[PATCH] D94946: [ARM] Expand vXi1 VSELECT's

[Dave Green via Phabricator via llvm-commits]

dmgreen added inline comments.


================
Comment at: llvm/test/Analysis/CostModel/ARM/arith-overflow.ll:387
 ; MVE-RECIP-NEXT:  Cost Model: Found an estimated cost of 74 for instruction: %V2I64 = call { <2 x i64>, <2 x i1> } @llvm.ssub.with.overflow.v2i64(<2 x i64> undef, <2 x i64> undef)
-; MVE-RECIP-NEXT:  Cost Model: Found an estimated cost of 166 for instruction: %V4I64 = call { <4 x i64>, <4 x i1> } @llvm.ssub.with.overflow.v4i64(<4 x i64> undef, <4 x i64> undef)
-; MVE-RECIP-NEXT:  Cost Model: Found an estimated cost of 582 for instruction: %V8I64 = call { <8 x i64>, <8 x i1> } @llvm.ssub.with.overflow.v8i64(<8 x i64> undef, <8 x i64> undef)
+; MVE-RECIP-NEXT:  Cost Model: Found an estimated cost of 242 for instruction: %V4I64 = call { <4 x i64>, <4 x i1> } @llvm.ssub.with.overflow.v4i64(<4 x i64> undef, <4 x i64> undef)
+; MVE-RECIP-NEXT:  Cost Model: Found an estimated cost of 866 for instruction: %V8I64 = call { <8 x i64>, <8 x i1> } @llvm.ssub.with.overflow.v8i64(<8 x i64> undef, <8 x i64> undef)
----------------
SjoerdMeijer wrote:
> More of a curiosity, how was this lowered before? Given the added codegen tests, I guess this is not only a cost-model tweak? And some of these costs are extremely high. Do they make sense?
Yep. The costs are based on whether the operation is legal and add.with.overflow checks the cost of selects of predicates (v4i1). There is no legal instruction for them for MVE, so a high cost is probably fine.

D94958 updates the similar costs for sadd.sat and friends, which also have costs based in sadd.with.overflow by default.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D94946/new/

https://reviews.llvm.org/D94946