1; RUN: opt -mtriple=thumbv8m.main -mcpu=cortex-m33 -arm-parallel-dsp %s -S -o - | FileCheck %s 2; RUN: opt -mtriple=arm-arm-eabi -mcpu=cortex-m0 < %s -arm-parallel-dsp -S | FileCheck %s --check-prefix=CHECK-UNSUPPORTED 3; RUN: opt -mtriple=arm-arm-eabi -mcpu=cortex-m33 -mattr=-dsp < %s -arm-parallel-dsp -S | FileCheck %s --check-prefix=CHECK-UNSUPPORTED 4; RUN: opt -mtriple=armeb-arm-eabi -mcpu=cortex-m33 < %s -arm-parallel-dsp -S | FileCheck %s --check-prefix=CHECK-UNSUPPORTED 5 6define i32 @smladx(i16* nocapture readonly %pIn1, i16* nocapture readonly %pIn2, i32 %j, i32 %limit) { 7 8; CHECK-LABEL: smladx 9; CHECK: = phi i32 [ 0, %for.body.preheader.new ], 10; CHECK: [[ACC0:%[^ ]+]] = phi i32 [ 0, %for.body.preheader.new ], [ [[ACC2:%[^ ]+]], %for.body ] 11; CHECK: [[PIN21:%[^ ]+]] = bitcast i16* %pIn2.1 to i32* 12; CHECK: [[IN21:%[^ ]+]] = load i32, i32* [[PIN21]], align 2 13; CHECK: [[PIN10:%[^ ]+]] = bitcast i16* %pIn1.0 to i32* 14; CHECK: [[IN10:%[^ ]+]] = load i32, i32* [[PIN10]], align 2 15; CHECK: [[ACC1:%[^ ]+]] = call i32 @llvm.arm.smladx(i32 [[IN21]], i32 [[IN10]], i32 [[ACC0]]) 16 17; CHECK: [[PIN23:%[^ ]+]] = bitcast i16* %pIn2.3 to i32* 18; CHECK: [[IN23:%[^ ]+]] = load i32, i32* [[PIN23]], align 2 19; CHECK: [[PIN12:%[^ ]+]] = bitcast i16* %pIn1.2 to i32* 20; CHECK: [[IN12:%[^ ]+]] = load i32, i32* [[PIN12]], align 2 21; CHECK: [[ACC2]] = call i32 @llvm.arm.smladx(i32 [[IN23]], i32 [[IN12]], i32 [[ACC1]]) 22; CHECK-NOT: call i32 @llvm.arm.smlad 23; CHECK-UNSUPPORTED-NOT: call i32 @llvm.arm.smlad 24 25entry: 26 %cmp9 = icmp eq i32 %limit, 0 27 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader 28 29for.body.preheader: 30 %0 = add i32 %limit, -1 31 %xtraiter = and i32 %limit, 3 32 %1 = icmp ult i32 %0, 3 33 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new 34 35for.body.preheader.new: 36 %unroll_iter = sub i32 %limit, %xtraiter 37 br label %for.body 38 39for.cond.cleanup.loopexit.unr-lcssa: 40 %add.lcssa.ph = phi i32 [ undef, %for.body.preheader ], [ %add.3, %for.body ] 41 %i.011.unr = phi i32 [ 0, %for.body.preheader ], [ %inc.3, %for.body ] 42 %sum.010.unr = phi i32 [ 0, %for.body.preheader ], [ %add.3, %for.body ] 43 %lcmp.mod = icmp eq i32 %xtraiter, 0 44 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil 45 46for.body.epil: 47 %i.011.epil = phi i32 [ %inc.epil, %for.body.epil ], [ %i.011.unr, %for.cond.cleanup.loopexit.unr-lcssa ] 48 %sum.010.epil = phi i32 [ %add.epil, %for.body.epil ], [ %sum.010.unr, %for.cond.cleanup.loopexit.unr-lcssa ] 49 %epil.iter = phi i32 [ %epil.iter.sub, %for.body.epil ], [ %xtraiter, %for.cond.cleanup.loopexit.unr-lcssa ] 50 %sub.epil = sub i32 %j, %i.011.epil 51 %arrayidx.epil = getelementptr inbounds i16, i16* %pIn2, i32 %sub.epil 52 %2 = load i16, i16* %arrayidx.epil, align 2 53 %conv.epil = sext i16 %2 to i32 54 %arrayidx1.epil = getelementptr inbounds i16, i16* %pIn1, i32 %i.011.epil 55 %3 = load i16, i16* %arrayidx1.epil, align 2 56 %conv2.epil = sext i16 %3 to i32 57 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil 58 %add.epil = add nsw i32 %mul.epil, %sum.010.epil 59 %inc.epil = add nuw i32 %i.011.epil, 1 60 %epil.iter.sub = add i32 %epil.iter, -1 61 %epil.iter.cmp = icmp eq i32 %epil.iter.sub, 0 62 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil 63 64for.cond.cleanup: 65 %sum.0.lcssa = phi i32 [ 0, %entry ], [ %add.lcssa.ph, %for.cond.cleanup.loopexit.unr-lcssa ], [ %add.epil, %for.body.epil ] 66 ret i32 %sum.0.lcssa 67 68for.body: 69 %i.011 = phi i32 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ] 70 %sum.010 = phi i32 [ 0, %for.body.preheader.new ], [ %add.3, %for.body ] 71 %niter = phi i32 [ %unroll_iter, %for.body.preheader.new ], [ %niter.nsub.3, %for.body ] 72 %pIn2Base = phi i16* [ %pIn2, %for.body.preheader.new ], [ %pIn2.4, %for.body ] 73 %pIn2.0 = getelementptr inbounds i16, i16* %pIn2Base, i32 0 74 %In2 = load i16, i16* %pIn2.0, align 2 75 %pIn1.0 = getelementptr inbounds i16, i16* %pIn1, i32 %i.011 76 %In1 = load i16, i16* %pIn1.0, align 2 77 %inc = or i32 %i.011, 1 78 %pIn2.1 = getelementptr inbounds i16, i16* %pIn2Base, i32 -1 79 %In2.1 = load i16, i16* %pIn2.1, align 2 80 %pIn1.1 = getelementptr inbounds i16, i16* %pIn1, i32 %inc 81 %In1.1 = load i16, i16* %pIn1.1, align 2 82 %inc.1 = or i32 %i.011, 2 83 %pIn2.2 = getelementptr inbounds i16, i16* %pIn2Base, i32 -2 84 %In2.2 = load i16, i16* %pIn2.2, align 2 85 %pIn1.2 = getelementptr inbounds i16, i16* %pIn1, i32 %inc.1 86 %In1.2 = load i16, i16* %pIn1.2, align 2 87 %inc.2 = or i32 %i.011, 3 88 %pIn2.3 = getelementptr inbounds i16, i16* %pIn2Base, i32 -3 89 %In2.3 = load i16, i16* %pIn2.3, align 2 90 %pIn1.3 = getelementptr inbounds i16, i16* %pIn1, i32 %inc.2 91 %In1.3 = load i16, i16* %pIn1.3, align 2 92 %sextIn1 = sext i16 %In1 to i32 93 %sextIn1.1 = sext i16 %In1.1 to i32 94 %sextIn1.2 = sext i16 %In1.2 to i32 95 %sextIn1.3 = sext i16 %In1.3 to i32 96 %sextIn2 = sext i16 %In2 to i32 97 %sextIn2.1 = sext i16 %In2.1 to i32 98 %sextIn2.2 = sext i16 %In2.2 to i32 99 %sextIn2.3 = sext i16 %In2.3 to i32 100 %mul = mul nsw i32 %sextIn1, %sextIn2 101 %mul.1 = mul nsw i32 %sextIn1.1, %sextIn2.1 102 %mul.2 = mul nsw i32 %sextIn1.2, %sextIn2.2 103 %mul.3 = mul nsw i32 %sextIn1.3, %sextIn2.3 104 %add = add nsw i32 %mul, %sum.010 105 %add.1 = add nsw i32 %mul.1, %add 106 %add.2 = add nsw i32 %mul.2, %add.1 107 %add.3 = add nsw i32 %mul.3, %add.2 108 %inc.3 = add i32 %i.011, 4 109 %pIn2.4 = getelementptr inbounds i16, i16* %pIn2Base, i32 -4 110 %niter.nsub.3 = add i32 %niter, -4 111 %niter.ncmp.3 = icmp eq i32 %niter.nsub.3, 0 112 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body 113} 114 115define i32 @smladx_swap(i16* nocapture readonly %pIn1, i16* nocapture readonly %pIn2, i32 %j, i32 %limit) { 116 117; CHECK-LABEL: smladx_swap 118; CHECK: for.body.preheader.new: 119; CHECK: [[PIN1Base:[^ ]+]] = getelementptr i16, i16* %pIn1 120; CHECK: [[PIN2Base:[^ ]+]] = getelementptr i16, i16* %pIn2 121 122; CHECK: for.body: 123; CHECK: [[PIN2:%[^ ]+]] = phi i16* [ [[PIN2_NEXT:%[^ ]+]], %for.body ], [ [[PIN2Base]], %for.body.preheader.new ] 124; CHECK: [[PIN1:%[^ ]+]] = phi i16* [ [[PIN1_NEXT:%[^ ]+]], %for.body ], [ [[PIN1Base]], %for.body.preheader.new ] 125; CHECK: [[IV:%[^ ]+]] = phi i32 126; CHECK: [[ACC0:%[^ ]+]] = phi i32 [ 0, %for.body.preheader.new ], [ [[ACC2:%[^ ]+]], %for.body ] 127 128; CHECK: [[PIN2_CAST:%[^ ]+]] = bitcast i16* [[PIN2]] to i32* 129; CHECK: [[IN2:%[^ ]+]] = load i32, i32* [[PIN2_CAST]], align 2 130 131; CHECK: [[PIN1_2:%[^ ]+]] = getelementptr i16, i16* [[PIN1]], i32 -2 132; CHECK: [[PIN1_2_CAST:%[^ ]+]] = bitcast i16* [[PIN1_2]] to i32* 133; CHECK: [[IN1_2:%[^ ]+]] = load i32, i32* [[PIN1_2_CAST]], align 2 134; CHECK: [[ACC1:%[^ ]+]] = call i32 @llvm.arm.smladx(i32 [[IN2]], i32 [[IN1_2]], i32 [[ACC0]]) 135 136; CHECK: [[PIN2_2:%[^ ]+]] = getelementptr i16, i16* [[PIN2]], i32 -2 137; CHECK: [[PIN2_2_CAST:%[^ ]+]] = bitcast i16* [[PIN2_2]] to i32* 138; CHECK: [[IN2_2:%[^ ]+]] = load i32, i32* [[PIN2_2_CAST]], align 2 139 140; CHECK: [[PIN1_CAST:%[^ ]+]] = bitcast i16* [[PIN1]] to i32* 141; CHECK: [[IN1:%[^ ]+]] = load i32, i32* [[PIN1_CAST]], align 2 142 143; CHECK: [[ACC2]] = call i32 @llvm.arm.smladx(i32 [[IN2_2]], i32 [[IN1]], i32 [[ACC1]]) 144 145; CHECK: [[PIN1_NEXT]] = getelementptr i16, i16* [[PIN1]], i32 4 146; CHECK: [[PIN2_NEXT]] = getelementptr i16, i16* [[PIN2]], i32 -4 147 148; CHECK-NOT: call i32 @llvm.arm.smlad 149; CHECK-UNSUPPORTED-NOT: call i32 @llvm.arm.smlad 150 151entry: 152 %cmp9 = icmp eq i32 %limit, 0 153 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader 154 155for.body.preheader: 156 %0 = add i32 %limit, -1 157 %xtraiter = and i32 %limit, 3 158 %1 = icmp ult i32 %0, 3 159 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new 160 161for.body.preheader.new: 162 %unroll_iter = sub i32 %limit, %xtraiter 163 %scevgep6 = getelementptr i16, i16* %pIn1, i32 2 164 %2 = add i32 %j, -1 165 %scevgep11 = getelementptr i16, i16* %pIn2, i32 %2 166 br label %for.body 167 168for.cond.cleanup.loopexit.unr-lcssa: 169 %add.lcssa.ph = phi i32 [ undef, %for.body.preheader ], [ %add.3, %for.body ] 170 %i.011.unr = phi i32 [ 0, %for.body.preheader ], [ %inc.3, %for.body ] 171 %sum.010.unr = phi i32 [ 0, %for.body.preheader ], [ %add.3, %for.body ] 172 %lcmp.mod = icmp eq i32 %xtraiter, 0 173 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil.preheader 174 175for.body.epil.preheader: 176 %scevgep = getelementptr i16, i16* %pIn1, i32 %i.011.unr 177 %3 = sub i32 %j, %i.011.unr 178 %scevgep2 = getelementptr i16, i16* %pIn2, i32 %3 179 %4 = sub i32 0, %xtraiter 180 br label %for.body.epil 181 182for.body.epil: 183 %lsr.iv5 = phi i32 [ %4, %for.body.epil.preheader ], [ %lsr.iv.next, %for.body.epil ] 184 %lsr.iv3 = phi i16* [ %scevgep2, %for.body.epil.preheader ], [ %scevgep4, %for.body.epil ] 185 %lsr.iv = phi i16* [ %scevgep, %for.body.epil.preheader ], [ %scevgep1, %for.body.epil ] 186 %sum.010.epil = phi i32 [ %add.epil, %for.body.epil ], [ %sum.010.unr, %for.body.epil.preheader ] 187 %5 = load i16, i16* %lsr.iv3, align 2 188 %conv.epil = sext i16 %5 to i32 189 %6 = load i16, i16* %lsr.iv, align 2 190 %conv2.epil = sext i16 %6 to i32 191 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil 192 %add.epil = add nsw i32 %mul.epil, %sum.010.epil 193 %scevgep1 = getelementptr i16, i16* %lsr.iv, i32 1 194 %scevgep4 = getelementptr i16, i16* %lsr.iv3, i32 -1 195 %lsr.iv.next = add nsw i32 %lsr.iv5, 1 196 %epil.iter.cmp = icmp eq i32 %lsr.iv.next, 0 197 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil 198 199for.cond.cleanup: 200 %sum.0.lcssa = phi i32 [ 0, %entry ], [ %add.lcssa.ph, %for.cond.cleanup.loopexit.unr-lcssa ], [ %add.epil, %for.body.epil ] 201 ret i32 %sum.0.lcssa 202 203for.body: 204 %pin2 = phi i16* [ %pin2_sub4, %for.body ], [ %scevgep11, %for.body.preheader.new ] 205 %pin1 = phi i16* [ %pin1_add4, %for.body ], [ %scevgep6, %for.body.preheader.new ] 206 %i.011 = phi i32 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ] 207 %sum.010 = phi i32 [ 0, %for.body.preheader.new ], [ %add.3, %for.body ] 208 %pin2_add1 = getelementptr i16, i16* %pin2, i32 1 209 %In2 = load i16, i16* %pin2_add1, align 2 210 %pin1_sub2 = getelementptr i16, i16* %pin1, i32 -2 211 %In1 = load i16, i16* %pin1_sub2, align 2 212 %In2.1 = load i16, i16* %pin2, align 2 213 %pin1_sub1 = getelementptr i16, i16* %pin1, i32 -1 214 %In1.1 = load i16, i16* %pin1_sub1, align 2 215 %pin2_sub1 = getelementptr i16, i16* %pin2, i32 -1 216 %In2.2 = load i16, i16* %pin2_sub1, align 2 217 %In1.2 = load i16, i16* %pin1, align 2 218 %pin2_sub2 = getelementptr i16, i16* %pin2, i32 -2 219 %In2.3 = load i16, i16* %pin2_sub2, align 2 220 %pin1_add1 = getelementptr i16, i16* %pin1, i32 1 221 %In1.3 = load i16, i16* %pin1_add1, align 2 222 %sextIn2 = sext i16 %In2 to i32 223 %sextIn1 = sext i16 %In1 to i32 224 %sextIn2.1 = sext i16 %In2.1 to i32 225 %sextIn1.1 = sext i16 %In1.1 to i32 226 %sextIn2.2 = sext i16 %In2.2 to i32 227 %sextIn1.2 = sext i16 %In1.2 to i32 228 %sextIn2.3 = sext i16 %In2.3 to i32 229 %sextIn1.3 = sext i16 %In1.3 to i32 230 %mul = mul nsw i32 %sextIn2, %sextIn1 231 %add = add nsw i32 %mul, %sum.010 232 %mul.1 = mul nsw i32 %sextIn2.1, %sextIn1.1 233 %add.1 = add nsw i32 %mul.1, %add 234 %mul.2 = mul nsw i32 %sextIn2.2, %sextIn1.2 235 %add.2 = add nsw i32 %mul.2, %add.1 236 %mul.3 = mul nsw i32 %sextIn2.3, %sextIn1.3 237 %add.3 = add nsw i32 %mul.3, %add.2 238 %inc.3 = add i32 %i.011, 4 239 %pin1_add4 = getelementptr i16, i16* %pin1, i32 4 240 %pin2_sub4 = getelementptr i16, i16* %pin2, i32 -4 241 %niter.ncmp.3 = icmp eq i32 %unroll_iter, %inc.3 242 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body 243} 244