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