[llvm-testresults] Grue-x86-64-O0-pic nightly tester results
Chris Lattner
clattner at apple.com
Wed Mar 3 17:59:03 PST 2010
This looks much better for me on mainline now, but I'm continuing to work on it:
TOT:
$ llvm/Release-Asserts/bin/llc ~/Desktop/oggenc.llvm.bc -asm-verbose=false -regalloc=local -O0 -time-passes -o /dev/null
===-------------------------------------------------------------------------===
Instruction Selection and Scheduling
===-------------------------------------------------------------------------===
Total Execution Time: 0.0801 seconds (0.0801 wall clock)
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
0.0148 ( 20.7%) 0.0008 ( 9.7%) 0.0156 ( 19.5%) 0.0156 ( 19.5%) Instruction Selection
0.0105 ( 14.7%) 0.0009 ( 10.8%) 0.0114 ( 14.3%) 0.0114 ( 14.3%) Instruction Scheduling
0.0079 ( 11.0%) 0.0017 ( 19.0%) 0.0095 ( 11.9%) 0.0097 ( 12.2%) DAG Combining 1
0.0086 ( 12.1%) 0.0009 ( 10.5%) 0.0096 ( 11.9%) 0.0096 ( 11.9%) DAG Legalization
0.0081 ( 11.3%) 0.0010 ( 11.2%) 0.0091 ( 11.3%) 0.0090 ( 11.3%) Type Legalization
0.0081 ( 11.3%) 0.0009 ( 10.2%) 0.0090 ( 11.2%) 0.0089 ( 11.1%) Instruction Creation
0.0058 ( 8.1%) 0.0008 ( 9.0%) 0.0065 ( 8.2%) 0.0065 ( 8.2%) Vector Legalization
0.0044 ( 6.2%) 0.0008 ( 9.6%) 0.0052 ( 6.6%) 0.0052 ( 6.5%) DAG Combining 2
0.0020 ( 2.8%) 0.0008 ( 9.2%) 0.0028 ( 3.5%) 0.0028 ( 3.5%) Instruction Scheduling Cleanup
0.0012 ( 1.6%) 0.0001 ( 0.9%) 0.0013 ( 1.6%) 0.0012 ( 1.6%) DAG Combining after legalize types
0.0714 (100.0%) 0.0087 (100.0%) 0.0801 (100.0%) 0.0801 (100.0%) TOTAL
...
Before:
$ llvm-97474/Release-Asserts/bin/llc ~/Desktop/oggenc.llvm.bc -asm-verbose=false -regalloc=local -O0 -time-passes -o /dev/null
===-------------------------------------------------------------------------===
Instruction Selection and Scheduling
===-------------------------------------------------------------------------===
Total Execution Time: 0.0757 seconds (0.0755 wall clock)
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
0.0104 ( 15.5%) 0.0011 ( 12.0%) 0.0114 ( 15.1%) 0.0114 ( 15.1%) Instruction Selection
0.0103 ( 15.5%) 0.0009 ( 10.2%) 0.0112 ( 14.8%) 0.0112 ( 14.8%) Instruction Scheduling
0.0078 ( 11.7%) 0.0017 ( 18.5%) 0.0095 ( 12.5%) 0.0096 ( 12.8%) DAG Combining 1
0.0086 ( 13.0%) 0.0009 ( 10.3%) 0.0096 ( 12.6%) 0.0095 ( 12.6%) DAG Legalization
0.0081 ( 12.1%) 0.0010 ( 11.4%) 0.0091 ( 12.0%) 0.0090 ( 11.9%) Type Legalization
0.0081 ( 12.2%) 0.0009 ( 9.4%) 0.0090 ( 11.9%) 0.0089 ( 11.8%) Instruction Creation
0.0058 ( 8.7%) 0.0008 ( 9.3%) 0.0066 ( 8.7%) 0.0066 ( 8.7%) Vector Legalization
0.0044 ( 6.6%) 0.0008 ( 9.3%) 0.0052 ( 6.9%) 0.0052 ( 6.9%) DAG Combining 2
0.0020 ( 3.0%) 0.0008 ( 8.9%) 0.0028 ( 3.7%) 0.0028 ( 3.7%) Instruction Scheduling Cleanup
0.0012 ( 1.7%) 0.0001 ( 0.9%) 0.0012 ( 1.6%) 0.0012 ( 1.6%) DAG Combining after legalize types
0.0667 (100.0%) 0.0091 (100.0%) 0.0757 (100.0%) 0.0755 (100.0%) TOTAL
Incidentally, the major issue here is that fast isel doesn't handle llvm.objectsize.
-Chris
On Mar 3, 2010, at 11:35 AM, Evan Cheng wrote:
> I am seeing some isel slow down. Here is the numbers I got for ldecod and oggenc:
>
> tot:
> 0.0330 ( 22.1%) 0.0014 ( 9.7%) 0.0344 ( 21.0%) 0.0343 ( 21.0%) Instruction Selection
> 0.0216 ( 24.5%) 0.0008 ( 9.9%) 0.0224 ( 23.3%) 0.0223 ( 23.3%) Instruction Selection
>
> 97474:
> 0.0197 ( 14.8%) 0.0016 ( 10.6%) 0.0212 ( 14.3%) 0.0212 ( 14.3%) Instruction Scheduling
> 0.0120 ( 15.2%) 0.0010 ( 12.6%) 0.0130 ( 15.0%) 0.0130 ( 14.9%) Instruction Selection
>
> This is how to test it
> llc -asm-verbose=false -regalloc=local -O0 Output/oggenc.llvm.bc -o Output/oggenc.llc.s -time-passes
>
> Evan
> <oggenc.llvm.bc>
> On Mar 3, 2010, at 11:04 AM, Evan Cheng wrote:
>
>> I'll take a look.
>>
>> Evan
>> On Mar 3, 2010, at 11:00 AM, Chris Lattner wrote:
>>
>>> On Mar 3, 2010, at 10:53 AM, Evan Cheng wrote:
>>>> This is an -O0 tester so the compile time regressions are really bad. Today's tester shows these have not recovered. Chris, I am assuming these are yours since codegen still uses dag isel even at -O0? If I remember correctly, we disable the more expensive isel optimizations at -O0. Have these changed?
>>>
>>> I'd be really surprised if they were mine, I'm not seeing anywhere near that slowdown and DAG ISel stuff doesn't get run much at -O0. Also, this run was from before I enabled the new isel (it was from Mar 1 @ 6am, I turned it on at Mar 1 @ 11am). I don't see a way to extract a rev # from the nightly test run, but the new isel got enabled in r97475.
>>>
>>> -Chris
>>>
>>>
>>>> On Mar 2, 2010, at 6:02 AM, daniel_dunbar at apple.com wrote:
>>>>
>>>>> http://smooshlab.apple.com/perf/db_nt_internal/nightlytest/5896/
>>>>> Name: grue.apple.com
>>>>> Nickname: Grue-x86-64-O0-pic:20
>>>>>
>>>>> Run: 5896, Start Time: 2010-03-02 01:19:08, End Time: 2010-03-02 06:02:20
>>>>> Comparing To: 5864, Start Time: 2010-03-01 01:19:13, End Time: 2010-03-01 06:04:01
>>>>>
>>>>> --- Changes Summary ---
>>>>> New Test Passes: 0
>>>>> New Test Failures: 0
>>>>> Added Tests: 0
>>>>> Removed Tests: 0
>>>>> Significant Changes: 26
>>>>>
>>>>> --- Tests Summary ---
>>>>> Total Tests: 3582
>>>>> Total Test Failures: 1654
>>>>>
>>>>> Total Test Failures By Type:
>>>>> CBE: 62
>>>>> JIT: 398
>>>>> JIT codegen: 398
>>>>> LLC-BETA: 398
>>>>> LLC_BETA compile: 398
>>>>>
>>>>> --- Changes Detail ---
>>>>> New Test Passes:
>>>>>
>>>>> New Test Failures:
>>>>>
>>>>> Added Tests:
>>>>>
>>>>> Removed Tests:
>>>>>
>>>>> Significant Changes in Test Results:
>>>>> LLC compile:
>>>>> MultiSource/Applications/ClamAV/clamscan: 9.14% (1.5122 => 1.6504)
>>>>> MultiSource/Applications/JM/ldecod/ldecod: 11.83% (0.8775 => 0.9813)
>>>>> MultiSource/Applications/JM/lencod/lencod: 12.72% (1.8677 => 2.1053)
>>>>> MultiSource/Applications/SPASS/SPASS: 6.94% (2.5081 => 2.6822)
>>>>> MultiSource/Applications/d/make_dparser: 10.85% (0.3430 => 0.3802)
>>>>> MultiSource/Applications/kimwitu++/kc: 7.37% (3.6318 => 3.8996)
>>>>> MultiSource/Applications/lua/lua: 8.47% (0.5208 => 0.5649)
>>>>> MultiSource/Applications/oggenc/oggenc: 12.05% (0.4921 => 0.5514)
>>>>> MultiSource/Applications/siod/siod: 7.74% (0.3655 => 0.3938)
>>>>> MultiSource/Applications/sqlite3/sqlite3: 10.21% (1.0976 => 1.2097)
>>>>> MultiSource/Applications/treecc/treecc: 7.31% (0.3776 => 0.4052)
>>>>> MultiSource/Benchmarks/ASCI_Purple/SMG2000/smg2000: 12.35% (0.4509 => 0.5066)
>>>>> MultiSource/Benchmarks/Bullet/bullet: 9.40% (3.7108 => 4.0595)
>>>>> MultiSource/Benchmarks/MallocBench/espresso/espresso: 7.81% (0.5695 => 0.6140)
>>>>> MultiSource/Benchmarks/MallocBench/gs/gs: 10.09% (0.5749 => 0.6329)
>>>>> MultiSource/Benchmarks/MiBench/consumer-jpeg/consumer-jpeg: 11.78% (0.3547 => 0.3965)
>>>>> MultiSource/Benchmarks/MiBench/consumer-lame/consumer-lame: 12.35% (0.3660 => 0.4112)
>>>>> MultiSource/Benchmarks/MiBench/consumer-typeset/consumer-typeset: 14.89% (2.1545 => 2.4753)
>>>>> MultiSource/Benchmarks/PAQ8p/paq8p: 8.80% (0.3297 => 0.3587)
>>>>> MultiSource/Benchmarks/Prolangs-C/TimberWolfMC/timberwolfmc: 11.54% (0.6094 => 0.6797)
>>>>> MultiSource/Benchmarks/mafft/pairlocalalign: 11.28% (0.7232 => 0.8048)
>>>>> MultiSource/Benchmarks/mediabench/jpeg/jpeg-6a/cjpeg: 11.86% (0.3693 => 0.4131)
>>>>> MultiSource/Benchmarks/tramp3d-v4/tramp3d-v4: 6.60% (4.8926 => 5.2154)
>>>>> SingleSource/Benchmarks/Adobe-C++/loop_unroll: 6.84% (0.3157 => 0.3373)
>>>>> SingleSource/Benchmarks/Adobe-C++/simple_types_constant_folding: 6.46% (0.5105 => 0.5435)
>>>>> SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant: 7.76% (0.3530 => 0.3804)
>>>>> _______________________________________________
>>>>> llvm-testresults mailing list
>>>>> llvm-testresults at cs.uiuc.edu
>>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-testresults
>>>>
>>>
>>
>>
>> _______________________________________________
>> llvm-testresults mailing list
>> llvm-testresults at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-testresults
>
More information about the llvm-testresults
mailing list