[PATCH] D135991: [AArch64] Fix cost model for `udiv` instruction when one of the operands is a uniform constant

[Dave Green via Phabricator via llvm-commits]

dmgreen added inline comments.


================
Comment at: llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp:2095
         // into scalar divisions of each pair of elements.
         Cost += getArithmeticInstrCost(Instruction::ExtractElement, Ty,
                                        CostKind, Op1Info, Op2Info);
----------------
Using getArithmeticInstrCost here seems very odd.

I would expect the base cost to be `N*getVectorInstrCost(Instruction::ExtractElement) + N* getVectorInstrCost(Instruction::InsertElement) + N*getArithmeticInstrCost(Div)`. Sometimes we use a cheaper cross-register-bank copy cost than getVectorInstrCost would give.

I would expect the getVectorInstrCost's to become removed if one of the operations was uniform or constant. And maybe the Cost += Cost shouldn't be needed if the vector instructions have correct costs. All the code here seems to be trying to give a very high cost for vector divides, as they will just be scalarized.


================
Comment at: llvm/test/Analysis/CostModel/AArch64/div.ll:181
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %I64 = sdiv i64 undef, 7
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V2i64 = sdiv <2 x i64> undef, <i64 7, i64 7>
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 96 for instruction: %V4i64 = sdiv <4 x i64> undef, <i64 7, i64 7, i64 7, i64 7>
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 192 for instruction: %V8i64 = sdiv <8 x i64> undef, <i64 7, i64 7, i64 7, i64 7, i64 7, i64 7, i64 7, i64 7>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2i64 = sdiv <2 x i64> undef, <i64 7, i64 7>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4i64 = sdiv <4 x i64> undef, <i64 7, i64 7, i64 7, i64 7>
----------------
It seems odd that doing 2 divides, plus some cross-register-bank copies, would be cheaper than doing 1


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135991/new/

https://reviews.llvm.org/D135991