1; RUN: opt -mtriple=armv8m.main-none-none-eabi -mattr=+dsp -arm-parallel-dsp %s -S -o - | FileCheck %s 2; RUN: opt -mtriple=armv6m-none-none-eabi < %s -arm-parallel-dsp -S | FileCheck %s --check-prefix=CHECK-UNSUPPORTED 3; RUN: opt -mtriple=armv8m.main-none-none-eabi -mattr=-dsp < %s -arm-parallel-dsp -S | FileCheck %s --check-prefix=CHECK-UNSUPPORTED 4; RUN: opt -mtriple=armv8m.maineb-none-none-eabi -mattr=+dsp < %s -arm-parallel-dsp -S | FileCheck %s --check-prefix=CHECK-UNSUPPORTED 5 6define i32 @smladx(ptr nocapture readonly %pIn1, ptr 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: [[IN21:%[^ ]+]] = load i32, ptr %pIn2.1, align 2 12; CHECK: [[IN10:%[^ ]+]] = load i32, ptr %pIn1.0, align 2 13; CHECK: [[ACC1:%[^ ]+]] = call i32 @llvm.arm.smladx(i32 [[IN21]], i32 [[IN10]], i32 [[ACC0]]) 14 15; CHECK: [[IN23:%[^ ]+]] = load i32, ptr %pIn2.3, align 2 16; CHECK: [[IN12:%[^ ]+]] = load i32, ptr %pIn1.2, align 2 17; CHECK: [[ACC2]] = call i32 @llvm.arm.smladx(i32 [[IN23]], i32 [[IN12]], i32 [[ACC1]]) 18; CHECK-NOT: call i32 @llvm.arm.smlad 19; CHECK-UNSUPPORTED-NOT: call i32 @llvm.arm.smlad 20 21entry: 22 %cmp9 = icmp eq i32 %limit, 0 23 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader 24 25for.body.preheader: 26 %0 = add i32 %limit, -1 27 %xtraiter = and i32 %limit, 3 28 %1 = icmp ult i32 %0, 3 29 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new 30 31for.body.preheader.new: 32 %unroll_iter = sub i32 %limit, %xtraiter 33 br label %for.body 34 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, 0 40 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil 41 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.epil 47 %arrayidx.epil = getelementptr inbounds i16, ptr %pIn2, i32 %sub.epil 48 %2 = load i16, ptr %arrayidx.epil, align 2 49 %conv.epil = sext i16 %2 to i32 50 %arrayidx1.epil = getelementptr inbounds i16, ptr %pIn1, i32 %i.011.epil 51 %3 = load i16, ptr %arrayidx1.epil, align 2 52 %conv2.epil = sext i16 %3 to i32 53 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil 54 %add.epil = add nsw i32 %mul.epil, %sum.010.epil 55 %inc.epil = add nuw i32 %i.011.epil, 1 56 %epil.iter.sub = add i32 %epil.iter, -1 57 %epil.iter.cmp = icmp eq i32 %epil.iter.sub, 0 58 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil 59 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.lcssa 63 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 2 70 %pIn1.0 = getelementptr inbounds i16, ptr %pIn1, i32 %i.011 71 %In1 = load i16, ptr %pIn1.0, align 2 72 %inc = or disjoint i32 %i.011, 1 73 %pIn2.1 = getelementptr inbounds i16, ptr %pIn2Base, i32 -1 74 %In2.1 = load i16, ptr %pIn2.1, align 2 75 %pIn1.1 = getelementptr inbounds i16, ptr %pIn1, i32 %inc 76 %In1.1 = load i16, ptr %pIn1.1, align 2 77 %inc.1 = or disjoint i32 %i.011, 2 78 %pIn2.2 = getelementptr inbounds i16, ptr %pIn2Base, i32 -2 79 %In2.2 = load i16, ptr %pIn2.2, align 2 80 %pIn1.2 = getelementptr inbounds i16, ptr %pIn1, i32 %inc.1 81 %In1.2 = load i16, ptr %pIn1.2, align 2 82 %inc.2 = or disjoint i32 %i.011, 3 83 %pIn2.3 = getelementptr inbounds i16, ptr %pIn2Base, i32 -3 84 %In2.3 = load i16, ptr %pIn2.3, align 2 85 %pIn1.3 = getelementptr inbounds i16, ptr %pIn1, i32 %inc.2 86 %In1.3 = load i16, ptr %pIn1.3, align 2 87 %sextIn1 = sext i16 %In1 to i32 88 %sextIn1.1 = sext i16 %In1.1 to i32 89 %sextIn1.2 = sext i16 %In1.2 to i32 90 %sextIn1.3 = sext i16 %In1.3 to i32 91 %sextIn2 = sext i16 %In2 to i32 92 %sextIn2.1 = sext i16 %In2.1 to i32 93 %sextIn2.2 = sext i16 %In2.2 to i32 94 %sextIn2.3 = sext i16 %In2.3 to i32 95 %mul = mul nsw i32 %sextIn1, %sextIn2 96 %mul.1 = mul nsw i32 %sextIn1.1, %sextIn2.1 97 %mul.2 = mul nsw i32 %sextIn1.2, %sextIn2.2 98 %mul.3 = mul nsw i32 %sextIn1.3, %sextIn2.3 99 %add = add nsw i32 %mul, %sum.010 100 %add.1 = add nsw i32 %mul.1, %add 101 %add.2 = add nsw i32 %mul.2, %add.1 102 %add.3 = add nsw i32 %mul.3, %add.2 103 %inc.3 = add i32 %i.011, 4 104 %pIn2.4 = getelementptr inbounds i16, ptr %pIn2Base, i32 -4 105 %niter.nsub.3 = add i32 %niter, -4 106 %niter.ncmp.3 = icmp eq i32 %niter.nsub.3, 0 107 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body 108} 109 110define i32 @smladx_swap(ptr nocapture readonly %pIn1, ptr nocapture readonly %pIn2, i32 %j, i32 %limit) { 111 112; CHECK-LABEL: smladx_swap 113; CHECK: for.body.preheader.new: 114; CHECK: [[PIN1Base:[^ ]+]] = getelementptr i16, ptr %pIn1 115; CHECK: [[PIN2Base:[^ ]+]] = getelementptr i16, ptr %pIn2 116 117; 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 i32 121; CHECK: [[ACC0:%[^ ]+]] = phi i32 [ 0, %for.body.preheader.new ], [ [[ACC2:%[^ ]+]], %for.body ] 122 123; CHECK: [[IN2:%[^ ]+]] = load i32, ptr [[PIN2]], align 2 124 125; CHECK: [[PIN1_2:%[^ ]+]] = getelementptr i16, ptr [[PIN1]], i32 -2 126; CHECK: [[IN1_2:%[^ ]+]] = load i32, ptr [[PIN1_2]], align 2 127; CHECK: [[ACC1:%[^ ]+]] = call i32 @llvm.arm.smladx(i32 [[IN2]], i32 [[IN1_2]], i32 [[ACC0]]) 128 129; CHECK: [[PIN2_2:%[^ ]+]] = getelementptr i16, ptr [[PIN2]], i32 -2 130; CHECK: [[IN2_2:%[^ ]+]] = load i32, ptr [[PIN2_2]], align 2 131 132; CHECK: [[IN1:%[^ ]+]] = load i32, ptr [[PIN1]], align 2 133 134; CHECK: [[ACC2]] = call i32 @llvm.arm.smladx(i32 [[IN2_2]], i32 [[IN1]], i32 [[ACC1]]) 135 136; CHECK: [[PIN1_NEXT]] = getelementptr i16, ptr [[PIN1]], i32 4 137; CHECK: [[PIN2_NEXT]] = getelementptr i16, ptr [[PIN2]], i32 -4 138 139; CHECK-NOT: call i32 @llvm.arm.smlad 140; CHECK-UNSUPPORTED-NOT: call i32 @llvm.arm.smlad 141 142entry: 143 %cmp9 = icmp eq i32 %limit, 0 144 br i1 %cmp9, label %for.cond.cleanup, label %for.body.preheader 145 146for.body.preheader: 147 %0 = add i32 %limit, -1 148 %xtraiter = and i32 %limit, 3 149 %1 = icmp ult i32 %0, 3 150 br i1 %1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new 151 152for.body.preheader.new: 153 %unroll_iter = sub i32 %limit, %xtraiter 154 %scevgep6 = getelementptr i16, ptr %pIn1, i32 2 155 %2 = add i32 %j, -1 156 %scevgep11 = getelementptr i16, ptr %pIn2, i32 %2 157 br label %for.body 158 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, 0 164 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil.preheader 165 166for.body.epil.preheader: 167 %scevgep = getelementptr i16, ptr %pIn1, i32 %i.011.unr 168 %3 = sub i32 %j, %i.011.unr 169 %scevgep2 = getelementptr i16, ptr %pIn2, i32 %3 170 %4 = sub i32 0, %xtraiter 171 br label %for.body.epil 172 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 2 179 %conv.epil = sext i16 %5 to i32 180 %6 = load i16, ptr %lsr.iv, align 2 181 %conv2.epil = sext i16 %6 to i32 182 %mul.epil = mul nsw i32 %conv2.epil, %conv.epil 183 %add.epil = add nsw i32 %mul.epil, %sum.010.epil 184 %scevgep1 = getelementptr i16, ptr %lsr.iv, i32 1 185 %scevgep4 = getelementptr i16, ptr %lsr.iv3, i32 -1 186 %lsr.iv.next = add nsw i32 %lsr.iv5, 1 187 %epil.iter.cmp = icmp eq i32 %lsr.iv.next, 0 188 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil 189 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.lcssa 193 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 1 200 %In2 = load i16, ptr %pin2_add1, align 2 201 %pin1_sub2 = getelementptr i16, ptr %pin1, i32 -2 202 %In1 = load i16, ptr %pin1_sub2, align 2 203 %In2.1 = load i16, ptr %pin2, align 2 204 %pin1_sub1 = getelementptr i16, ptr %pin1, i32 -1 205 %In1.1 = load i16, ptr %pin1_sub1, align 2 206 %pin2_sub1 = getelementptr i16, ptr %pin2, i32 -1 207 %In2.2 = load i16, ptr %pin2_sub1, align 2 208 %In1.2 = load i16, ptr %pin1, align 2 209 %pin2_sub2 = getelementptr i16, ptr %pin2, i32 -2 210 %In2.3 = load i16, ptr %pin2_sub2, align 2 211 %pin1_add1 = getelementptr i16, ptr %pin1, i32 1 212 %In1.3 = load i16, ptr %pin1_add1, align 2 213 %sextIn2 = sext i16 %In2 to i32 214 %sextIn1 = sext i16 %In1 to i32 215 %sextIn2.1 = sext i16 %In2.1 to i32 216 %sextIn1.1 = sext i16 %In1.1 to i32 217 %sextIn2.2 = sext i16 %In2.2 to i32 218 %sextIn1.2 = sext i16 %In1.2 to i32 219 %sextIn2.3 = sext i16 %In2.3 to i32 220 %sextIn1.3 = sext i16 %In1.3 to i32 221 %mul = mul nsw i32 %sextIn2, %sextIn1 222 %add = add nsw i32 %mul, %sum.010 223 %mul.1 = mul nsw i32 %sextIn2.1, %sextIn1.1 224 %add.1 = add nsw i32 %mul.1, %add 225 %mul.2 = mul nsw i32 %sextIn2.2, %sextIn1.2 226 %add.2 = add nsw i32 %mul.2, %add.1 227 %mul.3 = mul nsw i32 %sextIn2.3, %sextIn1.3 228 %add.3 = add nsw i32 %mul.3, %add.2 229 %inc.3 = add i32 %i.011, 4 230 %pin1_add4 = getelementptr i16, ptr %pin1, i32 4 231 %pin2_sub4 = getelementptr i16, ptr %pin2, i32 -4 232 %niter.ncmp.3 = icmp eq i32 %unroll_iter, %inc.3 233 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body 234} 235