[Bug runtime/6028] ia64_fetch_register is very slow

[Bug runtime/6028] ia64_fetch_register is very slow

[mhiramat at redhat dot com]

------- Additional Comments From mhiramat at redhat dot com  2008-04-03 21:40 -------
Created an attachment (id=2350)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2350&action=view)
ia64 fast(cached) fetch_register() patch

Introduce cached ia64_fetch_register() code.
- (tapsets.cxx) add context.raddr for caching unwound address
- (regs-ia64.c) don't unwind stack if it has already been unwound.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6028

------- You are receiving this mail because: -------
You are the assignee for the bug, or are watching the assignee.

[Bug runtime/6028] ia64_fetch_register is very slow

[mhiramat at redhat dot com]

------- Additional Comments From mhiramat at redhat dot com  2008-04-03 21:42 -------
(In reply to comment #1)
> Created an attachment (id=2350)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2350&action=view)
> ia64 fast(cached) fetch_register() patch

Here is a test result with this patch.
---
access 0 flags
probe syscall.open (<input>:1:1), hits: 18, cycles: 837min/1796avg/17953max
access 1 flags
probe syscall.open (<input>:1:1), hits: 18, cycles: 1086min/1779avg/12808max
access 2 flags
probe syscall.open (<input>:1:1), hits: 19, cycles: 1167min/2778avg/21327max
access 3 flags
probe syscall.open (<input>:1:1), hits: 18, cycles: 1252min/2272avg/17831max
---

Thanks,

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|ia64_fetch_register is very |ia64_fetch_register is very
                   |slow                        |slow


http://sourceware.org/bugzilla/show_bug.cgi?id=6028

------- You are receiving this mail because: -------
You are the assignee for the bug, or are watching the assignee.

[Bug runtime/6028] ia64_fetch_register is very slow

[fche at redhat dot com]

------- Additional Comments From fche at redhat dot com  2008-04-03 22:04 -------
Interesting approach.  It should work OK.  After committing, please
add a comment to the translator to explain the context "raddr" field.

We should consider lumping into a union (perhaps an anonymous one)
per-architecture or per-probe-point-type context fields, and using
plain memset to clear the thing at probe entry.


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6028

------- You are receiving this mail because: -------
You are the assignee for the bug, or are watching the assignee.

[Bug runtime/6028] ia64_fetch_register is very slow

[mhiramat at redhat dot com]

------- Additional Comments From mhiramat at redhat dot com  2008-04-04 21:09 -------
committed the patch.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED


http://sourceware.org/bugzilla/show_bug.cgi?id=6028

------- You are receiving this mail because: -------
You are the assignee for the bug, or are watching the assignee.

[Bug runtime/6028] ia64_fetch_register is very slow

[mhiramat at redhat dot com]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |mhiramat at redhat dot com


http://sourceware.org/bugzilla/show_bug.cgi?id=6028

------- You are receiving this mail because: -------
You are the assignee for the bug, or are watching the assignee.

[Bug runtime/6028] ia64_fetch_register is very slow

[mhiramat at redhat dot com]

------- Additional Comments From mhiramat at redhat dot com  2008-04-09 21:31 -------
I'm still considering whether we can remove all overheads which come from
unwinding stack...

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |REOPENED
         Resolution|FIXED                       |


http://sourceware.org/bugzilla/show_bug.cgi?id=6028

------- You are receiving this mail because: -------
You are the assignee for the bug, or are watching the assignee.

[Bug runtime/6028] ia64_fetch_register is very slow

[mhiramat at redhat dot com]

------- Additional Comments From mhiramat at redhat dot com  2008-04-18 02:36 -------
I made a set of patches which is a proof of concept.
But it showed a good result of benchmark:

---
access 0 flags
probe syscall.open (<input>:1:1), hits: 16, cycles: 63min/1019avg/15115max
access 1 flags
probe syscall.open (<input>:1:1), hits: 16, cycles: 259min/1642avg/20710max
access 2 flags
probe syscall.open (<input>:1:1), hits: 16, cycles: 415min/1179avg/11588max
access 3 flags
probe syscall.open (<input>:1:1), hits: 16, cycles: 504min/1201avg/10136max
---


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6028

------- You are receiving this mail because: -------
You are the assignee for the bug, or are watching the assignee.