This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH 0/9] RFC: selftests based on RTL dumps
- From: David Malcolm <dmalcolm at redhat dot com>
- To: gcc-patches at gcc dot gnu dot org
- Cc: David Malcolm <dmalcolm at redhat dot com>
- Date: Thu, 8 Sep 2016 20:30:44 -0400
- Subject: [PATCH 0/9] RFC: selftests based on RTL dumps
- Authentication-results: sourceware.org; auth=none
The current selftest code is adequate for testing individual
instructions, but most interesting passes have logic involving the
interaction of multiple instructions, or require a CFG and function
to be runnable. In theory we could write selftests by programatically
constructing a function and CFG "by hand" via API calls, but this is
awkward, tedious, and the resulting code is unnatural. Examples can
be seen in function-tests.c; that file constructs trivial 3-block
functions, and is pushing the limits of what I'd want to write
"by hand".
This patch kit provides a more natural way to write RTL selftests,
by providing a parser for RTL dumps (or fragments of RTL dumps). You
can copy and paste fragments of RTL dump into the source for a pass
and then have selftests that run part of the pass on that dump,
asserting that desired outcomes occur.
This is an updated and rewritten version of the RTL frontend work I
posted in May (c.f. "[PATCH 0/4] RFC: RTL frontend"
https://gcc.gnu.org/ml/gcc-patches/2016-05/msg00352.html).
In that patch kit, I introduced an rtl1 frontend, capable of parsing
RTL dumps, and running just one RTL pass on them, in the hope of
better testing individual RTL passes.
This patch kit takes a somewhat different approach: it provides
the infrastructure for parsing RTL dumps, but doesn't wire it up as
a frontend. Instead, it just provides a set of classes for use when
writing selftests. It would be possible to build out an "rtl1"
frontend as a followup to this kit.
The advantage of this approach is that it's possible to run and test
passes at finer granularity: for example, rather than being limited
to testing all of, say, pass "final", we can also write tests that
run just final_scan_insn on individual rtx_insn *, and verify that
the expected output is emitted. Tests can be written at various
different levels, testing how the components of a pass handle fragments
of an insn, how they handle entire insns, basic blocks, and so on up
to running a whole pass on a whole function.
The disadvantage is that changing a selftest requires recompilation
of cc1 (though that drawback affects selftests in general).
An overview of the kit follows; patches 6-9 are probably the most
interesting, as they show example of the kinds of selftest that can
be written using this approach:
* Patch 1 tidies up some global state in .md parsing, wrapping it up in
an rtx_reader class.
* Patches 2-4 add some selftest support code.
* Patch 5 introduces a function_reader subclass of patch 1's rtx_reader,
capable of parsing RTL function dumps (or fragments thereof), and
generalizes things so that rtx parsing can be run from the host, rather
than just at build time.
* Patch 6 uses the infrastructure to write a dataflow selftest. It's
a trivial example, but hopefully shows how more interesting selftests
could be written. (it's much easier to write a selftest if a similar
one already exists)
* Patch 7 does the same, but for combine.c.
* Patch 8 does the same, but for final.c, showing examples of both
assembling an entire function, and of assembling individual rtx_insn *
(to exercise the .md files and asm output code)
* Patch 9 does the same, but for cse.c. I attempted to create a
selftest that directly reproduces PR 71779, though I haven't yet
been able to to reproduce the issue (just load the insns and run cse
on them).
These tests are very target-specific and were developed mostly for
target==x86_64, and a little for target==aarch64.
I put clauses like this in the various test functions, which is a kludge:
/* Only run these tests for i386. */
#ifndef I386_OPTS_H
return;
#endif
Is there a better way to express this condition? (I guess I could
add a selftest::target_is_x86_p () predicate).
Posting for comment (doesn't fully bootstrap yet due to a few stray
warnings). Patches 1-4 could probably be applicable even without
the rest of the kit.
Thoughts?
David Malcolm (9):
Introduce class rtx_reader
Add selftest::read_file
selftest.h: add temp_override fixture
Expose forcibly_ggc_collect and run it after all selftests
Introduce class function_reader
df selftests
combine.c selftests
final.c selftests
cse.c selftests
gcc/Makefile.in | 5 +
gcc/cfgexpand.c | 7 +-
gcc/combine.c | 155 ++++++
gcc/cse.c | 109 +++++
gcc/df-core.c | 77 +++
gcc/emit-rtl.c | 70 ++-
gcc/emit-rtl.h | 2 +
gcc/errors.c | 23 +-
gcc/errors.h | 13 +
gcc/final.c | 271 +++++++++++
gcc/function-tests.c | 2 +-
gcc/function.c | 41 +-
gcc/function.h | 4 +-
gcc/genconstants.c | 3 +-
gcc/genenums.c | 3 +-
gcc/genmddeps.c | 3 +-
gcc/genpreds.c | 9 +-
gcc/gensupport.c | 29 +-
gcc/gensupport.h | 6 +-
gcc/ggc-tests.c | 28 +-
gcc/print-rtl.c | 4 +-
gcc/read-md.c | 328 +++++++++----
gcc/read-md.h | 192 ++++++--
gcc/read-rtl-function.c | 1197 ++++++++++++++++++++++++++++++++++++++++++++++
gcc/read-rtl-function.h | 29 ++
gcc/read-rtl.c | 760 ++++++++++++++++++++++++++++-
gcc/rtl.h | 4 +
gcc/selftest-rtl.c | 81 ++++
gcc/selftest-rtl.h | 66 +++
gcc/selftest-run-tests.c | 11 +
gcc/selftest.c | 60 +++
gcc/selftest.h | 46 ++
gcc/tree-dfa.c | 5 +
33 files changed, 3410 insertions(+), 233 deletions(-)
create mode 100644 gcc/read-rtl-function.c
create mode 100644 gcc/read-rtl-function.h
create mode 100644 gcc/selftest-rtl.c
create mode 100644 gcc/selftest-rtl.h
--
1.8.5.3