diff mbox

V2 [PATCH] x86: Use _rdtsc intrinsic for HP_TIMING_NOW

Message ID CAMe9rOq9fxtsfP6a+3t1J2US6xBK_GUkvxhw4KQRQBgs=m_0gQ@mail.gmail.com
State New, archived
Headers

Commit Message

H.J. Lu Oct. 20, 2018, 10:27 p.m. UTC 
  On 10/20/18, Florian Weimer <fw@deneb.enyo.de> wrote:
> * Florian Weimer:
>
>
>> I'm trying to remove the #include <hp-timing.h> from
>> <libc-internal.h>, but I don't know how hard this will be.
>
> That doesn't help.  It's pulled in via <tls.h> and <descr.h> as well.
> We would need a separate <hp_timing_t.h> header file to avoid that
> because hp_timing_t is required in a few central places (thread
> descriptor and link map).
>

Please try this.

Comments

Florian Weimer Oct. 21, 2018, 6:57 a.m. UTC | #1 
* H. J. Lu:

> +   NB: Use __builtin_ia32_rdtsc directly since including <x86intrin.h>
> +   is very slow with GCC 6.  */

Thanks for the patch.

It fixes the issue, but the comment is wrong: I see the same slowdown
with GCC 8.
diff mbox

Patch

From a4fac17d8bc95fc405faa1fac0a383719b72bccf Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <hjl.tools@gmail.com>
Date: Sat, 20 Oct 2018 15:22:01 -0700
Subject: [PATCH] x86: Don't include <x86intrin.h>

Use __builtin_ia32_rdtsc directly since including <x86intrin.h> is very
slow with GCC 6.

	* sysdeps/x86/hp-timing.h: Don't include <x86intrin.h>.
	(HP_TIMING_NOW): Replace _rdtsc with __builtin_ia32_rdtsc.
---
 sysdeps/x86/hp-timing.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/sysdeps/x86/hp-timing.h b/sysdeps/x86/hp-timing.h
index 1c20e9d828..950a2feb6b 100644
--- a/sysdeps/x86/hp-timing.h
+++ b/sysdeps/x86/hp-timing.h
@@ -22,8 +22,6 @@ 
 #include <isa.h>
 
 #if MINIMUM_ISA == 686 || MINIMUM_ISA == 8664
-# include <x86intrin.h>
-
 /* We always assume having the timestamp register.  */
 # define HP_TIMING_AVAIL	(1)
 # define HP_SMALL_TIMING_AVAIL	(1)
@@ -38,8 +36,11 @@  typedef unsigned long long int hp_timing_t;
    might not be 100% accurate since there might be some more instructions
    running in this moment.  This could be changed by using a barrier like
    'cpuid' right before the `rdtsc' instruciton.  But we are not interested
-   in accurate clock cycles here so we don't do this.  */
-# define HP_TIMING_NOW(Var)	((Var) = _rdtsc ())
+   in accurate clock cycles here so we don't do this.
+
+   NB: Use __builtin_ia32_rdtsc directly since including <x86intrin.h>
+   is very slow with GCC 6.  */
+# define HP_TIMING_NOW(Var)	((Var) = __builtin_ia32_rdtsc ())
 
 # include <hp-timing-common.h>
 #else
-- 
2.17.2