ArtBlnd commented on issue #4920:
Kinda struggling testing with NaNs. I think I can do in 3 days.
ArtBlnd commented on issue #4920:
Ok, I've try testing this many times but I think its better testing with i32 to f32 bitcasts. to do that. we need to test
bitcast.i32
andbitcast.f32
first before we go.Thoughts?
ArtBlnd edited a comment on issue #4920:
Ok, I've try testing this many times but I think its better testing with i32 to f32 bitcasts. to do that. we need to implement and test
bitcast.i32
andbitcast.f32
first before we go.Thoughts?
ArtBlnd edited a comment on issue #4920:
Ok, I've try testing this many times but I think its better testing with i32 to f32 bitcasts. to do that. we need to implement and test
bitcast.ixx
andbitcast.fxx
first before we go.Thoughts?
ArtBlnd edited a comment on issue #4920:
Ok, I've try testing this many times but I think its better testing with i32 to f32 bitcasts. to do that. we need to implement and test
bitcast.uxx
andbitcast.fxx
first before we go.Thoughts?
ArtBlnd edited a comment on issue #4920:
Ok, I've try testing this many times but I think its better testing with i32 to f32 bitcasts. to do that. we need to implement and test
bitcast.ixx
andbitcast.fxx
first before we go.Thoughts?
ArtBlnd edited a comment on issue #4920:
Ok, I've try testing this many times but I think its better testing with i32 to f32 bitcasts. to do that. we need to implement and test
bitcast.ixx
andbitcast.fxx
first before we go.also pretty much sure that fp bit ops should work like integer bit ops.
Thoughts?
afonso360 commented on issue #4920:
Ok, I've try testing this many times but I think its better testing with i32 to f32 bitcasts. to do that. we need to implement and test bitcast.ixx and bitcast.fxx first before we go.
I don't mind the tests with bitcast -> op -> bitcast. But what issues are you seeing with just using floats in and out?
also pretty much sure that fp bit ops should work like integer bit ops.
Agreed! I don't think any special input should cause us to deviate from that.
afonso360 edited a comment on issue #4920:
Ok, I've try testing this many times but I think its better testing with i32 to f32 bitcasts. to do that. we need to implement and test bitcast.ixx and bitcast.fxx first before we go.
I don't mind the tests with bitcast -> op -> bitcast (although I would slightly prefer them without). But what issues are you seeing with just using floats in and out?
also pretty much sure that fp bit ops should work like integer bit ops.
Agreed! I don't think any special input should cause us to deviate from that.
ArtBlnd commented on issue #4920:
I don't mind the tests with bitcast -> op -> bitcast (although I would slightly prefer them without). But what issues are you seeing with just using floats in and out?
We can't check sign value of float in NaN payload. I think its better using
bitcast
to test it.
ArtBlnd edited a comment on issue #4920:
I don't mind the tests with bitcast -> op -> bitcast (although I would slightly prefer them without). But what issues are you seeing with just using floats in and out?
We can't check sign flag of float in NaN payload. I think its better using
bitcast
to test it.
ArtBlnd edited a comment on issue #4920:
I don't mind the tests with bitcast -> op -> bitcast (although I would slightly prefer them without). But what issues are you seeing with just using floats in and out?
We can't check sign flag of float in NaN payload. I think its better using
bitcast
to test it. because its bitcast. the sign flag should not optimized away.
ArtBlnd edited a comment on issue #4920:
I don't mind the tests with bitcast -> op -> bitcast (although I would slightly prefer them without). But what issues are you seeing with just using floats in and out?
We can't check sign flag of float in NaN payload. I think its better using
bitcast
to test it. because its bitcast. the sign flag should not optimized away. (while NaN -> Integer casting?)
ArtBlnd edited a comment on issue #4920:
I don't mind the tests with bitcast -> op -> bitcast (although I would slightly prefer them without). But what issues are you seeing with just using floats in and out?
We can't check sign flag of float in NaN payload. I think its better using
bitcast
to test it. because its bitcast. the sign flag should not optimized away. (while NaN -> Integer casting?)thats why I added
i32 -> f32 -> i32
test on cast implementation.
afonso360 commented on issue #4920:
In our runtest suite we do bit comparisions of float output, so you should be able to pass in a
sNaN
and compare asNaN
output.We do that in other tests, see for example
fcopysign
which is also a "bitwise" op:
https://github.com/bytecodealliance/wasmtime/blob/a434f43d2276dd30dab26cfeaa4103f337557f17/cranelift/filetests/filetests/runtests/fcopysign.clif#L39-L55
afonso360 edited a comment on issue #4920:
In our runtest suite we do bit comparisions of float output, so you should be able to pass in a
sNaN
and compare asNaN
output.We do that in other tests, see for example
fcopysign
which is also a "bitwise" op:
https://github.com/bytecodealliance/wasmtime/blob/a434f43d2276dd30dab26cfeaa4103f337557f17/cranelift/filetests/filetests/runtests/fcopysign.clif#L39-L55Edit:
I think its better using bitcast to test it. because its bitcast. the sign flag should not optimized away. (while NaN -> Integer casting?)
Although yes, we should also check that for
bitcasts
, they also shouldn't silence any signal bits.
Last updated: Dec 23 2024 at 12:05 UTC