The patch is modified based on the review comments from Richard, Bernd and David.
The following changes are done to incorporate the comments received on the
previous mail.
1. With this patch the liveness of the Loop is not stored on the LOOPDATA
structures. The liveness is calculated based on the current loops
for which the invariant is checked and the regs_used is calculated.
2. Memory leaks are fixed.
3. Reworked on the comments section based on Bernd's comments.
Bootstrapped and regtested for i386 and Microblaze target.
SPEC CPU 2000 benchmarks are run on i386 target and following is the summary of
the results.
SPEC CPU 2000 INT benchmarks.
( Gemoean Score without the change vs Geomean score with reg pressure change =
3745.193 vs 3745.328)
SPEC CPU 2000 FP benchmarks.
( Gemoean Score without the change vs Geomean score with reg pressure change =
4741.825 vs 4748.364).
[Patch,rtl Optimization]: Better register pressure estimate for loop
invariant code motion
Calculate the loop liveness used for regs for calculating the register
pressure
in the cost estimation. Loop liveness is based on the following properties.
We only need to find the set of objects that are live at the birth or the
header
of the loop. We don't need to calculate the live through the loop by
considering
live in and live out of all the basic blocks of the loop. This is based on
the
point that the set of objects that are live-in at the birth or header of
the loop
will be live-in at every node in the loop.
If a v live is out at the header of the loop then the variable is live-in
at every node
in the loop. To prove this, consider a loop L with header h such that the
variable v
defined at d is live-in at h. Since v is live at h, d is not part of L.
This follows i
from the dominance property, i.e. h is strictly dominated by d.
Furthermore, there
exists a path from h to a use of v which does not go through d. For every
node p in
the loop, since the loop is strongly connected and node is a component of
the CFG,
there exists a path, consisting only of nodes of L from p to h.
Concatenating these
two paths proves that v is live-in and live-out of p.
Calculate the live-out and live-in for the exit edge of the loop. This
patch considers
liveness for not only the loop latch but also the liveness outside the
loops.
ChangeLog:
2015-12-15 Ajit Agarwal <ajit...@xilinx.com>
* loop-invariant.c
(find_invariants_to_move): Add the logic of regs_used based
on liveness.
* cfgloopanal.c
(estimate_reg_pressure_cost): Update the heuristics in presence
of call_p.
Signed-off-by:Ajit Agarwal ajit...@xilinx.com.
Thanks & Regards
Ajit
-----Original Message-----
From: Bernd Schmidt [mailto:bschm...@redhat.com]
Sent: Wednesday, December 09, 2015 7:34 PM
To: Ajit Kumar Agarwal; Richard Biener
Cc: Jeff Law; GCC Patches; Vinod Kathail; Shail Aditya Gupta; Vidhumouli
Hunsigida; Nagaraju Mekala
Subject: Re: [Patch,rtl Optimization]: Better register pressure estimate for
Loop Invariant Code Motion.
On 12/09/2015 12:22 PM, Ajit Kumar Agarwal wrote:
>
> This is because the available_regs = 6 and the regs_needed = 1 and
> new_regs = 0 and the regs_used = 10. As the reg_used that are based
> on the Liveness given above is greater than the available_regs, then
> it's candidate of spill and the estimate_register_pressure calculates > the
> spill cost. This spill cost is greater than inv_cost and gain > comes to be
> negative. The disables the loop invariant for the above > testcase.
As far as I can tell this loop does not lead to a spill. Hence, failure of this
testcase would suggest there is something wrong with your idea.
>> + FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) {
Formatting.
>> + /* Loop Liveness is based on the following proprties.
"properties"
>> + we only require to calculate the set of objects that are live at
>> + the birth or the header of the loop.
>> + We don't need to calculate the live through the Loop considering
>> + Live in and Live out of all the basic blocks of the Loop. This is
>> + because the set of objects. That are live-in at the birth or header
>> + of the loop will be live-in at every node in the Loop.
>> + If a v live out at the header of the loop then the variable is
>> live-in
>> + at every node in the Loop. To prove this, Consider a Loop L with
>> header
>> + h such that The variable v defined at d is live-in at h. Since v is
>> live
>> + at h, d is not part of L. This follows from the dominance property,
>> i.e.
>> + h is strictly dominated by d. Furthermore, there exists a path from
>> h to
>> + a use of v which does not go through d. For every node of the loop,
>> p,
>> + since the loop is strongly connected Component of the CFG, there
>> exists
>> + a path, consisting only of nodes of L from p to h. Concatenating
>> those
>> + two paths prove that v is live-in and live-out Of p. */
Please Refrain From Randomly Capitalizing Words, and please also fix the
grammar ("set of objects. That are live-in"). These problems make this comment
(and also your emails) extremely hard to read.
Partly for that reason, I can't quite make up my mind whether this patch makes
things better or just different. The testcase failure makes me think it's
probably not an improvement.
>> + bitmap_ior_into (&LOOP_DATA (loop)->regs_live, DF_LR_IN (loop->header));
>> + bitmap_ior_into (&LOOP_DATA (loop)->regs_live, DF_LR_OUT
>> + (loop->header));
Formatting.
Bernd
pressure-licm.patch
Description: pressure-licm.patch
