[PATCH] D76956: [TTI][SLP] Add TTI interface to estimate cost of chain of vector inserts/extracts.

[Valeriy Dmitriev via Phabricator via llvm-commits]

vdmitrie marked an inline comment as done.
vdmitrie added inline comments.


================
Comment at: llvm/test/Transforms/SLPVectorizer/X86/resched.ll:22-29
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <8 x i32> undef, i32 [[CONV31_I]], i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <8 x i32> [[TMP1]], i32 [[CONV31_I]], i32 1
+; CHECK-NEXT:    [[TMP3:%.*]] = insertelement <8 x i32> [[TMP2]], i32 [[CONV31_I]], i32 2
+; CHECK-NEXT:    [[TMP4:%.*]] = insertelement <8 x i32> [[TMP3]], i32 [[CONV31_I]], i32 3
+; CHECK-NEXT:    [[TMP5:%.*]] = insertelement <8 x i32> [[TMP4]], i32 [[CONV31_I]], i32 4
+; CHECK-NEXT:    [[TMP6:%.*]] = insertelement <8 x i32> [[TMP5]], i32 [[CONV31_I]], i32 5
+; CHECK-NEXT:    [[TMP7:%.*]] = insertelement <8 x i32> [[TMP6]], i32 [[CONV31_I]], i32 6
----------------
spatel wrote:
> Did SLP fail to recognize that this is a splat shuffle? I would have expected it to produce splat IR:
> (shuffle (insert X, 0), zeroinitializer)
> ...instead of a chain of inserts.
No. It did not fail to recognize a splat. As I see from code single element is not shuffled deliberately:

Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
...
    // Do not shuffle single element or if number of unique values is not power
    // of 2.
    if (UniqueValues.size() == VL.size() || UniqueValues.size() <= 1 ||
        !llvm::isPowerOf2_32(UniqueValues.size()))
      ReuseShuffleIndicies.clear();
...



Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D76956/new/

https://reviews.llvm.org/D76956