[DeadFunctionElimination] Pass on keypaths to foreign DeclRef properties #12625
Conversation
|
@swift-ci please test |
|
Can you add a sil test case for DFE? |
|
@gottesmm I think I did? you'd prefer one added to dead_function_elimination.swift as well? (Note the extra -O execution of the test there and the enclosing function being made-public: means DFE kicks in and asserts with failure if the fix isn't present) |
|
I meant one with optimization enabled. The tests you added AFAIKT are SILGen tests that do not actually test the optimizer.
… On Oct 25, 2017, at 5:58 PM, Graydon Hoare ***@***.*** ***@***.***>> wrote:
@gottesmm <https://github.com/gottesmm> I think I did? you'd prefer one added to dead_function_elimination.swift as well?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub <#12625 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAee39gpOaCA-q5WEAkztx9bH-42b7PDks5sv9ksgaJpZM4QG3PL>.
|
|
@gottesmm I added a secondary copy of the -emit-ir RUN line with -O turned on. It does run the optimizer as far as I know (it crashes with the motivating assert in DFE when I run the test but undo the fix). Again, I can add another if this is insufficiently obvious, but I think it does test the thing it's trying to test. |
…ies. These show up specifically when someone forms a keypath to an ObjC property in a category. They're no-ops from the perspective of DFE but the code has to intentionally skip the case if it wants to avoid the subsequent llvm_unreachable. rdar://34913689
graydon
force-pushed
the
rdar-34913689-dead-function-elim-vs-foreign-declref-keypaths
branch
from
94143b5 to
3ef686b
Compare
|
@gottesmm split test out into its own file in SILOptimizer. |
|
@swift-ci please test |
|
Build failed |
|
Build failed |
|
@graydon Thanks! |
graydon
deleted the
rdar-34913689-dead-function-elim-vs-foreign-declref-keypaths
branch
These show up specifically when someone forms a keypath to an ObjC
property in a category. They're no-ops from the perspective of
DFE but the code has to intentionally skip the case if it wants to
avoid the subsequent llvm_unreachable (in an assertion build).
rdar://34913689