bjorn3 commented on issue #1906:
I believe the
fmin_pseudo
andfmax_pseudo
instructions do this, but they are currently unimplemented for scalar types.
bjorn3 commented on issue #1906:
Nope, they are not the right instructions. They still propagate NaN.
akirilov-arm commented on issue #1906:
This proposal feels a bit redundant - is there any architecture that has a 1:1 mapping to this operation? Otherwise it is straightforward to pattern-match the equivalent operation sequences if the default lowering is not good enough. E.g. the minimum operation for 32-bit inputs:
v2 = fcmp eq v0, v0 v3 = select.f32 v2, v0, v1 v4 = fmin_pseudo v3, v1
Last updated: Dec 23 2024 at 12:05 UTC