-
-
Notifications
You must be signed in to change notification settings - Fork 5.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Bad PhiC node placement/usage corrupts gc-preserve use #34247
Comments
How coincidental: this is a duplicate of #34229 that I opened the day before.
|
I'm thinking the way to fix this would be to allow ssa uses across try/catch regions, coupled with an LLVM pass that either manually spills all state to the stack or outlines the body of the try/catch block. |
Hm, yeah, I think yuyichao was exploring something like that in #24573, but I don't know how feasible that is. In theory, the C standard says that LLVM is supposed to be OK with any variable that isn't modified in the try block being correct afterwards, and we seemed to be mostly OK for some years with assuming that. I still hope that someday we could move to using the lower-cost and more correct invoke instruction handling though. |
FWIW, bisected to cd92f60 (the upgrade to LLVM 8.0.1). |
This fixes #34247 by changing the way gc preserve is lowered. Instead of lowering it in a macro, lower it in the frontend. This allows us to use an SSA value directly for the return token of the gc begin expression. This bypasses the slot-renaming pass of the compiler, thus preventing the compiler from trying to save and restore the token. Of course, this kind of code would generally not be legal (because it uses an SSA value outside of the regular domination relation), but since this is a julia restriction, not an LLVM restriction, we can simply exempt gc_begin tokens from this particular validation. This works fine at the LLVM level also, because it doesn't have this particular restriction. It also doesn't have the same correctness problems as doing the same for non-token values, as the tokens get lowered away by the try/catch lowering before reaching the LLVM backend.
This fixes #34247 by changing the way gc preserve is lowered. Instead of lowering it in a macro, lower it in the frontend. This allows us to use an SSA value directly for the return token of the gc begin expression. This bypasses the slot-renaming pass of the compiler, thus preventing the compiler from trying to save and restore the token. Of course, this kind of code would generally not be legal (because it uses an SSA value outside of the regular domination relation), but since this is a julia restriction, not an LLVM restriction, we can simply exempt gc_begin tokens from this particular validation. This works fine at the LLVM level also, because it doesn't have this particular restriction. It also doesn't have the same correctness problems as doing the same for non-token values, as the tokens get lowered away by the try/catch lowering before reaching the LLVM backend.
This fixes #34247 by changing the way gc preserve is lowered. Instead of lowering it in a macro, lower it in the frontend. This allows us to use an SSA value directly for the return token of the gc begin expression. This bypasses the slot-renaming pass of the compiler, thus preventing the compiler from trying to save and restore the token. Of course, this kind of code would generally not be legal (because it uses an SSA value outside of the regular domination relation), but since this is a julia restriction, not an LLVM restriction, we can simply exempt gc_begin tokens from this particular validation. This works fine at the LLVM level also, because it doesn't have this particular restriction. It also doesn't have the same correctness problems as doing the same for non-token values, as the tokens get lowered away by the try/catch lowering before reaching the LLVM backend. (cherry picked from commit 07a16d6)
This fixes #34247 by changing the way gc preserve is lowered. Instead of lowering it in a macro, lower it in the frontend. This allows us to use an SSA value directly for the return token of the gc begin expression. This bypasses the slot-renaming pass of the compiler, thus preventing the compiler from trying to save and restore the token. Of course, this kind of code would generally not be legal (because it uses an SSA value outside of the regular domination relation), but since this is a julia restriction, not an LLVM restriction, we can simply exempt gc_begin tokens from this particular validation. This works fine at the LLVM level also, because it doesn't have this particular restriction. It also doesn't have the same correctness problems as doing the same for non-token values, as the tokens get lowered away by the try/catch lowering before reaching the LLVM backend. (cherry picked from commit 07a16d6)
This fixes #34247 by changing the way gc preserve is lowered. Instead of lowering it in a macro, lower it in the frontend. This allows us to use an SSA value directly for the return token of the gc begin expression. This bypasses the slot-renaming pass of the compiler, thus preventing the compiler from trying to save and restore the token. Of course, this kind of code would generally not be legal (because it uses an SSA value outside of the regular domination relation), but since this is a julia restriction, not an LLVM restriction, we can simply exempt gc_begin tokens from this particular validation. This works fine at the LLVM level also, because it doesn't have this particular restriction. It also doesn't have the same correctness problems as doing the same for non-token values, as the tokens get lowered away by the try/catch lowering before reaching the LLVM backend. (cherry picked from commit 07a16d6)
This fixes #34247 by changing the way gc preserve is lowered. Instead of lowering it in a macro, lower it in the frontend. This allows us to use an SSA value directly for the return token of the gc begin expression. This bypasses the slot-renaming pass of the compiler, thus preventing the compiler from trying to save and restore the token. Of course, this kind of code would generally not be legal (because it uses an SSA value outside of the regular domination relation), but since this is a julia restriction, not an LLVM restriction, we can simply exempt gc_begin tokens from this particular validation. This works fine at the LLVM level also, because it doesn't have this particular restriction. It also doesn't have the same correctness problems as doing the same for non-token values, as the tokens get lowered away by the try/catch lowering before reaching the LLVM backend.
This fixes JuliaLang#34247 by changing the way gc preserve is lowered. Instead of lowering it in a macro, lower it in the frontend. This allows us to use an SSA value directly for the return token of the gc begin expression. This bypasses the slot-renaming pass of the compiler, thus preventing the compiler from trying to save and restore the token. Of course, this kind of code would generally not be legal (because it uses an SSA value outside of the regular domination relation), but since this is a julia restriction, not an LLVM restriction, we can simply exempt gc_begin tokens from this particular validation. This works fine at the LLVM level also, because it doesn't have this particular restriction. It also doesn't have the same correctness problems as doing the same for non-token values, as the tokens get lowered away by the try/catch lowering before reaching the LLVM backend. (cherry picked from commit 07a16d6)
Report: https://github.com/maleadt/BasePkgEvalReports/blob/0bb12607c245153c8e7215788e580d43d7d2d51a/pkgeval-9bd498d_vs_2e6715c/logs/Strided/1.4.0-DEV-082f446f52.log
The text was updated successfully, but these errors were encountered: