Steven Johnson
Steven Johnson
> To consider: you want to inspect these symbolic bounds from python, which may make a C++ template approach questionable. It might be helpful to offer a strawman representation that...
(Additionally, it appears that when targeting Windows and using emulated, the relevant glue code for f16f32 conversions is referenced by LLVM but not emitted by the Linker, leading to confusing...
Yeah, I didn't think about that part. Full f16 requires AVX10 or similar, right?
I believe so
Oh wait, this test is weird, stand by
Testing this in Google, I get many many failures of the form `Error: Buffer foo may be accessed in an unbounded way in dimension 0`
I'll have to figure out how many pipelines it is. (there are *lots* of failures, but maybe it's just a handful of shared pipelines.) What is the likely failure pattern...
> Access to an input named "foo" using a computed uint32 index. This is a bit of a guess but I think the needed patch is something like: The first...
Running the failure case with HL_DEBUG_CODEGEN=4 reveals something possibly interesting: just before the failure, I see ``` simplify let t749 = (int16)1023 in... new let t749.s = (undefined) in ......
The immediate failure looks like this: ``` Input to CSE ((uint16)2048 == (uint16)2048) After removing lets: ((uint16)2048 == (uint16)2048) Include: ((uint16)2048 == (uint16)2048); should extract: 1 Include: (uint16)2048; should extract:...