[wasm] Fix for getelema1 crash in specific scenarios #87246
Conversation
If we ever see garbage in the zero page disable it permanently Log an error if garbage appears in the zero page after startup
|
Tagging subscribers to 'arch-wasm': @lewing Issue DetailsThere was a latent bug in getelema1 that for some reason triggers only in BenchmarkDotNet, this PR should fix it. This is what I get for using 'const' like eslint constantly reminds me to. I also improved on the zero page reserved checks so that we permanently disable use of the zero page if we ever see nonzero data there, and if nonzero data appears there after startup we log an error (if this happened your app is probably going to crash.)
|
ghost
assigned 
|
This is apparently still not enough to fix the two BDN benchmarks that broke. |
Fix count not being initialized for cpblk in some cases
|
There was a second bug (in CPBLK) hidden until recently where under certain circumstances the count local used to control the number of bytes copied wasn't getting initialized, and due to local reuse in traces it could have an arbitrarily large value in it. This was hard to find at first so I ended up just refactoring how traces use locals, renaming some of the scratch locals to have more semantically meaningful names and adding more of them to make it less likely that uses of them will collide. I also changed the memmove and memcpy unrollers to use dedicated scratch locals so it's safe to call them without worrying about them trampling on your locals. As a result the "math_..." locals should mostly be used for math operations now, and for things like index computations and memory accesses "dest_ptr", "src_ptr", "index" and "count" should be getting used instead. I probably need to add a system for allocating locals on demand... I don't know if adding these additional locals will make trace stack frames bigger or cause traces to get slower. |
|
Determined that the old leave-on-stack behavior in the two places I changed was actually correct, so I changed it back. I misunderstood how WebAssembly blocks interact with the stack. |
|
Previous commit passed on linux (didn't fail this time, just never ran), and the windows lanes passed in both AOT and interp modes, so merging. |
There was a latent bug in getelema1 that for some reason triggers only in BenchmarkDotNet, this PR should fix it. This is what I get for using 'const' like eslint constantly reminds me to.
I also improved on the zero page reserved checks so that we permanently disable use of the zero page if we ever see nonzero data there, and if nonzero data appears there after startup we log an error (if this happened your app is probably going to crash.)