Add a FutureFeatures.md section discussing full IEEE-754 conformance. #171
Conversation
|
|
||
| ## Full IEEE-754 conformance | ||
|
|
||
| IEEE-754 NaN [bit pattern] propagation is presently permitted but not |
There was a problem hiding this comment.
What are the square brackets for?
There was a problem hiding this comment.
I may have been thinking of adding a link, but didn't. I'll remove them.
|
Link to #148? |
sunfishcode
force-pushed
the
full-ieee754
branch
2 times, most recently
from
8f4009e to
f8ea8d5
Compare
|
It's possible we'll resolve #148 by having a mode where subnormals are supported, if nothing else, in which case we won't need a link. Let's see where that discussion goes first. |
|
lgtm, but I imagine @kripken may want to take a look. |
|
|
||
| Debugging techniques are also important, but they don't necessarily need to be | ||
| in the spec itself. Implementations are welcome (and encouraged) to support | ||
| execution modes with alternate default rounding modes, or alternate execution |
There was a problem hiding this comment.
Perhaps specify "non-standard execution modes enabled only from developer tools"?
|
lgtm |
sunfishcode
force-pushed
the
full-ieee754
branch
2 times, most recently
from
fad4535 to
dc5ff3e
Compare
| in the spec itself. Implementations are welcome (and encouraged) to support | ||
| non-standard execution modes, enabled only from developer tools, such as modes | ||
| with alternate rounding, or evaluation of floating point expressions at greater | ||
| precision, to support [techniques for detecting numerical instability] |
There was a problem hiding this comment.
Maybe this is too new, but Herbie sounds interesting too!
There was a problem hiding this comment.
These paragraphs are addressing important use cases on popular hardware platforms that typically involve global (per-thread) floating point control/status registers, which is something that we may wish to avoid. The previous paragraph covers local override use cases, which covers most of the use cases I'm aware of. This paragraph explains how we can address the main use case that actually wants a global mode, by observing that debug techniques don't necessarily need to be specified in the spec. And I tossed in the idea of greater precision evaluation because it's another technique mentioned in the same paper which doesn't end up being completely futile and which would be straight-forward to implement.
Herbie looks neat, and correct me if I'm wrong, but it doesn't appear to require any special platform support, so there'd be no reason why users can't use it on WebAssembly implementations.
sunfishcode
force-pushed
the
full-ieee754
branch
from
dc5ff3e to
6ab650b
Compare
|
@jfbastien I added a mention that subnormals are still under discussion to the main AstSemantics, with a link to #148. |
And provide a link to background reading about numerical debugging techniques.
sunfishcode
force-pushed
the
full-ieee754
branch
from
6ab650b to
b771579
Compare
|
Taking the liberty of merging based on LGTMs above, even though I added a few extra minor sentences. I don't expect they're controversial. |
Add a FutureFeatures.md section discussing full IEEE-754 conformance.
Some people I spoke to had concerns about the IEEE-754 section. This patch adds a FutureFeatures.md section discussing a possible path to full conformance in the future, to help people understand where we are. As with all features in FutureFeatures.md, we'd need further discussion before we'd actually standardize anything here.