|From:||Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>|
|To:||Thomas Munro <thomas(dot)munro(at)enterprisedb(dot)com>|
|Cc:||Alvaro Herrera <alvherre(at)2ndquadrant(dot)com>, Andrew Gierth <andrew(at)tao11(dot)riddles(dot)org(dot)uk>, David Rowley <david(dot)rowley(at)2ndquadrant(dot)com>, PostgreSQL Hackers <pgsql-hackers(at)lists(dot)postgresql(dot)org>|
|Subject:||Re: Using POPCNT and other advanced bit manipulation instructions|
|Views:||Raw Message | Whole Thread | Download mbox | Resend email|
Some further thoughts here ...
Does the "lzcnt" runtime probe actually do anything useful?
On the x86_64 compilers I tried (gcc 8.2.1 and 4.4.7), __builtin_clz
and __builtin_ctz compile to sequences involving bsrq and bsfq
regardless of -mpopcnt. It's fairly hard to see how lzcnt would
buy anything over those sequences even if there were zero overhead
involved in using it.
Alvaro noted that the test programs used by c-compiler.m4 fail
to produce any actual code involving the builtin, because of
compile-time constant folding. This seems pretty unwise.
I see that on my x86_64 compilers, without -mpopcnt,
__builtin_popcnt compiles to a call of some libgcc function
or other. It's conceivable that on an (arguably misconfigured)
platform, these c-compiler.m4 tests would pass yet the build fails
at link because libgcc lacks the needed infrastructure. These tests
should be coded in a way that doesn't allow the call to be optimized
away -- cf comments for PGAC_C_BUILTIN_OP_OVERFLOW.
Also, it's starting to seem like we have enough probes for compiler
builtins that we should fold them to use one set of infrastructure.
There are some like __builtin_constant_p that probably do need their
own custom tests, but these ones that just verify that a call
compiles seem pretty duplicative ...
regards, tom lane
|Next Message||Gavin Flower||2019-02-14 05:29:24||Re: Using POPCNT and other advanced bit manipulation instructions|
|Previous Message||Tom Lane||2019-02-14 04:44:08||Re: Using POPCNT and other advanced bit manipulation instructions|