[PATCH] D80682: [NFC][SLP] Add test case exposing SLP cost model bug.

[Valeriy Dmitriev via Phabricator via llvm-commits]

vdmitrie created this revision.
vdmitrie added reviewers: ABataev, RKSimon, spatel.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

The bug is related to aggregate build cost model adjustment that
adds a bias to cost triggering vectorization of actually unprofitable
to vectorize tree.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D80682

Files:
  llvm/test/Transforms/SLPVectorizer/X86/slp_vec_list_bias.ll


Index: llvm/test/Transforms/SLPVectorizer/X86/slp_vec_list_bias.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/SLPVectorizer/X86/slp_vec_list_bias.ll
@@ -0,0 +1,64 @@
+; RUN: opt < %s -slp-vectorizer -mtriple=x86_64-unknown-linux-gnu -mcpu=skylake -tti -S | FileCheck %s
+
+; Check no vectorization triggered with any portion of
+; insertelement <8 x i32> instructions that build entire vector.
+; Vectorization triggered by cost bias caused by substructing
+; the cost of entire "aggregate build" sequence while
+; building vectorizable tree from only a portion of it.
+; FIXME: this is unprofitable to vectorize.
+; CHECK:  insertelement <4 x i32>
+; CHECK:  extractelement <4 x i32>
+; CHECK-NOT:  insertelement <2 x i32>
+; CHECK-NOT:  extractelement <2 x i32>
+
+
+define void @test(i32* nocapture %tmp2) {
+
+  %tmp3 = load i32, i32* %tmp2, align 4
+  %tmp4 = getelementptr inbounds i32, i32* %tmp2, i64 7
+  %tmp5 = load i32, i32* %tmp4, align 4
+  %tmp8 = getelementptr inbounds i32, i32* %tmp2, i64 1
+  %tmp9 = load i32, i32* %tmp8, align 4
+  %tmp10 = getelementptr inbounds i32, i32* %tmp2, i64 6
+  %tmp11 = load i32, i32* %tmp10, align 4
+  %tmp14 = getelementptr inbounds i32, i32* %tmp2, i64 2
+  %tmp15 = load i32, i32* %tmp14, align 4
+  %tmp16 = getelementptr inbounds i32, i32* %tmp2, i64 5
+  %tmp17 = load i32, i32* %tmp16, align 4
+  %tmp20 = getelementptr inbounds i32, i32* %tmp2, i64 3
+  %tmp21 = load i32, i32* %tmp20, align 4
+  %tmp22 = getelementptr inbounds i32, i32* %tmp2, i64 4
+  %tmp23 = load i32, i32* %tmp22, align 4
+  %tmp24 = add nsw i32 %tmp23, %tmp21
+  %tmp25 = sub nsw i32 %tmp21, %tmp23
+  %tmp27 = sub nsw i32 %tmp3, %tmp24
+  %tmp28 = add nsw i32 %tmp15, %tmp9
+  %tmp29 = sub nsw i32 %tmp9, %tmp15
+  %tmp30 = add nsw i32 %tmp27, %tmp29
+  %tmp31 = mul nsw i32 %tmp30, 4433
+  %tmp32 = mul nsw i32 %tmp27, 6270
+  %tmp34 = mul nsw i32 %tmp29, -15137
+  %tmp37 = add nsw i32 %tmp25, %tmp11
+  %tmp38 = add nsw i32 %tmp17, %tmp5
+  %tmp39 = add nsw i32 %tmp37, %tmp38
+  %tmp40 = mul nsw i32 %tmp39, 9633
+  %tmp41 = mul nsw i32 %tmp25, 2446
+  %tmp42 = mul nsw i32 %tmp17, 16819
+  %tmp47 = mul nsw i32 %tmp37, -16069
+  %tmp48 = mul nsw i32 %tmp38, -3196
+  %tmp49 = add nsw i32 %tmp40, %tmp47
+  %tmp50 = add nsw i32 %tmp40, %tmp48
+  %tmp65 = insertelement <8 x i32> undef, i32 %tmp28, i32 0
+  %tmp66 = insertelement <8 x i32> %tmp65, i32 %tmp50, i32 1
+  %tmp67 = insertelement <8 x i32> %tmp66, i32 %tmp32, i32 2
+  %tmp68 = insertelement <8 x i32> %tmp67, i32 %tmp49, i32 3
+  %tmp69 = insertelement <8 x i32> %tmp68, i32 %tmp28, i32 4
+  %tmp70 = insertelement <8 x i32> %tmp69, i32 %tmp50, i32 5
+  %tmp71 = insertelement <8 x i32> %tmp70, i32 %tmp34, i32 6
+  %tmp72 = insertelement <8 x i32> %tmp71, i32 %tmp49, i32 7
+  %tmp76 = shl <8 x i32> %tmp72, <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>
+  %tmp79 = bitcast i32* %tmp2 to <8 x i32>*
+  store <8 x i32> %tmp76, <8 x i32>* %tmp79, align 4
+
+  unreachable
+}


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D80682.266712.patch
Type: text/x-patch
Size: 3051 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200528/2eff12b5/attachment.bin>