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 5define i64 @smlaldx(i16* nocapture readonly %pIn1, i16* nocapture readonly %pIn2, i32 %j, i32 %limit) { 6 7; CHECK-LABEL: smlaldx 8; CHECK: = phi i32 [ 0, %for.body.preheader.new ], 9; CHECK: [[ACC0:%[^ ]+]] = phi i64 [ 0, %for.body.preheader.new ], [ [[ACC2:%[^ ]+]], %for.body ] 10; CHECK: [[PIN21:%[^ ]+]] = bitcast i16* %pIn2.1 to i32* 11; CHECK: [[IN21:%[^ ]+]] = load i32, i32* [[PIN21]], align 2 12; CHECK: [[PIN10:%[^ ]+]] = bitcast i16* %pIn1.0 to i32* 13; CHECK: [[IN10:%[^ ]+]] = load i32, i32* [[PIN10]], align 2 14; CHECK: [[ACC1:%[^ ]+]] = call i64 @llvm.arm.smlaldx(i32 [[IN21]], i32 [[IN10]], i64 [[ACC0]]) 15; CHECK: [[PIN23:%[^ ]+]] = bitcast i16* %pIn2.3 to i32* 16; CHECK: [[IN23:%[^ ]+]] = load i32, i32* [[PIN23]], align 2 17; CHECK: [[PIN12:%[^ ]+]] = bitcast i16* %pIn1.2 to i32* 18; CHECK: [[IN12:%[^ ]+]] = load i32, i32* [[PIN12]], align 2 19; CHECK: [[ACC2]] = call i64 @llvm.arm.smlaldx(i32 [[IN23]], i32 [[IN12]], i64 [[ACC1]]) 20; CHECK-NOT: call i64 @llvm.arm.smlad 21; CHECK-UNSUPPORTED-NOT: call i64 @llvm.arm.smlad 22 23entry: 24 %cmp9 = icmp eq i32 %limit, 0 25 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader 26 27for.body.preheader: 28 %0 = add i32 %limit, -1 29 %xtraiter = and i32 %limit, 3 30 %1 = icmp ult i32 %0, 3 31 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new 32 33for.body.preheader.new: 34 %unroll_iter = sub i32 %limit, %xtraiter 35 br label %for.body 36 37for.cond.cleanup.loopexit.unr-lcssa: 38 %add.lcssa.ph = phi i64 [ undef, %for.body.preheader ], [ %add.3, %for.body ] 39 %i.011.unr = phi i32 [ 0, %for.body.preheader ], [ %inc.3, %for.body ] 40 %sum.010.unr = phi i64 [ 0, %for.body.preheader ], [ %add.3, %for.body ] 41 %lcmp.mod = icmp eq i32 %xtraiter, 0 42 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil 43 44for.body.epil: 45 %i.011.epil = phi i32 [ %inc.epil, %for.body.epil ], [ %i.011.unr, %for.cond.cleanup.loopexit.unr-lcssa ] 46 %sum.010.epil = phi i64 [ %add.epil, %for.body.epil ], [ %sum.010.unr, %for.cond.cleanup.loopexit.unr-lcssa ] 47 %epil.iter = phi i32 [ %epil.iter.sub, %for.body.epil ], [ %xtraiter, %for.cond.cleanup.loopexit.unr-lcssa ] 48 %sub.epil = sub i32 %j, %i.011.epil 49 %arrayidx.epil = getelementptr inbounds i16, i16* %pIn2, i32 %sub.epil 50 %2 = load i16, i16* %arrayidx.epil, align 2 51 %conv.epil = sext i16 %2 to i32 52 %arrayidx1.epil = getelementptr inbounds i16, i16* %pIn1, i32 %i.011.epil 53 %3 = load i16, i16* %arrayidx1.epil, align 2 54 %conv2.epil = sext i16 %3 to i32 55 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil 56 %sext.mul.epil = sext i32 %mul.epil to i64 57 %add.epil = add nsw i64 %sext.mul.epil, %sum.010.epil 58 %inc.epil = add nuw i32 %i.011.epil, 1 59 %epil.iter.sub = add i32 %epil.iter, -1 60 %epil.iter.cmp = icmp eq i32 %epil.iter.sub, 0 61 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil 62 63for.cond.cleanup: 64 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add.lcssa.ph, %for.cond.cleanup.loopexit.unr-lcssa ], [ %add.epil, %for.body.epil ] 65 ret i64 %sum.0.lcssa 66 67for.body: 68 %i.011 = phi i32 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ] 69 %sum.010 = phi i64 [ 0, %for.body.preheader.new ], [ %add.3, %for.body ] 70 %niter = phi i32 [ %unroll_iter, %for.body.preheader.new ], [ %niter.nsub.3, %for.body ] 71 %pIn2Base = phi i16* [ %pIn2, %for.body.preheader.new ], [ %pIn2.4, %for.body ] 72 %pIn2.0 = getelementptr inbounds i16, i16* %pIn2Base, i32 0 73 %In2 = load i16, i16* %pIn2.0, align 2 74 %pIn1.0 = getelementptr inbounds i16, i16* %pIn1, i32 %i.011 75 %In1 = load i16, i16* %pIn1.0, align 2 76 %inc = or i32 %i.011, 1 77 %pIn2.1 = getelementptr inbounds i16, i16* %pIn2Base, i32 -1 78 %In2.1 = load i16, i16* %pIn2.1, align 2 79 %pIn1.1 = getelementptr inbounds i16, i16* %pIn1, i32 %inc 80 %In1.1 = load i16, i16* %pIn1.1, align 2 81 %inc.1 = or i32 %i.011, 2 82 %pIn2.2 = getelementptr inbounds i16, i16* %pIn2Base, i32 -2 83 %In2.2 = load i16, i16* %pIn2.2, align 2 84 %pIn1.2 = getelementptr inbounds i16, i16* %pIn1, i32 %inc.1 85 %In1.2 = load i16, i16* %pIn1.2, align 2 86 %inc.2 = or i32 %i.011, 3 87 %pIn2.3 = getelementptr inbounds i16, i16* %pIn2Base, i32 -3 88 %In2.3 = load i16, i16* %pIn2.3, align 2 89 %pIn1.3 = getelementptr inbounds i16, i16* %pIn1, i32 %inc.2 90 %In1.3 = load i16, i16* %pIn1.3, align 2 91 %sextIn1 = sext i16 %In1 to i32 92 %sextIn1.1 = sext i16 %In1.1 to i32 93 %sextIn1.2 = sext i16 %In1.2 to i32 94 %sextIn1.3 = sext i16 %In1.3 to i32 95 %sextIn2 = sext i16 %In2 to i32 96 %sextIn2.1 = sext i16 %In2.1 to i32 97 %sextIn2.2 = sext i16 %In2.2 to i32 98 %sextIn2.3 = sext i16 %In2.3 to i32 99 %mul = mul nsw i32 %sextIn1, %sextIn2 100 %mul.1 = mul nsw i32 %sextIn1.1, %sextIn2.1 101 %mul.2 = mul nsw i32 %sextIn1.2, %sextIn2.2 102 %mul.3 = mul nsw i32 %sextIn1.3, %sextIn2.3 103 %sext.mul = sext i32 %mul to i64 104 %sext.mul.1 = sext i32 %mul.1 to i64 105 %sext.mul.2 = sext i32 %mul.2 to i64 106 %sext.mul.3 = sext i32 %mul.3 to i64 107 %add = add nsw i64 %sum.010, %sext.mul 108 %add.1 = add nsw i64 %sext.mul.1, %add 109 %add.2 = add nsw i64 %add.1, %sext.mul.2 110 %add.3 = add nsw i64 %sext.mul.3, %add.2 111 %inc.3 = add i32 %i.011, 4 112 %pIn2.4 = getelementptr inbounds i16, i16* %pIn2Base, i32 -4 113 %niter.nsub.3 = add i32 %niter, -4 114 %niter.ncmp.3 = icmp eq i32 %niter.nsub.3, 0 115 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body 116} 117 118define i64 @smlaldx_swap(i16* nocapture readonly %pIn1, i16* nocapture readonly %pIn2, i32 %j, i32 %limit) { 119 120entry: 121 %cmp9 = icmp eq i32 %limit, 0 122 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader 123 124for.body.preheader: 125 %0 = add i32 %limit, -1 126 %xtraiter = and i32 %limit, 3 127 %1 = icmp ult i32 %0, 3 128 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new 129 130for.body.preheader.new: 131 %unroll_iter = sub i32 %limit, %xtraiter 132 %scevgep6 = getelementptr i16, i16* %pIn1, i32 2 133 %2 = add i32 %j, -1 134 %scevgep11 = getelementptr i16, i16* %pIn2, i32 %2 135 br label %for.body 136 137for.cond.cleanup.loopexit.unr-lcssa: 138 %add.lcssa.ph = phi i64 [ undef, %for.body.preheader ], [ %add.3, %for.body ] 139 %i.011.unr = phi i32 [ 0, %for.body.preheader ], [ %inc.3, %for.body ] 140 %sum.010.unr = phi i64 [ 0, %for.body.preheader ], [ %add.3, %for.body ] 141 %lcmp.mod = icmp eq i32 %xtraiter, 0 142 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil.preheader 143 144for.body.epil.preheader: 145 %scevgep = getelementptr i16, i16* %pIn1, i32 %i.011.unr 146 %3 = sub i32 %j, %i.011.unr 147 %scevgep2 = getelementptr i16, i16* %pIn2, i32 %3 148 %4 = sub i32 0, %xtraiter 149 br label %for.body.epil 150 151for.body.epil: 152 %lsr.iv5 = phi i32 [ %4, %for.body.epil.preheader ], [ %lsr.iv.next, %for.body.epil ] 153 %lsr.iv3 = phi i16* [ %scevgep2, %for.body.epil.preheader ], [ %scevgep4, %for.body.epil ] 154 %lsr.iv = phi i16* [ %scevgep, %for.body.epil.preheader ], [ %scevgep1, %for.body.epil ] 155 %sum.010.epil = phi i64 [ %add.epil, %for.body.epil ], [ %sum.010.unr, %for.body.epil.preheader ] 156 %5 = load i16, i16* %lsr.iv3, align 2 157 %conv.epil = sext i16 %5 to i32 158 %6 = load i16, i16* %lsr.iv, align 2 159 %conv2.epil = sext i16 %6 to i32 160 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil 161 %sext.mul.epil = sext i32 %mul.epil to i64 162 %add.epil = add nsw i64 %sext.mul.epil, %sum.010.epil 163 %scevgep1 = getelementptr i16, i16* %lsr.iv, i32 1 164 %scevgep4 = getelementptr i16, i16* %lsr.iv3, i32 -1 165 %lsr.iv.next = add nsw i32 %lsr.iv5, 1 166 %epil.iter.cmp = icmp eq i32 %lsr.iv.next, 0 167 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil 168 169for.cond.cleanup: 170 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add.lcssa.ph, %for.cond.cleanup.loopexit.unr-lcssa ], [ %add.epil, %for.body.epil ] 171 ret i64 %sum.0.lcssa 172 173; CHECK-LABEL: smlaldx_swap 174; CHECK: for.body.preheader.new: 175; CHECK: [[PIN1Base:[^ ]+]] = getelementptr i16, i16* %pIn1 176; CHECK: [[PIN2Base:[^ ]+]] = getelementptr i16, i16* %pIn2 177 178; CHECK: for.body: 179; CHECK: [[PIN2:%[^ ]+]] = phi i16* [ [[PIN2_NEXT:%[^ ]+]], %for.body ], [ [[PIN2Base]], %for.body.preheader.new ] 180; CHECK: [[PIN1:%[^ ]+]] = phi i16* [ [[PIN1_NEXT:%[^ ]+]], %for.body ], [ [[PIN1Base]], %for.body.preheader.new ] 181; CHECK: [[IV:%[^ ]+]] = phi i32 182; CHECK: [[ACC0:%[^ ]+]] = phi i64 [ 0, %for.body.preheader.new ], [ [[ACC2:%[^ ]+]], %for.body ] 183 184; CHECK: [[PIN2_CAST:%[^ ]+]] = bitcast i16* [[PIN2]] to i32* 185; CHECK: [[IN2:%[^ ]+]] = load i32, i32* [[PIN2_CAST]], align 2 186 187; CHECK: [[PIN1_2:%[^ ]+]] = getelementptr i16, i16* [[PIN1]], i32 -2 188; CHECK: [[PIN1_2_CAST:%[^ ]+]] = bitcast i16* [[PIN1_2]] to i32* 189; CHECK: [[IN1_2:%[^ ]+]] = load i32, i32* [[PIN1_2_CAST]], align 2 190; CHECK: [[ACC1:%[^ ]+]] = call i64 @llvm.arm.smlaldx(i32 [[IN2]], i32 [[IN1_2]], i64 [[ACC0]]) 191 192; CHECK: [[PIN2_2:%[^ ]+]] = getelementptr i16, i16* [[PIN2]], i32 -2 193; CHECK: [[PIN2_2_CAST:%[^ ]+]] = bitcast i16* [[PIN2_2]] to i32* 194; CHECK: [[IN2_2:%[^ ]+]] = load i32, i32* [[PIN2_2_CAST]], align 2 195 196; CHECK: [[PIN1_CAST:%[^ ]+]] = bitcast i16* [[PIN1]] to i32* 197; CHECK: [[IN1:%[^ ]+]] = load i32, i32* [[PIN1_CAST]], align 2 198; CHECK: [[ACC2]] = call i64 @llvm.arm.smlaldx(i32 [[IN2_2]], i32 [[IN1]], i64 [[ACC1]]) 199 200; CHECK: [[PIN1_NEXT]] = getelementptr i16, i16* [[PIN1]], i32 4 201; CHECK: [[PIN2_NEXT]] = getelementptr i16, i16* [[PIN2]], i32 -4 202 203; CHECK-NOT: call i64 @llvm.arm.smlad 204; CHECK-UNSUPPORTED-NOT: call i64 @llvm.arm.smlad 205 206for.body: 207 %pin2 = phi i16* [ %pin2.sub4, %for.body ], [ %scevgep11, %for.body.preheader.new ] 208 %pin1 = phi i16* [ %pin1.add4, %for.body ], [ %scevgep6, %for.body.preheader.new ] 209 %i.011 = phi i32 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ] 210 %sum.010 = phi i64 [ 0, %for.body.preheader.new ], [ %add.3, %for.body ] 211 %pin2.add1 = getelementptr i16, i16* %pin2, i32 1 212 %In2 = load i16, i16* %pin2.add1, align 2 213 %pin1.sub2 = getelementptr i16, i16* %pin1, i32 -2 214 %In1 = load i16, i16* %pin1.sub2, align 2 215 %In2.1 = load i16, i16* %pin2, align 2 216 %pin1.sub1 = getelementptr i16, i16* %pin1, i32 -1 217 %In1.1 = load i16, i16* %pin1.sub1, align 2 218 %pin2.sub1 = getelementptr i16, i16* %pin2, i32 -1 219 %In2.2 = load i16, i16* %pin2.sub1, align 2 220 %In1.2 = load i16, i16* %pin1, align 2 221 %pin2.sub2 = getelementptr i16, i16* %pin2, i32 -2 222 %In2.3 = load i16, i16* %pin2.sub2, align 2 223 %pin1.add1 = getelementptr i16, i16* %pin1, i32 1 224 %In1.3 = load i16, i16* %pin1.add1, align 2 225 %sextIn2 = sext i16 %In2 to i32 226 %sextIn1 = sext i16 %In1 to i32 227 %sextIn2.1 = sext i16 %In2.1 to i32 228 %sextIn1.1 = sext i16 %In1.1 to i32 229 %sextIn2.2 = sext i16 %In2.2 to i32 230 %sextIn1.2 = sext i16 %In1.2 to i32 231 %sextIn2.3 = sext i16 %In2.3 to i32 232 %sextIn1.3 = sext i16 %In1.3 to i32 233 %mul = mul nsw i32 %sextIn2, %sextIn1 234 %sext.mul = sext i32 %mul to i64 235 %add = add nsw i64 %sext.mul, %sum.010 236 %mul.1 = mul nsw i32 %sextIn2.1, %sextIn1.1 237 %sext.mul.1 = sext i32 %mul.1 to i64 238 %add.1 = add nsw i64 %sext.mul.1, %add 239 %mul.2 = mul nsw i32 %sextIn2.2, %sextIn1.2 240 %sext.mul.2 = sext i32 %mul.2 to i64 241 %add.2 = add nsw i64 %add.1, %sext.mul.2 242 %mul.3 = mul nsw i32 %sextIn2.3, %sextIn1.3 243 %sext.mul.3 = sext i32 %mul.3 to i64 244 %add.3 = add nsw i64 %add.2, %sext.mul.3 245 %inc.3 = add i32 %i.011, 4 246 %pin1.add4 = getelementptr i16, i16* %pin1, i32 4 247 %pin2.sub4 = getelementptr i16, i16* %pin2, i32 -4 248 %niter.ncmp.3 = icmp eq i32 %unroll_iter, %inc.3 249 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body 250} 251