235 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; RUN: opt -mtriple=armv8m.maineb-none-none-eabi -mattr=+dsp < %s -arm-parallel-dsp -S | FileCheck %s --check-prefix=CHECK-UNSUPPORTED5 6define i32 @smladx(ptr nocapture readonly %pIn1, ptr nocapture readonly %pIn2, i32 %j, i32 %limit) {7 8; CHECK-LABEL: smladx9; CHECK: = phi i32 [ 0, %for.body.preheader.new ],10; CHECK: [[ACC0:%[^ ]+]] = phi i32 [ 0, %for.body.preheader.new ], [ [[ACC2:%[^ ]+]], %for.body ]11; CHECK: [[IN21:%[^ ]+]] = load i32, ptr %pIn2.1, align 212; CHECK: [[IN10:%[^ ]+]] = load i32, ptr %pIn1.0, align 213; CHECK: [[ACC1:%[^ ]+]] = call i32 @llvm.arm.smladx(i32 [[IN21]], i32 [[IN10]], i32 [[ACC0]])14 15; CHECK: [[IN23:%[^ ]+]] = load i32, ptr %pIn2.3, align 216; CHECK: [[IN12:%[^ ]+]] = load i32, ptr %pIn1.2, align 217; CHECK: [[ACC2]] = call i32 @llvm.arm.smladx(i32 [[IN23]], i32 [[IN12]], i32 [[ACC1]])18; CHECK-NOT: call i32 @llvm.arm.smlad19; CHECK-UNSUPPORTED-NOT: call i32 @llvm.arm.smlad20 21entry:22 %cmp9 = icmp eq i32 %limit, 023 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader24 25for.body.preheader:26 %0 = add i32 %limit, -127 %xtraiter = and i32 %limit, 328 %1 = icmp ult i32 %0, 329 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new30 31for.body.preheader.new:32 %unroll_iter = sub i32 %limit, %xtraiter33 br label %for.body34 35for.cond.cleanup.loopexit.unr-lcssa:36 %add.lcssa.ph = phi i32 [ undef, %for.body.preheader ], [ %add.3, %for.body ]37 %i.011.unr = phi i32 [ 0, %for.body.preheader ], [ %inc.3, %for.body ]38 %sum.010.unr = phi i32 [ 0, %for.body.preheader ], [ %add.3, %for.body ]39 %lcmp.mod = icmp eq i32 %xtraiter, 040 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil41 42for.body.epil:43 %i.011.epil = phi i32 [ %inc.epil, %for.body.epil ], [ %i.011.unr, %for.cond.cleanup.loopexit.unr-lcssa ]44 %sum.010.epil = phi i32 [ %add.epil, %for.body.epil ], [ %sum.010.unr, %for.cond.cleanup.loopexit.unr-lcssa ]45 %epil.iter = phi i32 [ %epil.iter.sub, %for.body.epil ], [ %xtraiter, %for.cond.cleanup.loopexit.unr-lcssa ]46 %sub.epil = sub i32 %j, %i.011.epil47 %arrayidx.epil = getelementptr inbounds i16, ptr %pIn2, i32 %sub.epil48 %2 = load i16, ptr %arrayidx.epil, align 249 %conv.epil = sext i16 %2 to i3250 %arrayidx1.epil = getelementptr inbounds i16, ptr %pIn1, i32 %i.011.epil51 %3 = load i16, ptr %arrayidx1.epil, align 252 %conv2.epil = sext i16 %3 to i3253 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil54 %add.epil = add nsw i32 %mul.epil, %sum.010.epil55 %inc.epil = add nuw i32 %i.011.epil, 156 %epil.iter.sub = add i32 %epil.iter, -157 %epil.iter.cmp = icmp eq i32 %epil.iter.sub, 058 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil59 60for.cond.cleanup:61 %sum.0.lcssa = phi i32 [ 0, %entry ], [ %add.lcssa.ph, %for.cond.cleanup.loopexit.unr-lcssa ], [ %add.epil, %for.body.epil ]62 ret i32 %sum.0.lcssa63 64for.body:65 %i.011 = phi i32 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ]66 %sum.010 = phi i32 [ 0, %for.body.preheader.new ], [ %add.3, %for.body ]67 %niter = phi i32 [ %unroll_iter, %for.body.preheader.new ], [ %niter.nsub.3, %for.body ]68 %pIn2Base = phi ptr [ %pIn2, %for.body.preheader.new ], [ %pIn2.4, %for.body ]69 %In2 = load i16, ptr %pIn2Base, align 270 %pIn1.0 = getelementptr inbounds i16, ptr %pIn1, i32 %i.01171 %In1 = load i16, ptr %pIn1.0, align 272 %inc = or disjoint i32 %i.011, 173 %pIn2.1 = getelementptr inbounds i16, ptr %pIn2Base, i32 -174 %In2.1 = load i16, ptr %pIn2.1, align 275 %pIn1.1 = getelementptr inbounds i16, ptr %pIn1, i32 %inc76 %In1.1 = load i16, ptr %pIn1.1, align 277 %inc.1 = or disjoint i32 %i.011, 278 %pIn2.2 = getelementptr inbounds i16, ptr %pIn2Base, i32 -279 %In2.2 = load i16, ptr %pIn2.2, align 280 %pIn1.2 = getelementptr inbounds i16, ptr %pIn1, i32 %inc.181 %In1.2 = load i16, ptr %pIn1.2, align 282 %inc.2 = or disjoint i32 %i.011, 383 %pIn2.3 = getelementptr inbounds i16, ptr %pIn2Base, i32 -384 %In2.3 = load i16, ptr %pIn2.3, align 285 %pIn1.3 = getelementptr inbounds i16, ptr %pIn1, i32 %inc.286 %In1.3 = load i16, ptr %pIn1.3, align 287 %sextIn1 = sext i16 %In1 to i3288 %sextIn1.1 = sext i16 %In1.1 to i3289 %sextIn1.2 = sext i16 %In1.2 to i3290 %sextIn1.3 = sext i16 %In1.3 to i3291 %sextIn2 = sext i16 %In2 to i3292 %sextIn2.1 = sext i16 %In2.1 to i3293 %sextIn2.2 = sext i16 %In2.2 to i3294 %sextIn2.3 = sext i16 %In2.3 to i3295 %mul = mul nsw i32 %sextIn1, %sextIn296 %mul.1 = mul nsw i32 %sextIn1.1, %sextIn2.197 %mul.2 = mul nsw i32 %sextIn1.2, %sextIn2.298 %mul.3 = mul nsw i32 %sextIn1.3, %sextIn2.399 %add = add nsw i32 %mul, %sum.010100 %add.1 = add nsw i32 %mul.1, %add101 %add.2 = add nsw i32 %mul.2, %add.1102 %add.3 = add nsw i32 %mul.3, %add.2103 %inc.3 = add i32 %i.011, 4104 %pIn2.4 = getelementptr inbounds i16, ptr %pIn2Base, i32 -4105 %niter.nsub.3 = add i32 %niter, -4106 %niter.ncmp.3 = icmp eq i32 %niter.nsub.3, 0107 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body108}109 110define i32 @smladx_swap(ptr nocapture readonly %pIn1, ptr nocapture readonly %pIn2, i32 %j, i32 %limit) {111 112; CHECK-LABEL: smladx_swap113; CHECK: for.body.preheader.new:114; CHECK: [[PIN1Base:[^ ]+]] = getelementptr i16, ptr %pIn1115; CHECK: [[PIN2Base:[^ ]+]] = getelementptr i16, ptr %pIn2116 117; CHECK: for.body:118; CHECK: [[PIN2:%[^ ]+]] = phi ptr [ [[PIN2_NEXT:%[^ ]+]], %for.body ], [ [[PIN2Base]], %for.body.preheader.new ]119; CHECK: [[PIN1:%[^ ]+]] = phi ptr [ [[PIN1_NEXT:%[^ ]+]], %for.body ], [ [[PIN1Base]], %for.body.preheader.new ]120; CHECK: [[IV:%[^ ]+]] = phi i32121; CHECK: [[ACC0:%[^ ]+]] = phi i32 [ 0, %for.body.preheader.new ], [ [[ACC2:%[^ ]+]], %for.body ]122 123; CHECK: [[IN2:%[^ ]+]] = load i32, ptr [[PIN2]], align 2124 125; CHECK: [[PIN1_2:%[^ ]+]] = getelementptr i16, ptr [[PIN1]], i32 -2126; CHECK: [[IN1_2:%[^ ]+]] = load i32, ptr [[PIN1_2]], align 2127; CHECK: [[ACC1:%[^ ]+]] = call i32 @llvm.arm.smladx(i32 [[IN2]], i32 [[IN1_2]], i32 [[ACC0]])128 129; CHECK: [[PIN2_2:%[^ ]+]] = getelementptr i16, ptr [[PIN2]], i32 -2130; CHECK: [[IN2_2:%[^ ]+]] = load i32, ptr [[PIN2_2]], align 2131 132; CHECK: [[IN1:%[^ ]+]] = load i32, ptr [[PIN1]], align 2133 134; CHECK: [[ACC2]] = call i32 @llvm.arm.smladx(i32 [[IN2_2]], i32 [[IN1]], i32 [[ACC1]])135 136; CHECK: [[PIN1_NEXT]] = getelementptr i16, ptr [[PIN1]], i32 4137; CHECK: [[PIN2_NEXT]] = getelementptr i16, ptr [[PIN2]], i32 -4138 139; CHECK-NOT: call i32 @llvm.arm.smlad140; CHECK-UNSUPPORTED-NOT: call i32 @llvm.arm.smlad141 142entry:143 %cmp9 = icmp eq i32 %limit, 0144 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader145 146for.body.preheader:147 %0 = add i32 %limit, -1148 %xtraiter = and i32 %limit, 3149 %1 = icmp ult i32 %0, 3150 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new151 152for.body.preheader.new:153 %unroll_iter = sub i32 %limit, %xtraiter154 %scevgep6 = getelementptr i16, ptr %pIn1, i32 2155 %2 = add i32 %j, -1156 %scevgep11 = getelementptr i16, ptr %pIn2, i32 %2157 br label %for.body158 159for.cond.cleanup.loopexit.unr-lcssa:160 %add.lcssa.ph = phi i32 [ undef, %for.body.preheader ], [ %add.3, %for.body ]161 %i.011.unr = phi i32 [ 0, %for.body.preheader ], [ %inc.3, %for.body ]162 %sum.010.unr = phi i32 [ 0, %for.body.preheader ], [ %add.3, %for.body ]163 %lcmp.mod = icmp eq i32 %xtraiter, 0164 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil.preheader165 166for.body.epil.preheader:167 %scevgep = getelementptr i16, ptr %pIn1, i32 %i.011.unr168 %3 = sub i32 %j, %i.011.unr169 %scevgep2 = getelementptr i16, ptr %pIn2, i32 %3170 %4 = sub i32 0, %xtraiter171 br label %for.body.epil172 173for.body.epil:174 %lsr.iv5 = phi i32 [ %4, %for.body.epil.preheader ], [ %lsr.iv.next, %for.body.epil ]175 %lsr.iv3 = phi ptr [ %scevgep2, %for.body.epil.preheader ], [ %scevgep4, %for.body.epil ]176 %lsr.iv = phi ptr [ %scevgep, %for.body.epil.preheader ], [ %scevgep1, %for.body.epil ]177 %sum.010.epil = phi i32 [ %add.epil, %for.body.epil ], [ %sum.010.unr, %for.body.epil.preheader ]178 %5 = load i16, ptr %lsr.iv3, align 2179 %conv.epil = sext i16 %5 to i32180 %6 = load i16, ptr %lsr.iv, align 2181 %conv2.epil = sext i16 %6 to i32182 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil183 %add.epil = add nsw i32 %mul.epil, %sum.010.epil184 %scevgep1 = getelementptr i16, ptr %lsr.iv, i32 1185 %scevgep4 = getelementptr i16, ptr %lsr.iv3, i32 -1186 %lsr.iv.next = add nsw i32 %lsr.iv5, 1187 %epil.iter.cmp = icmp eq i32 %lsr.iv.next, 0188 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil189 190for.cond.cleanup:191 %sum.0.lcssa = phi i32 [ 0, %entry ], [ %add.lcssa.ph, %for.cond.cleanup.loopexit.unr-lcssa ], [ %add.epil, %for.body.epil ]192 ret i32 %sum.0.lcssa193 194for.body:195 %pin2 = phi ptr [ %pin2_sub4, %for.body ], [ %scevgep11, %for.body.preheader.new ]196 %pin1 = phi ptr [ %pin1_add4, %for.body ], [ %scevgep6, %for.body.preheader.new ]197 %i.011 = phi i32 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ]198 %sum.010 = phi i32 [ 0, %for.body.preheader.new ], [ %add.3, %for.body ]199 %pin2_add1 = getelementptr i16, ptr %pin2, i32 1200 %In2 = load i16, ptr %pin2_add1, align 2201 %pin1_sub2 = getelementptr i16, ptr %pin1, i32 -2202 %In1 = load i16, ptr %pin1_sub2, align 2203 %In2.1 = load i16, ptr %pin2, align 2204 %pin1_sub1 = getelementptr i16, ptr %pin1, i32 -1205 %In1.1 = load i16, ptr %pin1_sub1, align 2206 %pin2_sub1 = getelementptr i16, ptr %pin2, i32 -1207 %In2.2 = load i16, ptr %pin2_sub1, align 2208 %In1.2 = load i16, ptr %pin1, align 2209 %pin2_sub2 = getelementptr i16, ptr %pin2, i32 -2210 %In2.3 = load i16, ptr %pin2_sub2, align 2211 %pin1_add1 = getelementptr i16, ptr %pin1, i32 1212 %In1.3 = load i16, ptr %pin1_add1, align 2213 %sextIn2 = sext i16 %In2 to i32214 %sextIn1 = sext i16 %In1 to i32215 %sextIn2.1 = sext i16 %In2.1 to i32216 %sextIn1.1 = sext i16 %In1.1 to i32217 %sextIn2.2 = sext i16 %In2.2 to i32218 %sextIn1.2 = sext i16 %In1.2 to i32219 %sextIn2.3 = sext i16 %In2.3 to i32220 %sextIn1.3 = sext i16 %In1.3 to i32221 %mul = mul nsw i32 %sextIn2, %sextIn1222 %add = add nsw i32 %mul, %sum.010223 %mul.1 = mul nsw i32 %sextIn2.1, %sextIn1.1224 %add.1 = add nsw i32 %mul.1, %add225 %mul.2 = mul nsw i32 %sextIn2.2, %sextIn1.2226 %add.2 = add nsw i32 %mul.2, %add.1227 %mul.3 = mul nsw i32 %sextIn2.3, %sextIn1.3228 %add.3 = add nsw i32 %mul.3, %add.2229 %inc.3 = add i32 %i.011, 4230 %pin1_add4 = getelementptr i16, ptr %pin1, i32 4231 %pin2_sub4 = getelementptr i16, ptr %pin2, i32 -4232 %niter.ncmp.3 = icmp eq i32 %unroll_iter, %inc.3233 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body234}235