Add extraction gym's ILP extractor to test in nightly #738
Conversation
oflatt
force-pushed
the
oflatt-ilp-extraction
branch
from
a0a7f74 to
96ca9c1
Compare
| @@ -4,45 +4,32 @@ expression: visualization.result | |||
| --- | |||
| # ARGS: 0 | |||
| @main(v0: int) { | |||
There was a problem hiding this comment.
This one looks suspicious. It seems the old one does loop unrolling and the new one does not.
| @@ -14,18 +14,15 @@ expression: visualization.result | |||
| v7_: int = id c3_; | |||
| v8_: int = id c4_; | |||
| .b9_: | |||
There was a problem hiding this comment.
Same here. The old one does loop unrolling (step by 4) and the new one step by 1.
The old one also does not do the optimization (x+2)+2 = x + 4
| @@ -14,16 +14,15 @@ expression: visualization.result | |||
| v7_: int = id c3_; | |||
| v8_: int = id c4_; | |||
| .b9_: | |||
There was a problem hiding this comment.
same as the sequential version
| @@ -16,451 +16,25 @@ expression: visualization.result | |||
| v8_: bool = eq c3_ v0; | |||
| br v8_ .b9_ .b10_; | |||
| .b9_: | |||
| v11_: bool = eq c1_ c1_; | |||
| v11_: int = call @fac c1_; | |||
There was a problem hiding this comment.
This PR seems to undo function inlining?
| @@ -4,7 +4,7 @@ expression: visualization.result | |||
| --- | |||
| # ARGS: | |||
There was a problem hiding this comment.
this seems to increase the number of inlining
| @@ -4,55 +4,32 @@ expression: visualization.result | |||
| --- | |||
| # ARGS: 30 10 | |||
| @main(v0: int, v1: int) { | |||
|
I am leaning towards keeping this PR as a separate branch given (1) the snapshot has some big changes, (2) the extractor diff is non-trivial, (3) ILP figures we want are one-off, and (4) the deadline's in five days. What do you think? |
|
This PR fixes Get nodes, as well as improving extraction performance by pruning the e-graph |
|
Ok let's keep in separate branch for now unless someone needs Get to work |
Draft for now until successful nightly