Gab has been debugging a failure that goes like this:
In testsuite/g++.dg/pph/x1functions.h:
struct type {
int mbr_decl_then_def(int);
int mbr_decl_inline(int i) { }
}
In testsuite/g++.dg/pph/x1functions.cc:
#include "x1functions.h"
int type::mbr_decl_then_def(int i)
{ return mbr_decl_inline( i ); }
We first compile x1functions.h into x1functions.pph. Then we compile
x1functions.cc with
$ gcc -fpph-hdr=x1functions x1functions.cc
This causes the #include directive to ignore x1functions.h and instead
load x1functions.pph. So, the parser really starts parsing at the
first token of x1functions.cc. The problem starts when it then tries
to lookup the IDENTIFIER_NODE mbr_decl_inline. It fails to realize
that the identifier is the FUNCTION_DECL type::mbr_decl_inline().
We have been tracing why the name lookup fails and it seems that this
is failing because mbr_decl_inline does not have a proper
IDENTIFIER_BINDINGS set. At first, we thought that maybe we were not
streaming this field out, but the parser removes and reattaches
different binding information using scope_chain->bindings.
In fact, scope_chain has a lot of state that I think needs to be saved
in each pph image. I am still not too sure how the parser uses the
various fields in scope_chain, but it seems to be keeping lists and
stacks as it goes in and out of classes and types during parsing.
Since we told it to skip the header file text, when it started parsing
the .cc file, none of the state that it would've collected while
parsing the .h file was present.
Am I on the right track? My impression is that for each header file
that we pre-parse we should save this state in scope_chain. The issue
then becomes, given more than one pph image, how should the different
scope_chains be merged?
Thanks. Diego.
