xref: /llvm-project/llvm/test/Transforms/LoopVectorize/scalable-loop-unpredicated-body-scalar-tail.ll (revision 56c091ea7106507b36015297ee9005c9d5fab0bf)
1; RUN: opt -S -passes=loop-vectorize,instcombine -force-vector-interleave=1 -force-vector-width=4 -force-target-supports-scalable-vectors=true -scalable-vectorization=on < %s | FileCheck %s --check-prefix=CHECKUF1
2; RUN: opt -S -passes=loop-vectorize,instcombine -force-vector-interleave=2 -force-vector-width=4 -force-target-supports-scalable-vectors=true -scalable-vectorization=on < %s | FileCheck %s --check-prefix=CHECKUF2
3
4; CHECKUF1: for.body.preheader:
5; CHECKUF1-DAG: %wide.trip.count = zext nneg i32 %N to i64
6; CHECKUF1-DAG: %[[VSCALE:.*]] = call i64 @llvm.vscale.i64()
7; CHECKUF1-DAG: %[[VSCALEX4:.*]] = shl i64 %[[VSCALE]], 2
8; CHECKUF1-DAG: %min.iters.check = icmp ugt i64 %[[VSCALEX4]], %wide.trip.count
9
10; CHECKUF1: vector.ph:
11; CHECKUF1-DAG:  %[[VSCALE:.*]] = call i64 @llvm.vscale.i64()
12; CHECKUF1-DAG:  %[[VSCALEX4:.*]] = shl i64 %[[VSCALE]], 2
13; CHECKUF1-DAG:  %n.mod.vf = urem i64 %wide.trip.count, %[[VSCALEX4]]
14; CHECKUF1:      %n.vec = sub nsw i64 %wide.trip.count, %n.mod.vf
15; CHECKUF1: %[[VSCALE:.*]] = call i64 @llvm.vscale.i64()
16; CHECKUF1: %[[VSCALEX4:.*]] = shl i64 %[[VSCALE]], 2
17
18; CHECKUF1: vector.body:
19; CHECKUF1: %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
20; CHECKUF1: %[[IDXB:.*]] = getelementptr inbounds double, ptr %b, i64 %index
21; CHECKUF1: %wide.load = load <vscale x 4 x double>, ptr %[[IDXB]], align 8
22; CHECKUF1: %[[FADD:.*]] = fadd <vscale x 4 x double> %wide.load, splat (double 1.000000e+00)
23; CHECKUF1: %[[IDXA:.*]] = getelementptr inbounds double, ptr %a, i64 %index
24; CHECKUF1: store <vscale x 4 x double> %[[FADD]], ptr %[[IDXA]], align 8
25; CHECKUF1: %index.next = add nuw i64 %index, %[[VSCALEX4]]
26; CHECKUF1: %[[CMP:.*]] = icmp eq i64 %index.next, %n.vec
27; CHECKUF1: br i1 %[[CMP]], label %middle.block, label %vector.body, !llvm.loop !0
28
29
30; For an interleave factor of 2, vscale is scaled by 8 instead of 4 (and thus shifted left by 3 instead of 2).
31; There is also the increment for the next iteration, e.g. instead of indexing IDXB, it indexes at IDXB + vscale * 4.
32
33; CHECKUF2: for.body.preheader:
34; CHECKUF2-DAG: %wide.trip.count = zext nneg i32 %N to i64
35; CHECKUF2-DAG: %[[VSCALE:.*]] = call i64 @llvm.vscale.i64()
36; CHECKUF2-DAG: %[[VSCALEX8:.*]] = shl i64 %[[VSCALE]], 3
37; CHECKUF2-DAG: %min.iters.check = icmp ugt i64 %[[VSCALEX8]], %wide.trip.count
38
39; CHECKUF2: vector.ph:
40; CHECKUF2-DAG:  %[[VSCALE:.*]] = call i64 @llvm.vscale.i64()
41; CHECKUF2-DAG:  %[[VSCALEX8:.*]] = shl i64 %[[VSCALE]], 3
42; CHECKUF2-DAG:  %n.mod.vf = urem i64 %wide.trip.count, %[[VSCALEX8]]
43; CHECKUF2:      %n.vec = sub nsw i64 %wide.trip.count, %n.mod.vf
44; CHECKUF2: %[[VSCALE:.*]] = call i64 @llvm.vscale.i64()
45; CHECKUF2: %[[VSCALEX8:.*]] = shl i64 %[[VSCALE]], 3
46
47; CHECKUF2: vector.body:
48; CHECKUF2: %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
49; CHECKUF2: %[[IDXB:.*]] = getelementptr inbounds double, ptr %b, i64 %index
50; CHECKUF2: %[[VSCALE:.*]] = call i64 @llvm.vscale.i64()
51; CHECKUF2: %[[VSCALE2:.*]] = shl i64 %[[VSCALE]], 5
52; CHECKUF2: %[[IDXB_NEXT:.*]] = getelementptr inbounds i8, ptr %[[IDXB]], i64 %[[VSCALE2]]
53; CHECKUF2: %wide.load = load <vscale x 4 x double>, ptr %[[IDXB]], align 8
54; CHECKUF2: %wide.load{{[0-9]+}} = load <vscale x 4 x double>, ptr %[[IDXB_NEXT]], align 8
55; CHECKUF2: %[[FADD:.*]] = fadd <vscale x 4 x double> %wide.load, splat (double 1.000000e+00)
56; CHECKUF2: %[[FADD_NEXT:.*]] = fadd <vscale x 4 x double> %wide.load{{[0-9]+}}, splat (double 1.000000e+00)
57; CHECKUF2: %[[IDXA:.*]] = getelementptr inbounds double, ptr %a, i64 %index
58; CHECKUF2: %[[VSCALE:.*]] = call i64 @llvm.vscale.i64()
59; CHECKUF2: %[[VSCALE2:.*]] = shl i64 %[[VSCALE]], 5
60; CHECKUF2: %[[IDXA_NEXT:.*]] = getelementptr inbounds i8, ptr %[[IDXA]], i64 %[[VSCALE2]]
61; CHECKUF2: store <vscale x 4 x double> %[[FADD]], ptr %[[IDXA]], align 8
62; CHECKUF2: store <vscale x 4 x double> %[[FADD_NEXT]], ptr %[[IDXA_NEXT]], align 8
63; CHECKUF2: %index.next = add nuw i64 %index, %[[VSCALEX8]]
64; CHECKUF2: %[[CMP:.*]] = icmp eq i64 %index.next, %n.vec
65; CHECKUF2: br i1 %[[CMP]], label %middle.block, label %vector.body, !llvm.loop !0
66
67define void @loop(i32 %N, ptr nocapture %a, ptr nocapture readonly %b) {
68entry:
69  %cmp7 = icmp sgt i32 %N, 0
70  br i1 %cmp7, label %for.body.preheader, label %for.cond.cleanup
71
72for.body.preheader:                               ; preds = %entry
73  %wide.trip.count = zext i32 %N to i64
74  br label %for.body
75
76for.cond.cleanup:                                 ; preds = %for.body, %entry
77  ret void
78
79for.body:                                         ; preds = %for.body.preheader, %for.body
80  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]
81  %arrayidx = getelementptr inbounds double, ptr %b, i64 %indvars.iv
82  %0 = load double, ptr %arrayidx, align 8
83  %add = fadd double %0, 1.000000e+00
84  %arrayidx2 = getelementptr inbounds double, ptr %a, i64 %indvars.iv
85  store double %add, ptr %arrayidx2, align 8
86  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
87  %exitcond.not = icmp eq i64 %indvars.iv.next, %wide.trip.count
88  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !1
89}
90
91!1 = distinct !{!1, !2}
92!2 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
93