On 3/24/23 18:25, Jakub Jelinek wrote:
On Fri, Mar 24, 2023 at 06:11:44PM -0400, Jason Merrill wrote:
When we touch this for COMPONENT_REF, shouldn't we then handle it as
unary given that the second operand is FIELD_DECL and third/fourth
will likely be NULL and even if not, aren't user expressions that should be
inspected?
So, instead of doing this do:
case COMPONENT_REF:
x = TREE_OPERAND (x, 0);
writer = 0;
goto restart;
?
Is clearing 'writer' what we want, since an access to COMPONENT_REF is an
access to (a subobject of) its op0?
I've just mindlessly copied the unary op case.
writer is set for pre/post increments and lhs of MODIFY_EXPR, and it is
true that VIEW_CONVERT_EXPR doesn't clear it, but e.g. ARRAY_REF clears it
for all operands.
For whatever reason leaving writer set led to lots of false positives,
so I've gone with your suggestion.
Currently with the fix it takes <1s while gcc12 takes ~80s.
Perfect.
PR c++/107163
gcc/c-family/ChangeLog:
* c-common.cc (verify_tree): Don't use sequenced handling
for COMPONENT_REF.
gcc/testsuite/ChangeLog:
* g++.dg/template/recurse5.C: New test.
LGTM, thanks. Maybe the testcase would be better as
warn/Wsequence-point-5.C, dunno.
Done.
Jason
