[AArch64][ARM] experimental.vector.reduce without FMF asserts

[nikic]

Bugzilla Link	44600
Resolution	FIXED
Resolved on	Feb 05, 2020 04:58
Version	10.0
OS	Linux
Blocks	#43900
CC	@alex,@Arnaud-de-Grandmaison-ARM,@zmodem,@jdm,@smithp35
Fixed by commit(s)	70d345e, 1cc4f8d, 362d00e

Extended Description

define float @​test_fadd_ordered(<3 x float> %v) {
%res = call float @​llvm.experimental.vector.reduce.v2.fadd.f32.v3f32(float 0.0, <3 x float> %v)
ret float %res
}

define float @​test_fmax_unordered_nan(<4 x float> %v) {
%res = call reassoc float @​llvm.experimental.vector.reduce.fmax.v4f32(<4 x float> %v)
ret float %res
}

declare float @​llvm.experimental.vector.reduce.v2.fadd.f32.v3f32(float %start_value, <3 x float> %a)
declare float @​llvm.experimental.vector.reduce.fmax.v4f32(<4 x float> %a)

The first case asserts on both AArch64 and ARM because there is no support for legalizing VECREDUCE_STRICT. The second one asserts on AArch64 because it requires fmax reductions to be nnan.

[nikic]

Candidate patch: https://reviews.llvm.org/D73135

[nikic]

Fixed by https://reviews.llvm.org/rG70d345e687caba4ac1f95655c6924dfa91e0083f. Keeping issue open to track LLVM 10 backport.

[nikic]

Additional related fixes were applied in https://reviews.llvm.org/D73854 and https://reviews.llvm.org/D73908.

[zmodem]

Candidate patch: https://reviews.llvm.org/D73135

Cherry-picked to 10.x in 0b8a540

Additional related fixes were applied in https://reviews.llvm.org/D73854

picked in 99c6a4e

and
https://reviews.llvm.org/D73908.

picked in 8195a96.