News on Clang

A couple of months back, Greg Stark reported mixed results on getting
Clang/LLVM to build Postgres. It could be done, but there were some
bugs, including a bug that caused certain grammar TUs to take way too
long to compile. 2 bug reports were filed. He said that the long
compile time problem was made a lot better by using SVN-tip as it
existed at the time - it brought compile time down from over a minute
to just 15 seconds for preproc.c, which was of particular concern
then.

Last night, I had a go at getting Postgres to build using my own build
of Clang SVN-tip (2.9). I was successful, but I too found certain
grammar TUs to be very slow to compile. Total times were 3m23s for
Clang, to GCC 4.6's 2m1s. I made a (perhaps duplicate) bug report,
which had a preprocessed gram.c as a testcase. Here was the compile
time that I saw for the file:

[peter(at)peter postgresql]$ time /home/peter/build/Release/bin/clang -O2
-Wall -Wmissing-prototypes -Wpointer-arith
-Wdeclaration-after-statement -Wendif-labels -Wformat-security
-fno-strict-aliasing -fwrapv -Wno-error -I. -I. -I
/home/peter/postgresql/src/include -D_GNU_SOURCE -c -o gram.o
/home/peter/postgresql/src/backend/parser/gram.c
In file included from gram.y:12939:
scan.c:16246:23: warning: unused variable 'yyg' [-Wunused-variable]
struct yyguts_t * yyg = (struct yyguts_t*)yyscanner; /* This var ...
^
1 warning generated.

real 1m5.780s
user 1m4.915s
sys 0m0.086s

The compile time is astronomically high, considering it takes about 2
seconds to compile gram.c on the same machine using GCC 4.6 with the
same flags.

This problem is reportedly a front-end issue. Observe what happens
when I omit the -Wall flag:

[peter(at)peter postgresql]$ time /home/peter/build/Release/bin/clang -O2
-Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement
-Wendif-labels -Wformat-security -fno-strict-aliasing -fwrapv
-Wno-error -I. -I. -I /home/peter/postgresql/src/include -D_GNU_SOURCE
-c -o gram.o /home/peter/postgresql/src/backend/parser/gram.c

real 0m2.153s
user 0m2.073s
sys 0m0.065s
[peter(at)peter postgresql]$

This is very close to the time for GCC. The problem has been further
isolated to the flag -Wuninitialized.

I hacked our configure file to omit -Wall in $CFLAGS . -Wall is just a
flag to have GCC/Clang "show all reasonable warnings" - it doesn't
affect binaries. I then measured the total build time for Postgres
using Clang SVN-tip. Here's what "time make" outputs:

* SNIP *
real 2m7.102s
user 1m49.015s
sys 0m10.635s

I'm very encouraged by this - Clang is snapping at the heels of GCC
here. I'd really like to see Clang as a better supported compiler,
because the whole LLVM infrastructure seems to have a lot to offer -
potentially improved compile times, better warnings/errors, a good
static analysis tool, a nice debugger, and a nice modular architecture
for integration with third party components. It is still a bit
immature, but it has the momentum.

At a large presentation that I and other PG community members were
present at during FOSDEM, Postgres was specifically cited as an
example of a medium-sized C program that had considerably improved
compile times on Clang. While I was obviously unable to reproduce the
very impressive compile-time numbers claimed (at -O0), I still think
that Clang has a lot of promise. Here are the slides from that
presentation:

http://www.scribd.com/doc/48921683/LLVM-Clang-Advancing-Compiler-Technology

--
Peter Geoghegan       http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training and Services