http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50889
Bug #: 50889
Summary: PLUGIN_FINISH_UNIT arrives a bit to late for really
high-level (AST-based) in-memory code
transformation/instrumentation
Classification: Unclassified
Product: gcc
Version: 4.7.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: plugins
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: krik_evryw...@hotmail.com
The PLUGIN_FINISH_UNIT callback is called in toplev.c at line 668 (in version
4.6.2 of gcc), at the very end of the compile_file () function.
Unfortunately, it seems to mean the heavy work of lowering the AST is already
done by the time it reaches this callback, and to the point to having almost
finished emitting the assembly code, as suggested by the comment and statement
just after the callback call:
/* This must be at the end. Some target ports emit end of file directives
into the assembly file here, and hence we can not output anything to the
assembly file after this point. */
targetm.asm_out.file_end ();
At least that's the way I understand it, I'm very new to gcc. That's way too
late to do some high-level transformations on the AST before it gets lowered
and, thus, the "new" (freshly in-memory modified) code gets included into the
lower phases of the process.
I've come to this basic conclusion after having checked the work the
interesting functions called by the parse_file () hook do, and having checked
against the way GCC-XML has modified gcc 4.2.1 to call its own hook to browse
through the high-level AST generated when gcc has finished parsing the source
file.
I would like to request adding a simple additional callback type, something
like "PLUGIN_PREFINISH_UNIT", "PLUGIN_PRE_FINISH_UNIT",
"PLUGIN_FINISH_PARSE_UNIT" or whatever name suits, which would be called at
line 585 (still talking against version 4.6.2), just before the "if
(flag_syntax_only || flag_wpa)" line, that's the best place I can spot (and the
one looking the more logical to me).
It's just a matter of adding a few lines here and there (code and docs). I
could appropriately alter a checked-out revision of version 4.7.0, or whatever
is called the "trunk" for now, if it's ok and it can help this gets
implemented.
As to the reason why, for example, I would like to automatically add a
companion variable to some variables based on their type, through a plugin.
Which is not safe to do with some naive look-up, and far from trivial to do
otherwise because of redoing what gcc already does very much that I could ever
do, that's parsing the source code. Doing this with a plugin would heavily
simplify and help the analysis, and thus the insertion at proper place in the
AST.
For completeness, I should mention it seems to me the bug 44968 ("structs saved
(in a vector) during PLUGIN_FINISH_TYPE are mangled by the time of
PLUGIN_FINISH_UNIT") is a bit related to this request, as the author seems to
have expected the AST would be the raw one, not an already lowered one.
