C++: Fix global-variable flow for array types #14822
Conversation
MathiasVP
force-pushed
the
fix-global-variable-flow-for-arrays
branch
from
22f25e8 to
8dc6c03
Compare
MathiasVP
force-pushed
the
fix-global-variable-flow-for-arrays
branch
from
8dc6c03 to
dcba8e5
Compare
| | test.cpp:864:44:864:58 | {...} | Node should have one enclosing callable but has 0. | | ||
| | test.cpp:864:47:864:54 | call to source | Node should have one enclosing callable but has 0. | | ||
| | test.cpp:872:46:872:51 | call to source | Node should have one enclosing callable but has 0. | | ||
| | test.cpp:872:53:872:56 | 1 | Node should have one enclosing callable but has 0. | | ||
| uniqueCallEnclosingCallable | ||
| | test.cpp:864:47:864:54 | call to source | Call should have one enclosing callable but has 0. | | ||
| | test.cpp:872:46:872:51 | call to source | Call should have one enclosing callable but has 0. | |
There was a problem hiding this comment.
Why are these suddenly showing up, while we don't have these for other globals for which we already had tests?
There was a problem hiding this comment.
No idea. Since this is AST dataflow I didn't really bother to investigate this.
There was a problem hiding this comment.
Ah, this is for AST dataflow, I missed that. Should have had a better look at the file names first. In that case, it seems safe to ignore.
There was a problem hiding this comment.
Note that line 864 is:
static const int global_array_dynamic[] = { ::source() };and line 872 is:
static const int* global_pointer_dynamic = source(true);So it's very possible that the code for these syntactic constructs simply was never generalized to handle these things appearing in global scope in AST dataflow (since AST dataflow never actually implementer proper support for global variables).
In #14736 we fixed a bunch of problems related to tracking indirections of global variables. One thing we forgot to handle was the subtle semantic difference between pointer types and array types. This PR fixes this by special-casing
ArrayTypesimilar to what we've been doing for SSA for a long time.The first commit adds a lot of tests. I basically threw in any combination of direct/indirect and static local/global tests I could think of. The second PR fixes the tests that were failing by doing the above handlign of array types.
DCA looks uneventful. The first run showed no changes in performance/results. The second run (which should have had no performance-relevant differences compared to the first one) shows a significant speedup for a couple of projects that I can't explain. Luckily, it also shows no result changes.