[PATCH] D107651: [AArch64][SVE] Teach cost model that masked loads/stores are cheap

[David Sherwood via Phabricator via llvm-commits]

david-arm added inline comments.


================
Comment at: llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp:1505
                                       TTI::TargetCostKind CostKind) {
-  if (!isa<ScalableVectorType>(Src))
-    return BaseT::getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
----------------
I don't think we should be removing this because it's needed when we don't know the SVE vector length, in which case we will have to scalarise the masked ops. I think you can add an extra check here to see if we're using SVE for fixed width vectors.


================
Comment at: llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp:1516
 
-  return LT.first * 2;
+  return isa<ScalableVectorType>(Src) ? LT.first * 2 : LT.first * 2 + LT.first;
 }
----------------
bsmith wrote:
> What's the rationale behind this cost for fixed types?
Is this due to the extra predicate we have to create? If so, that's also true for SVE since we'll need a ptrue and it's highly likely to get reused anyway or hoisted out of a loop.


================
Comment at: llvm/test/Analysis/CostModel/AArch64/masked_ldst.ll:6
 ; CHECK-LABEL: 'fixed'
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %v2i8 = call <2 x i8> @llvm.masked.load.v2i8.p0v2i8(<2 x i8>* undef, i32 8, <2 x i1> undef, <2 x i8> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 25 for instruction: %v4i8 = call <4 x i8> @llvm.masked.load.v4i8.p0v4i8(<4 x i8>* undef, i32 8, <4 x i1> undef, <4 x i8> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 53 for instruction: %v8i8 = call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* undef, i32 8, <8 x i1> undef, <8 x i8> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 109 for instruction: %v16i8 = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* undef, i32 8, <16 x i1> undef, <16 x i8> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %v2i16 = call <2 x i16> @llvm.masked.load.v2i16.p0v2i16(<2 x i16>* undef, i32 8, <2 x i1> undef, <2 x i16> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 25 for instruction: %v4i16 = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* undef, i32 8, <4 x i1> undef, <4 x i16> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 53 for instruction: %v8i16 = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* undef, i32 8, <8 x i1> undef, <8 x i16> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %v2i32 = call <2 x i32> @llvm.masked.load.v2i32.p0v2i32(<2 x i32>* undef, i32 8, <2 x i1> undef, <2 x i32> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 25 for instruction: %v4i32 = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* undef, i32 8, <4 x i1> undef, <4 x i32> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %v2i64 = call <2 x i64> @llvm.masked.load.v2i64.p0v2i64(<2 x i64>* undef, i32 8, <2 x i1> undef, <2 x i64> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %v2f16 = call <2 x half> @llvm.masked.load.v2f16.p0v2f16(<2 x half>* undef, i32 8, <2 x i1> undef, <2 x half> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 25 for instruction: %v4f16 = call <4 x half> @llvm.masked.load.v4f16.p0v4f16(<4 x half>* undef, i32 8, <4 x i1> undef, <4 x half> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 53 for instruction: %v8f16 = call <8 x half> @llvm.masked.load.v8f16.p0v8f16(<8 x half>* undef, i32 8, <8 x i1> undef, <8 x half> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %v2f32 = call <2 x float> @llvm.masked.load.v2f32.p0v2f32(<2 x float>* undef, i32 8, <2 x i1> undef, <2 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 25 for instruction: %v4f32 = call <4 x float> @llvm.masked.load.v4f32.p0v4f32(<4 x float>* undef, i32 8, <4 x i1> undef, <4 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %v2f64 = call <2 x double> @llvm.masked.load.v2f64.p0v2f64(<2 x double>* undef, i32 8, <2 x i1> undef, <2 x double> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 22 for instruction: %v4i64 = call <4 x i64> @llvm.masked.load.v4i64.p0v4i64(<4 x i64>* undef, i32 8, <4 x i1> undef, <4 x i64> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 212 for instruction: %v32f16 = call <32 x half> @llvm.masked.load.v32f16.p0v32f16(<32 x half>* undef, i32 8, <32 x i1> undef, <32 x half> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
-;
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction:   %v1i1 = call <1 x i1> @llvm.masked.load.v1i1.p0v1i1(<1 x i1>* undef, i32 8, <1 x i1> undef, <1 x i1> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction:   %v2i1 = call <2 x i1> @llvm.masked.load.v2i1.p0v2i1(<2 x i1>* undef, i32 8, <2 x i1> undef, <2 x i1> undef)
----------------
Again, I think these changes look wrong - they should be high because we are going to scalarise the operations, since we haven't specified the SVE vector length so we can't use SVE or NEON.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107651/new/

https://reviews.llvm.org/D107651