Use exactness to optimize GUFA cone depth (2nd alternative) #7582
Conversation
Exact references are known to point to exactly their heap type and not one of its subtypes. GUFA already analyzed exactness via the more general "cone" types that include an allowed subtyping depth. Exact types are equivalent to cone types of depth zero. Let GUFA take advantage of exactness information by normalizing cone depths to 0 whenever the cone type is exact.
tlively
changed the title
|
This is an alternative to #7538 based on the feedback from #7538 (comment). See commit cff042c for the diff from #7538 at time of writing. I don't feel strongly one way or another between the two approaches, but this one has to touch more code. It's unclear to me whether there is a functional difference between the two, although they pass the same tests. After having wrestled with this code for a while on multiple occasions, I'm sorely tempted to jump in and rewrite PossibleContents as a composition of lattices. |
src/ir/possible-contents.cpp
Outdated
| @@ -371,7 +357,18 @@ bool PossibleContents::isSubContents(const PossibleContents& a, | |||
| return false; | |||
| } | |||
|
|
|||
| WASM_UNREACHABLE("unhandled case of isSubContents"); | |||
| if (b.isGlobal()) { | |||
| // We've already ruled out anything but another global or cone type for a. | |||
There was a problem hiding this comment.
| // We've already ruled out anything but another global or cone type for a. | |
| // We've already ruled out anything but another global or (non-full) cone type for a. |
src/ir/possible-contents.h
Outdated
| @@ -128,7 +128,9 @@ class PossibleContents { | |||
|
|
|||
| // Internal convenience for creating a cone type of unbounded depth, i.e., the | |||
| // full cone of all subtypes for that type. | |||
src/ir/possible-contents.h
Outdated
| @@ -150,8 +152,8 @@ class PossibleContents { | |||
| } | |||
| // Helper for a cone with unbounded depth, i.e., the full cone of all subtypes | |||
| // for that type. | |||
I'd be curious to see what that looks like. Do you mean with the existing Lattice? |
Yes, exactly. The only missing piece preventing this from being a quick piece of work is a general |
Exact references are known to point to exactly their heap type and not
one of its subtypes. GUFA already analyzed exactness via the more
general "cone" types that include an allowed subtyping depth. Exact
types are equivalent to cone types of depth zero. Let GUFA take
advantage of exactness information by normalizing cone depths to 0
whenever the cone type is exact.