On 11/11/2017 19:40, Paul Smith wrote:
> On Wed, 2017-11-08 at 14:30 +0100, Paolo Bonzini wrote:
>> Another month, another email begging for an update??
>
> These changes have been pushed into the Git repository, thanks very
> much.
Thanks to you!
> I needed to adjust s
On 22/09/2017 12:46, Paolo Bonzini wrote:
> On 11/08/2017 13:44, Paolo Bonzini wrote:
>> Overall, they provide a 15% improvement for my favorite testcase (QEMU's
>> "no-op" build). These were first sent out last November, but there
>> have been quite a fe
> I'll need to think about the hash patches: the hash implementation in
> GNU make is taken pretty straightforwardly from id-utils. If we start
> to modify it significantly here we'd be cutting that relationship. That
> may be OK, but I'd need to think about it.
Maybe you want to backport those
On 11/08/2017 13:44, Paolo Bonzini wrote:
> Overall, they provide a 15% improvement for my favorite testcase (QEMU's
> "no-op" build). These were first sent out last November, but there
> have been quite a few changes since then:
>
> - I have dropped the SSE-opt
Compare against '$' directly rather than using MAP_VARIABLE. This
is a ~10% speedup for find_map_function, which is the top hotspot in
QEMU's no-op build. The build is sped up overall by about 1.5% more
(from 11.1 to 10.95 seconds).
* read.c (find_map_function): Do not compare against singleton
Use the stopchar map to quickly jump over everything that is not an
open/close brace, an open/close parenthesis or a comma.
This provides a roughly 1% improvement on QEMU's noop build (from
11.23 to 11.1 seconds).
* function.c (find_next_argument, handle_function): Check
with STOP_SET before comp
* read.c (find_percent_cached): Use strchr instead of STOP_SET
to find % or nul.
* makeint.h (MAP_PERCENT): Remove.
* main.c (initialize_stopchar_map): Remove.
---
main.c| 1 -
makeint.h | 1 -
read.c| 10 +-
3 files changed, 5 insertions(+), 7 deletions(-)
diff --git a/main.c b
In most cases, find_char_unquote has a single stopchar. In that
case we can look for it using strchr's optimized implementation.
The resulting speedup on QEMU's noop build is 3.5% (from 11.78 to 11.37
seconds).
* read.c (find_char_unquote): Rename to find_map_unquote. Replace
with an implementa
This is about twice as fast as the current hash, and removes the need
for double hashing (improving locality of reference). The hash function
is based on Bob Jenkins' design, slightly adapted wherever Make needs
to hash NUL-terminated strings. The old hash function is kept for
case-insensitive ha
Makefile. For this reason the times are lower and the speedup
from patch 5 is smaller than I had reported for v1.
Paolo
Paolo Bonzini (6):
use Jenkins hash
use strchr for simple case of find_char_unquote
use strchr/memmove in collapse_continuations
remove MAP_PERCENT
speedup parsing of
collapse_continuations is already using strchr to speed up the common
case of no backslash-newline sequence, but on modern processors it is
faster to scan the string twice with strchr+memmove (or strlen+memmove)
than to move bytes manually.
This saves about 1.5% on QEMU's no-op build (from 11.37 t
On 11/11/2016 18:42, Paolo Bonzini wrote:
>
>
> On 11/11/2016 17:44, Paul Smith wrote:
>> On Fri, 2016-11-11 at 12:28 +0100, Paolo Bonzini wrote:
>>> Paul, any news?
>>
>> I expect to carve out a chunk of time to work on GNU make maintenance
>> next mo
On 11/11/2016 17:44, Paul Smith wrote:
> On Fri, 2016-11-11 at 12:28 +0100, Paolo Bonzini wrote:
>> Paul, any news?
>
> I expect to carve out a chunk of time to work on GNU make maintenance
> next month, and will try to put out a new release. These updates are
> in my qu
On 02/11/2016 17:24, Paolo Bonzini wrote:
> +int needs_glob(const char *s)
> +{
> + return strpbrk (s, "?*[") == NULL;
My bad, this should be "!=" NULL. I will send a v2 when I get more
feedback.
Paolo
> +}
> +#endif
> diff --git a/read.c b/read.c
&
On 02/11/2016 17:24, Paolo Bonzini wrote:
> These patches optimize the hotspots in QEMU's "noop" build:
>
> - patch 1: parse_file_seq / strpbrk
>
> - patch 2: find_char_unquote
>
> - patch 3: variable_hash_1 / variable_hash_2
>
> - patches 4+5: h
On 03/11/2016 09:44, Andreas Schwab wrote:
> On Nov 02 2016, Paolo Bonzini wrote:
>
>> +#define sum_get_unaligned_32(r, p) \
>> + do { \
>> +unsigned int val; \
>
This is needed to free a stopchar_map bit for the next patch.
* read.c (find_percent_cached): Use strchr instead of STOP_SET
to find % or nul.
* makeint.h (MAP_PERCENT): Remove.
* main.c (initialize_stopchar_map): Remove.
---
main.c| 1 -
makeint.h | 1 -
read.c| 10 +-
3 files
This removes the hotspot in parse_file_seq's call to strpbrk, by using
SSE2 vector instructions. The resulting speedup on QEMU's noop build
is around 6% (15.4 seconds to 14.5).
The code is roughly based on GCC's similar optimizations in the lexer.
* read-opt.c: New.
* read.c (parse_file_seq): Us
In most cases, find_char_unquote has a single stopchar. In that
case we can look for it using strchr's optimized implementation.
The resulting speedup on QEMU's noop build is 4.4% (from 14.5 seconds
to 13.8).
* read.c (find_char_unquote): Rename to find_map_unquote. Replace
with an implementati
employer has a blanket assignment on file).
Thanks,
Paolo
Paolo Bonzini (5):
optimize checking for globs
use strchr for simple case of find_char_unquote
use jhash for STRING_N_HASH
remove MAP_PERCENT
speedup parsing of functions
Makefile.am | 6 ++---
Makefile.in | 19 ++---
f
Use the stopchar map to quickly jump over everything that is not an
open/close brace, an open/close parenthesis or a comma. This provides
an 1.5% improvement on QEMU's noop build (from 13.2 seconds to 13).
* function.c (find_next_argument, handle_function): Ignore
characters quickly using the sto
The hottest hash table in Make is the variable hash, and it is
easy to apply a better function (that works on bigger chunks
than just one byte) because the length of its key is known.
This is about twice as fast as the current hash, and removes
the need for double hashing (improving locality of ref
The manual's example of using $(findstring) to test MAKEFLAGS:
archive.a: ...
ifneq (,$(findstring t,$(MAKEFLAGS)))
+touch archive.a
+ranlib -t archive.a
else
ranlib archive.a
endif
was broken by the addition of --no-print-directory to $
GNU make 3.80 is a HUGE memory hog. It calls xstrdup to build
dependency list. gnu-src-gcc.deps in libjava has 3000+ targets depend
the same 3000+ files, whose filenames are more than 260K. For this
dependency alone, make takes 3000*260K == 761MB.
Then, you should make the 3000+ target depend
24 matches
Mail list logo
