1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mtriple=thumbv8.1m.main-none-eabi -mattr=+mve.fp -o - %s | FileCheck --check-prefix=CHECK %s 3 4define void @tailpred(half* nocapture readonly %pSrcA, half* nocapture readonly %pSrcB, half* nocapture %pDst, i32 %blockSize) { 5; CHECK-LABEL: tailpred: 6; CHECK: @ %bb.0: @ %entry 7; CHECK-NEXT: .save {r4, r5, r7, lr} 8; CHECK-NEXT: push {r4, r5, r7, lr} 9; CHECK-NEXT: cmp r3, #0 10; CHECK-NEXT: beq .LBB0_6 11; CHECK-NEXT: @ %bb.1: @ %vector.memcheck 12; CHECK-NEXT: add.w r5, r2, r3, lsl #1 13; CHECK-NEXT: add.w r4, r1, r3, lsl #1 14; CHECK-NEXT: cmp r5, r1 15; CHECK-NEXT: cset r12, hi 16; CHECK-NEXT: cmp r4, r2 17; CHECK-NEXT: cset lr, hi 18; CHECK-NEXT: cmp r5, r0 19; CHECK-NEXT: add.w r5, r0, r3, lsl #1 20; CHECK-NEXT: cset r4, hi 21; CHECK-NEXT: cmp r5, r2 22; CHECK-NEXT: cset r5, hi 23; CHECK-NEXT: ands r4, r5 24; CHECK-NEXT: lsls r4, r4, #31 25; CHECK-NEXT: itt eq 26; CHECK-NEXT: andeq.w r5, lr, r12 27; CHECK-NEXT: lslseq.w r5, r5, #31 28; CHECK-NEXT: beq .LBB0_4 29; CHECK-NEXT: @ %bb.2: @ %while.body.preheader 30; CHECK-NEXT: dls lr, r3 31; CHECK-NEXT: .LBB0_3: @ %while.body 32; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1 33; CHECK-NEXT: vldr.16 s0, [r0] 34; CHECK-NEXT: vldr.16 s2, [r1] 35; CHECK-NEXT: adds r1, #2 36; CHECK-NEXT: adds r0, #2 37; CHECK-NEXT: vadd.f16 s0, s2, s0 38; CHECK-NEXT: vstr.16 s0, [r2] 39; CHECK-NEXT: adds r2, #2 40; CHECK-NEXT: le lr, .LBB0_3 41; CHECK-NEXT: b .LBB0_6 42; CHECK-NEXT: .LBB0_4: @ %vector.ph 43; CHECK-NEXT: dlstp.16 lr, r3 44; CHECK-NEXT: .LBB0_5: @ %vector.body 45; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1 46; CHECK-NEXT: vldrh.u16 q0, [r0], #16 47; CHECK-NEXT: vldrh.u16 q1, [r1], #16 48; CHECK-NEXT: vadd.f16 q0, q1, q0 49; CHECK-NEXT: vstrh.16 q0, [r2], #16 50; CHECK-NEXT: letp lr, .LBB0_5 51; CHECK-NEXT: .LBB0_6: @ %while.end 52; CHECK-NEXT: pop {r4, r5, r7, pc} 53entry: 54 %cmp.not6 = icmp eq i32 %blockSize, 0 55 br i1 %cmp.not6, label %while.end, label %vector.memcheck 56 57vector.memcheck: ; preds = %entry 58 %scevgep = getelementptr half, half* %pDst, i32 %blockSize 59 %scevgep14 = getelementptr half, half* %pSrcA, i32 %blockSize 60 %scevgep17 = getelementptr half, half* %pSrcB, i32 %blockSize 61 %bound0 = icmp ugt half* %scevgep14, %pDst 62 %bound1 = icmp ugt half* %scevgep, %pSrcA 63 %found.conflict = and i1 %bound0, %bound1 64 %bound019 = icmp ugt half* %scevgep17, %pDst 65 %bound120 = icmp ugt half* %scevgep, %pSrcB 66 %found.conflict21 = and i1 %bound019, %bound120 67 %conflict.rdx = or i1 %found.conflict, %found.conflict21 68 br i1 %conflict.rdx, label %while.body, label %vector.ph 69 70vector.ph: ; preds = %vector.memcheck 71 %n.rnd.up = add i32 %blockSize, 7 72 %n.vec = and i32 %n.rnd.up, -8 73 br label %vector.body 74 75vector.body: ; preds = %vector.body, %vector.ph 76 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ] 77 %next.gep = getelementptr half, half* %pSrcA, i32 %index 78 %next.gep28 = getelementptr half, half* %pDst, i32 %index 79 %next.gep29 = getelementptr half, half* %pSrcB, i32 %index 80 %active.lane.mask = call <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32 %index, i32 %blockSize) 81 %0 = bitcast half* %next.gep to <8 x half>* 82 %wide.masked.load = call <8 x half> @llvm.masked.load.v8f16.p0v8f16(<8 x half>* %0, i32 2, <8 x i1> %active.lane.mask, <8 x half> undef) 83 %1 = bitcast half* %next.gep29 to <8 x half>* 84 %wide.masked.load32 = call <8 x half> @llvm.masked.load.v8f16.p0v8f16(<8 x half>* %1, i32 2, <8 x i1> %active.lane.mask, <8 x half> undef) 85 %2 = fadd fast <8 x half> %wide.masked.load32, %wide.masked.load 86 %3 = bitcast half* %next.gep28 to <8 x half>* 87 call void @llvm.masked.store.v8f16.p0v8f16(<8 x half> %2, <8 x half>* %3, i32 2, <8 x i1> %active.lane.mask) 88 %index.next = add i32 %index, 8 89 %4 = icmp eq i32 %index.next, %n.vec 90 br i1 %4, label %while.end, label %vector.body 91 92while.body: ; preds = %vector.memcheck, %while.body 93 %blkCnt.010 = phi i32 [ %dec, %while.body ], [ %blockSize, %vector.memcheck ] 94 %pSrcA.addr.09 = phi half* [ %incdec.ptr, %while.body ], [ %pSrcA, %vector.memcheck ] 95 %pDst.addr.08 = phi half* [ %incdec.ptr3, %while.body ], [ %pDst, %vector.memcheck ] 96 %pSrcB.addr.07 = phi half* [ %incdec.ptr1, %while.body ], [ %pSrcB, %vector.memcheck ] 97 %incdec.ptr = getelementptr inbounds half, half* %pSrcA.addr.09, i32 1 98 %5 = load half, half* %pSrcA.addr.09, align 2 99 %incdec.ptr1 = getelementptr inbounds half, half* %pSrcB.addr.07, i32 1 100 %6 = load half, half* %pSrcB.addr.07, align 2 101 %7 = fadd fast half %6, %5 102 %incdec.ptr3 = getelementptr inbounds half, half* %pDst.addr.08, i32 1 103 store half %7, half* %pDst.addr.08, align 2 104 %dec = add i32 %blkCnt.010, -1 105 %cmp.not = icmp eq i32 %dec, 0 106 br i1 %cmp.not, label %while.end, label %while.body 107 108while.end: ; preds = %vector.body, %while.body, %entry 109 ret void 110} 111 112define void @notailpred(half* nocapture readonly %pSrcA, half* nocapture readonly %pSrcB, half* nocapture %pDst, i32 %blockSize) { 113; CHECK-LABEL: notailpred: 114; CHECK: @ %bb.0: @ %entry 115; CHECK-NEXT: .save {r4, r5, r6, r7, lr} 116; CHECK-NEXT: push {r4, r5, r6, r7, lr} 117; CHECK-NEXT: cbz r3, .LBB1_6 118; CHECK-NEXT: @ %bb.1: @ %while.body.preheader 119; CHECK-NEXT: cmp r3, #8 120; CHECK-NEXT: blo .LBB1_3 121; CHECK-NEXT: @ %bb.2: @ %vector.memcheck 122; CHECK-NEXT: add.w r5, r2, r3, lsl #1 123; CHECK-NEXT: add.w r6, r1, r3, lsl #1 124; CHECK-NEXT: cmp r5, r1 125; CHECK-NEXT: add.w r4, r0, r3, lsl #1 126; CHECK-NEXT: cset r7, hi 127; CHECK-NEXT: cmp r6, r2 128; CHECK-NEXT: cset r6, hi 129; CHECK-NEXT: cmp r5, r0 130; CHECK-NEXT: cset r5, hi 131; CHECK-NEXT: cmp r4, r2 132; CHECK-NEXT: cset r4, hi 133; CHECK-NEXT: ands r5, r4 134; CHECK-NEXT: lsls r5, r5, #31 135; CHECK-NEXT: itt eq 136; CHECK-NEXT: andeq r7, r6 137; CHECK-NEXT: lslseq.w r7, r7, #31 138; CHECK-NEXT: beq .LBB1_7 139; CHECK-NEXT: .LBB1_3: 140; CHECK-NEXT: mov r5, r3 141; CHECK-NEXT: mov r12, r0 142; CHECK-NEXT: mov r7, r2 143; CHECK-NEXT: mov r4, r1 144; CHECK-NEXT: .LBB1_4: @ %while.body.preheader31 145; CHECK-NEXT: dls lr, r5 146; CHECK-NEXT: .LBB1_5: @ %while.body 147; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1 148; CHECK-NEXT: vldr.16 s0, [r12] 149; CHECK-NEXT: vldr.16 s2, [r4] 150; CHECK-NEXT: adds r4, #2 151; CHECK-NEXT: add.w r12, r12, #2 152; CHECK-NEXT: vadd.f16 s0, s2, s0 153; CHECK-NEXT: vstr.16 s0, [r7] 154; CHECK-NEXT: adds r7, #2 155; CHECK-NEXT: le lr, .LBB1_5 156; CHECK-NEXT: .LBB1_6: @ %while.end 157; CHECK-NEXT: pop {r4, r5, r6, r7, pc} 158; CHECK-NEXT: .LBB1_7: @ %vector.ph 159; CHECK-NEXT: bic r6, r3, #7 160; CHECK-NEXT: movs r5, #1 161; CHECK-NEXT: sub.w r7, r6, #8 162; CHECK-NEXT: add.w r4, r1, r6, lsl #1 163; CHECK-NEXT: add.w r12, r0, r6, lsl #1 164; CHECK-NEXT: add.w lr, r5, r7, lsr #3 165; CHECK-NEXT: add.w r7, r2, r6, lsl #1 166; CHECK-NEXT: dls lr, lr 167; CHECK-NEXT: and r5, r3, #7 168; CHECK-NEXT: .LBB1_8: @ %vector.body 169; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1 170; CHECK-NEXT: vldrh.u16 q0, [r0], #16 171; CHECK-NEXT: vldrh.u16 q1, [r1], #16 172; CHECK-NEXT: vadd.f16 q0, q1, q0 173; CHECK-NEXT: vstrb.8 q0, [r2], #16 174; CHECK-NEXT: le lr, .LBB1_8 175; CHECK-NEXT: @ %bb.9: @ %middle.block 176; CHECK-NEXT: cmp r6, r3 177; CHECK-NEXT: bne .LBB1_4 178; CHECK-NEXT: b .LBB1_6 179entry: 180 %cmp.not6 = icmp eq i32 %blockSize, 0 181 br i1 %cmp.not6, label %while.end, label %while.body.preheader 182 183while.body.preheader: ; preds = %entry 184 %min.iters.check = icmp ult i32 %blockSize, 8 185 br i1 %min.iters.check, label %while.body.preheader31, label %vector.memcheck 186 187vector.memcheck: ; preds = %while.body.preheader 188 %scevgep = getelementptr half, half* %pDst, i32 %blockSize 189 %scevgep14 = getelementptr half, half* %pSrcA, i32 %blockSize 190 %scevgep17 = getelementptr half, half* %pSrcB, i32 %blockSize 191 %bound0 = icmp ugt half* %scevgep14, %pDst 192 %bound1 = icmp ugt half* %scevgep, %pSrcA 193 %found.conflict = and i1 %bound0, %bound1 194 %bound019 = icmp ugt half* %scevgep17, %pDst 195 %bound120 = icmp ugt half* %scevgep, %pSrcB 196 %found.conflict21 = and i1 %bound019, %bound120 197 %conflict.rdx = or i1 %found.conflict, %found.conflict21 198 br i1 %conflict.rdx, label %while.body.preheader31, label %vector.ph 199 200vector.ph: ; preds = %vector.memcheck 201 %n.vec = and i32 %blockSize, -8 202 %ind.end = and i32 %blockSize, 7 203 %ind.end23 = getelementptr half, half* %pSrcA, i32 %n.vec 204 %ind.end25 = getelementptr half, half* %pDst, i32 %n.vec 205 %ind.end27 = getelementptr half, half* %pSrcB, i32 %n.vec 206 br label %vector.body 207 208vector.body: ; preds = %vector.body, %vector.ph 209 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ] 210 %next.gep = getelementptr half, half* %pSrcA, i32 %index 211 %next.gep28 = getelementptr half, half* %pDst, i32 %index 212 %next.gep29 = getelementptr half, half* %pSrcB, i32 %index 213 %0 = bitcast half* %next.gep to <8 x half>* 214 %wide.load = load <8 x half>, <8 x half>* %0, align 2 215 %1 = bitcast half* %next.gep29 to <8 x half>* 216 %wide.load30 = load <8 x half>, <8 x half>* %1, align 2 217 %2 = fadd fast <8 x half> %wide.load30, %wide.load 218 %3 = bitcast half* %next.gep28 to <8 x half>* 219 store <8 x half> %2, <8 x half>* %3, align 2 220 %index.next = add i32 %index, 8 221 %4 = icmp eq i32 %index.next, %n.vec 222 br i1 %4, label %middle.block, label %vector.body 223 224middle.block: ; preds = %vector.body 225 %cmp.n = icmp eq i32 %n.vec, %blockSize 226 br i1 %cmp.n, label %while.end, label %while.body.preheader31 227 228while.body.preheader31: ; preds = %middle.block, %vector.memcheck, %while.body.preheader 229 %blkCnt.010.ph = phi i32 [ %blockSize, %vector.memcheck ], [ %blockSize, %while.body.preheader ], [ %ind.end, %middle.block ] 230 %pSrcA.addr.09.ph = phi half* [ %pSrcA, %vector.memcheck ], [ %pSrcA, %while.body.preheader ], [ %ind.end23, %middle.block ] 231 %pDst.addr.08.ph = phi half* [ %pDst, %vector.memcheck ], [ %pDst, %while.body.preheader ], [ %ind.end25, %middle.block ] 232 %pSrcB.addr.07.ph = phi half* [ %pSrcB, %vector.memcheck ], [ %pSrcB, %while.body.preheader ], [ %ind.end27, %middle.block ] 233 br label %while.body 234 235while.body: ; preds = %while.body.preheader31, %while.body 236 %blkCnt.010 = phi i32 [ %dec, %while.body ], [ %blkCnt.010.ph, %while.body.preheader31 ] 237 %pSrcA.addr.09 = phi half* [ %incdec.ptr, %while.body ], [ %pSrcA.addr.09.ph, %while.body.preheader31 ] 238 %pDst.addr.08 = phi half* [ %incdec.ptr3, %while.body ], [ %pDst.addr.08.ph, %while.body.preheader31 ] 239 %pSrcB.addr.07 = phi half* [ %incdec.ptr1, %while.body ], [ %pSrcB.addr.07.ph, %while.body.preheader31 ] 240 %incdec.ptr = getelementptr inbounds half, half* %pSrcA.addr.09, i32 1 241 %5 = load half, half* %pSrcA.addr.09, align 2 242 %incdec.ptr1 = getelementptr inbounds half, half* %pSrcB.addr.07, i32 1 243 %6 = load half, half* %pSrcB.addr.07, align 2 244 %7 = fadd fast half %6, %5 245 %incdec.ptr3 = getelementptr inbounds half, half* %pDst.addr.08, i32 1 246 store half %7, half* %pDst.addr.08, align 2 247 %dec = add i32 %blkCnt.010, -1 248 %cmp.not = icmp eq i32 %dec, 0 249 br i1 %cmp.not, label %while.end, label %while.body 250 251while.end: ; preds = %while.body, %middle.block, %entry 252 ret void 253} 254 255declare <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32, i32) #1 256declare <8 x half> @llvm.masked.load.v8f16.p0v8f16(<8 x half>*, i32 immarg, <8 x i1>, <8 x half>) #2 257declare void @llvm.masked.store.v8f16.p0v8f16(<8 x half>, <8 x half>*, i32 immarg, <8 x i1>) #3 258