1 ; RUN: opt -S -passes=loop-vectorize -prefer-predicate-over-epilogue=predicate-else-scalar-epilogue <%s | FileCheck %s
3 target triple = "aarch64-unknown-linux-gnu"
5 ; The uniform load of %d in the following loop triggers the special
6 ; branch costing code in LoopVectorizationCostModel::getInstructionCost.
7 ; However, this should only affect the fixed-width cost because for
8 ; NEON it needs to scalarize the load, whereas for SVE it can use a predicated load.
9 ; Because of how the LoopVectorizer annotates the load to need scalarization with
10 ; predicated blocks, this leads to different costs for the branch instruction.
12 ; NOTE: This test assumes we will never use a fixed-width VF due to
13 ; the high cost of scalarizing the masked store, however this assumption may
14 ; break in future if we permit the use of SVE loads and stores to perform the
15 ; fixed-width operations.
16 define i32 @uniform_load(i64 %n, ptr readnone %c, ptr %d) #0 {
17 ; CHECK-LABEL: @uniform_load(
18 ; CHECK: call void @llvm.masked.store.nxv4f32.p0(<vscale x 4 x float>
22 for.body: ; preds = %entry, %for.body
23 %indvars.iv = phi i64 [ 1, %entry ], [ %indvars.iv.next, %for.body ]
24 %load2 = load float, ptr %d, align 4
25 %arrayidx2 = getelementptr inbounds float, ptr %c, i64 %indvars.iv
26 store float %load2, ptr %arrayidx2, align 4
27 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
28 %exitcond.not = icmp eq i64 %indvars.iv.next, %n
29 br i1 %exitcond.not, label %for.end, label %for.body
31 for.end: ; preds = %for.body
35 attributes #0 = { vscale_range(1,16) "target-features"="+sve" }