lkml.org 
[lkml]   [2015]   [Dec]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[RFC 00/17] ARC Dwarf unwinder improvements
Date
Hi guys,

In light of perf -g stalling (as unwinder was taking ~3million cycles for non
existent entries), I've revamped the dwarf unwinder.

There are some optim tweaks and much of it is "De-generalization" for things
which we can safely assume on ARC.

Crude Instrumentation shows following improvements per unwinder call:
- Avg time come down from ~4650 cycles to ~2794 cycles (+40%)
- Max time come down 9793 cycles to 5987 cycles

This is on a SMP FPGA config @ 75 MHz

It seems much of time (65%) is taken for binary lookup thru ~12k FDE entries,
roughly 13 lookups, each likely a dcache miss.


git://git.kernel.org/pub/scm/linux/kernel/git/vgupta/arc # topic-unwinder-rework-4-instrument

-Vineet

Vineet Gupta (17):
ARC: dw2 unwind: Elide generation of const propagated clones
ARC: dw2 unwind: remove unused cruft
ARC: dw2 unwind: Remove handling of for signal frame
ARC: dw2 unwind: Remove FP based unwinding
ARC: dw2 unwind: Better printing
ARC: dw2 unwind: Don't verify Main FDE Table size everytime
ARC: dw2 unwind: Refactor the FDE lookup table (eh_frame_header) code
ARC: dw2 unwind: Don't verify FDE lookup table metadata
ARC: dw2 unwind: Use striaght forward code to implement binary lookup
ARC: dw2 unwind: CIE parsing/validation done only once at startup
ARC: dw2 unwind: Elide REG_INVALID check
ARC: dw2 unwind: Elide a loop if DW_CFA_register not present
ARC: dw2 unwind: Assume all regs to be unsigned long
ARC: dw2 unwind: No need for __get_user
ARC: dw2 unwind: Single exit point for instrumentation
ARC: dw2 unwind: skip regs not updated
xxx: instrument

arch/arc/include/asm/unwind.h | 47 +--
arch/arc/kernel/Makefile | 1 +
arch/arc/kernel/unwind.c | 806 ++++++++++++++++--------------------------
3 files changed, 313 insertions(+), 541 deletions(-)

--
1.9.1



\
 
 \ /
  Last update: 2015-12-03 14:01    [W:0.078 / U:0.404 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site