[PATCH] D100486: [COST]Improve cost model for shuffles in SLP.

Thu Jun 24 12:57:57 PDT 2021

ABataev added inline comments.

================
Comment at: llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:3633-3637
+    auto *FinalVecTy =
+        FixedVectorType::get(VecTy->getElementType(), ReuseShuffleNumbers);
     ReuseShuffleCost =
-        TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, VecTy,
-                            E->ReuseShuffleIndices);
+        TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc,
+                            FinalVecTy, E->ReuseShuffleIndices);
----------------
We incorrectly calculate the cost here (using the wrong vector type), with this fix and without this patch at least 4 X86 tests are not vectorized anymore.

================
Comment at: llvm/test/Transforms/SLPVectorizer/AArch64/PR38339.ll:6-16
+; CHECK-NEXT:    [[T2:%.*]] = extractelement <2 x i16> [[X:%.*]], i32 0
+; CHECK-NEXT:    [[T3:%.*]] = extractelement <2 x i16> [[X]], i32 1
 ; CHECK-NEXT:    [[PTR0:%.*]] = getelementptr inbounds [4 x i16], [4 x i16]* undef, i16 0, i16 0
 ; CHECK-NEXT:    [[PTR1:%.*]] = getelementptr inbounds [4 x i16], [4 x i16]* undef, i16 0, i16 1
 ; CHECK-NEXT:    [[PTR2:%.*]] = getelementptr inbounds [4 x i16], [4 x i16]* undef, i16 0, i16 2
 ; CHECK-NEXT:    [[PTR3:%.*]] = getelementptr inbounds [4 x i16], [4 x i16]* undef, i16 0, i16 3
+; CHECK-NEXT:    store i16 [[T2]], i16* [[A:%.*]], align 2
----------------
These regressions are caused by the incomplete cost model for AArch64, no cost for PermuteSingleSrc shuffle kind for VFxi16

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D100486/new/

https://reviews.llvm.org/D100486