[IRGen] Mark inout ptrs noalias. #42503
Conversation
nate-chandler
force-pushed
the
rdar76540030
branch
from
dd502b3 to
9469330
Compare
|
@swift-ci please test |
There was a problem hiding this comment.
This should be applied to all indirect arguments except @inout_aliasable.
What's the purpose of doing this for "pointer types" only? What LLVM alias analysis issues need to be avoided, how does this check avoid them, and when can we stop hiding info from LLVM?
lib/IRGen/GenCall.cpp
Outdated
| // Thanks to exclusivity checking, it is not possible to alias inouts. | ||
| if (paramSILType.getASTType()->getAnyPointerElementType()) { | ||
| // To ward against issues with LLVM's alias analysis, for now, only add the | ||
| // attribute if it's a pointer being passed inout. | ||
| b.addAttribute(llvm::Attribute::NoAlias); | ||
| } | ||
| // Aliasing inouts can't be captured without doing unsafe stuff. |
There was a problem hiding this comment.
Indirect_InoutAliasable arguments do alias. So I think the code should check aliasable.
There was a problem hiding this comment.
Thanks! Fixed.
Thanks to exclusivity checking, inout arguments can be marked noalias. In the fullness of time, this should be applied in general. As an incremental step towards marking all such arguments noalias, start by marking inout arguments of pointer type noalias. Take this conservative step rather than always applying the attribute to ward against issues with LLVM's alias analysis. rdar://76540030
nate-chandler
force-pushed
the
rdar76540030
branch
from
9469330 to
d70d391
Compare
|
Deferring to @aschwaighofer to answer the questions about staging this in like this and starting with pointer types. |
|
@swift-ci please test |
There are reoccuring bugs in LLVM's alias analysis (as it is continuously "improved"). They cause hard to debug memory bugs and would impose a cost on those who have to debug them. This cost can be weight against (performance) benefits that we suspect to be getting. My suggestion was to introduce the use of At some later point we can remove the constraint and see where we end up with then. |
... |
|
When we have rewritten the code generation backend in Swift. |
Thanks to exclusivity checking, inout arguments can be marked noalias.
In the fullness of time, this should be applied in general.
As an incremental step towards marking all such arguments noalias, start
by marking inout arguments of pointer type noalias. Take this
conservative step rather than always applying the attribute to ward
against issues with LLVM's alias analysis.
rdar://76540030