Hi, gfortran's handling of .mod files leaves a lot to be desired. One problem being that the parsing of the module is done in such a way that we seek back and forth in the file; see e.g. comments in PR 25708. This not only causes us to spend unnecessary time executing lseek syscalls, but we also read a lot more data than needed, as the typical stdio implementation discards any buffered data when the file is seeked.
Furthermore, I had a patch a year or so ago to reduce the size of the
module files (PR 40958) by compressing them with zlib. Unfortunately
this turned out to be unworkable since seeking in a zlib file is
extremely inefficient (it appears to restart uncompression from the
beginning of the file in order to count up to the right offset). To
fix this I tried various approaches to parse module files without
requiring seeking, but I was not successful in this.
The attached patch takes the crude approach of first sequentially
reading the .mod file into a temporary buffer, then does the actual
parsing from that buffer. Testing on CP2K_2009-05-01.f90 (single-file
version of CP2K for benchmarking purposes), the -fsyntax-only compile
time drops from 55 seconds to 46, lseek syscalls drop from ~512k to
22k, read syscalls from 460k to 37k. This patch lays the groundwork
for
- zlib compression of .mod files (part 2/3).
- Caching module files, a crude implementation of the old "module
namespaces" idea. E.g. put the uncompressed module contents into a map
keyed by module name. (part 3/3).
Regtested on x86_64-unknown-linux-gnu, Ok for trunk?
2013-03-24 Janne Blomqvist <j...@gcc.gnu.org>
PR fortran/25708
* module.c (module_locus): Use long for position.
(module_content): New variable.
(module_pos): Likewise.
(prev_character): Remove.
(bad_module): Free data instead of closing mod file.
(set_module_locus): Use module_pos.
(get_module_locus): Likewise.
(module_char): use buffer rather than stdio file.
(module_unget_char): Likewise.
(read_module_to_tmpbuf): New function.
(gfc_use_module): Call read_module_to_tmpbuf.
--
Janne Blomqvist
tmpbuf.diff
Description: Binary data
