load i16, ptr %pIn2Base, align 269 %pIn1.0 = getelementptr inbounds i16, ptr %pIn1, i32 %i.01170 %In1 = load i16, ptr %pIn1.0, align 271 %inc = or disjoint i32 %i.011, 172 %pIn2.1 = getelementptr inbounds i16, ptr %pIn2Base, i32 -173 %In2.1 = load i16, ptr %pIn2.1, align 274 %pIn1.1 = getelementptr inbounds i16, ptr %pIn1, i32 %inc75 %In1.1 = load i16, ptr %pIn1.1, align 276 %inc.1 = or disjoint i32 %i.011, 277 %pIn2.2 = getelementptr inbounds i16, ptr %pIn2Base, i32 -278 %In2.2 = load i16, ptr %pIn2.2, align 279 %pIn1.2 = getelementptr inbounds i16, ptr %pIn1, i32 %inc.180 %In1.2 = load i16, ptr %pIn1.2, align 281 %inc.2 = or disjoint i32 %i.011, 382 %pIn2.3 = getelementptr inbounds i16, ptr %pIn2Base, i32 -383 %In2.3 = load i16, ptr %pIn2.3, align 284 %pIn1.3 = getelementptr inbounds i16, ptr %pIn1, i32 %inc.285 %In1.3 = load i16, ptr %pIn1.3, align 286 %sextIn1 = sext i16 %In1 to i3287 %sextIn1.1 = sext i16 %In1.1 to i3288 %sextIn1.2 = sext i16 %In1.2 to i3289 %sextIn1.3 = sext i16 %In1.3 to i3290 %sextIn2 = sext i16 %In2 to i3291 %sextIn2.1 = sext i16 %In2.1 to i3292 %sextIn2.2 = sext i16 %In2.2 to i3293 %sextIn2.3 = sext i16 %In2.3 to i3294 %mul = mul nsw i32 %sextIn1, %sextIn295 %mul.1 = mul nsw i32 %sextIn1.1, %sextIn2.196 %mul.2 = mul nsw i32 %sextIn1.2, %sextIn2.297 %mul.3 = mul nsw i32 %sextIn1.3, %sextIn2.398 %sext.mul = sext i32 %mul to i6499 %sext.mul.1 = sext i32 %mul.1 to i64100 %sext.mul.2 = sext i32 %mul.2 to i64101 %sext.mul.3 = sext i32 %mul.3 to i64102 %add = add nsw i64 %sext.mul, %sum.010103 %add.1 = add nsw i64 %sext.mul.1, %add104 %add.2 = add nsw i64 %sext.mul.2, %add.1105 %add.3 = add nsw i64 %sext.mul.3, %add.2106 %inc.3 = add i32 %i.011, 4107 %pIn2.4 = getelementptr inbounds i16, ptr %pIn2Base, i32 -4108 %niter.nsub.3 = add i32 %niter, -4109 %niter.ncmp.3 = icmp eq i32 %niter.nsub.3, 0110 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body111}112 113define i64 @smlaldx_swap(ptr nocapture readonly %pIn1, ptr nocapture readonly %pIn2, i32 %j, i32 %limit) {114 115entry:116 %cmp9 = icmp eq i32 %limit, 0117 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader118 119for.body.preheader:120 %0 = add i32 %limit, -1121 %xtraiter = and i32 %limit, 3122 %1 = icmp ult i32 %0, 3123 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new124 125for.body.preheader.new:126 %unroll_iter = sub i32 %limit, %xtraiter127 %scevgep6 = getelementptr i16, ptr %pIn1, i32 2128 %2 = add i32 %j, -1129 %scevgep11 = getelementptr i16, ptr %pIn2, i32 %2130 br label %for.body131 132for.cond.cleanup.loopexit.unr-lcssa:133 %add.lcssa.ph = phi i64 [ undef, %for.body.preheader ], [ %add.3, %for.body ]134 %i.011.unr = phi i32 [ 0, %for.body.preheader ], [ %inc.3, %for.body ]135 %sum.010.unr = phi i64 [ 0, %for.body.preheader ], [ %add.3, %for.body ]136 %lcmp.mod = icmp eq i32 %xtraiter, 0137 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil.preheader138 139for.body.epil.preheader:140 %scevgep = getelementptr i16, ptr %pIn1, i32 %i.011.unr141 %3 = sub i32 %j, %i.011.unr142 %scevgep2 = getelementptr i16, ptr %pIn2, i32 %3143 %4 = sub i32 0, %xtraiter144 br label %for.body.epil145 146for.body.epil:147 %lsr.iv5 = phi i32 [ %4, %for.body.epil.preheader ], [ %lsr.iv.next, %for.body.epil ]148 %lsr.iv3 = phi ptr [ %scevgep2, %for.body.epil.preheader ], [ %scevgep4, %for.body.epil ]149 %lsr.iv = phi ptr [ %scevgep, %for.body.epil.preheader ], [ %scevgep1, %for.body.epil ]150 %sum.010.epil = phi i64 [ %add.epil, %for.body.epil ], [ %sum.010.unr, %for.body.epil.preheader ]151 %5 = load i16, ptr %lsr.iv3, align 2152 %conv.epil = sext i16 %5 to i32153 %6 = load i16, ptr %lsr.iv, align 2154 %conv2.epil = sext i16 %6 to i32155 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil156 %sext.mul.epil = sext i32 %mul.epil to i64157 %add.epil = add nsw i64 %sext.mul.epil, %sum.010.epil158 %scevgep1 = getelementptr i16, ptr %lsr.iv, i32 1159 %scevgep4 = getelementptr i16, ptr %lsr.iv3, i32 -1160 %lsr.iv.next = add nsw i32 %lsr.iv5, 1161 %epil.iter.cmp = icmp eq i32 %lsr.iv.next, 0162 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil163 164for.cond.cleanup:165 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add.lcssa.ph, %for.cond.cleanup.loopexit.unr-lcssa ], [ %add.epil, %for.body.epil ]166 ret i64 %sum.0.lcssa167 168; CHECK-LABEL: smlaldx_swap169; CHECK: for.body.preheader.new:170; CHECK: [[PIN1Base:[^ ]+]] = getelementptr i16, ptr %pIn1171; CHECK: [[PIN2Base:[^ ]+]] = getelementptr i16, ptr %pIn2172 173; CHECK: for.body:174; CHECK: [[PIN2:%[^ ]+]] = phi ptr [ [[PIN2_NEXT:%[^ ]+]], %for.body ], [ [[PIN2Base]], %for.body.preheader.new ]175; CHECK: [[PIN1:%[^ ]+]] = phi ptr [ [[PIN1_NEXT:%[^ ]+]], %for.body ], [ [[PIN1Base]], %for.body.preheader.new ]176; CHECK: [[IV:%[^ ]+]] = phi i32177; CHECK: [[ACC0:%[^ ]+]] = phi i64 [ 0, %for.body.preheader.new ], [ [[ACC2:%[^ ]+]], %for.body ]178 179; CHECK: [[IN2:%[^ ]+]] = load i32, ptr [[PIN2]], align 2180 181; CHECK: [[PIN1_2:%[^ ]+]] = getelementptr i16, ptr [[PIN1]], i32 -2182; CHECK: [[IN1_2:%[^ ]+]] = load i32, ptr [[PIN1_2]], align 2183; CHECK: [[ACC1:%[^ ]+]] = call i64 @llvm.arm.smlaldx(i32 [[IN2]], i32 [[IN1_2]], i64 [[ACC0]])184 185; CHECK: [[PIN2_2:%[^ ]+]] = getelementptr i16, ptr [[PIN2]], i32 -2186; CHECK: [[IN2_2:%[^ ]+]] = load i32, ptr [[PIN2_2]], align 2187 188; CHECK: [[IN1:%[^ ]+]] = load i32, ptr [[PIN1]], align 2189; CHECK: [[ACC2]] = call i64 @llvm.arm.smlaldx(i32 [[IN2_2]], i32 [[IN1]], i64 [[ACC1]])190 191; CHECK: [[PIN1_NEXT]] = getelementptr i16, ptr [[PIN1]], i32 4192; CHECK: [[PIN2_NEXT]] = getelementptr i16, ptr [[PIN2]], i32 -4193 194; CHECK-NOT: call i64 @llvm.arm.smlad195; CHECK-UNSUPPORTED-NOT: call i64 @llvm.arm.smlad196 197for.body:198 %pin2 = phi ptr [ %pin2.sub4, %for.body ], [ %scevgep11, %for.body.preheader.new ]199 %pin1 = phi ptr [ %pin1.add4, %for.body ], [ %scevgep6, %for.body.preheader.new ]200 %i.011 = phi i32 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ]201 %sum.010 = phi i64 [ 0, %for.body.preheader.new ], [ %add.3, %for.body ]202 %pin2.add1 = getelementptr i16, ptr %pin2, i32 1203 %In2 = load i16, ptr %pin2.add1, align 2204 %pin1.sub2 = getelementptr i16, ptr %pin1, i32 -2205 %In1 = load i16, ptr %pin1.sub2, align 2206 %In2.1 = load i16, ptr %pin2, align 2207 %pin1.sub1 = getelementptr i16, ptr %pin1, i32 -1208 %In1.1 = load i16, ptr %pin1.sub1, align 2209 %pin2.sub1 = getelementptr i16, ptr %pin2, i32 -1210 %In2.2 = load i16, ptr %pin2.sub1, align 2211 %In1.2 = load i16, ptr %pin1, align 2212 %pin2.sub2 = getelementptr i16, ptr %pin2, i32 -2213 %In2.3 = load i16, ptr %pin2.sub2, align 2214 %pin1.add1 = getelementptr i16, ptr %pin1, i32 1215 %In1.3 = load i16, ptr %pin1.add1, align 2216 %sextIn2 = sext i16 %In2 to i32217 %sextIn1 = sext i16 %In1 to i32218 %sextIn2.1 = sext i16 %In2.1 to i32219 %sextIn1.1 = sext i16 %In1.1 to i32220 %sextIn2.2 = sext i16 %In2.2 to i32221 %sextIn1.2 = sext i16 %In1.2 to i32222 %sextIn2.3 = sext i16 %In2.3 to i32223 %sextIn1.3 = sext i16 %In1.3 to i32224 %mul = mul nsw i32 %sextIn2, %sextIn1225 %sext.mul = sext i32 %mul to i64226 %add = add nsw i64 %sext.mul, %sum.010227 %mul.1 = mul nsw i32 %sextIn2.1, %sextIn1.1228 %sext.mul.1 = sext i32 %mul.1 to i64229 %add.1 = add nsw i64 %sext.mul.1, %add230 %mul.2 = mul nsw i32 %sextIn2.2, %sextIn1.2231 %sext.mul.2 = sext i32 %mul.2 to i64232 %add.2 = add nsw i64 %sext.mul.2, %add.1233 %mul.3 = mul nsw i32 %sextIn2.3, %sextIn1.3234 %sext.mul.3 = sext i32 %mul.3 to i64235 %add.3 = add nsw i64 %sext.mul.3, %add.2236 %inc.3 = add i32 %i.011, 4237 %pin1.add4 = getelementptr i16, ptr %pin1, i32 4238 %pin2.sub4 = getelementptr i16, ptr %pin2, i32 -4239 %niter.ncmp.3 = icmp eq i32 %unroll_iter, %inc.3240 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body241}242