search

llvm / llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.
http://llvm.org
Other
29k stars 11.95k forks source link

Bitcode file with a large flat function takes a long time to compile with llc -O3 #32668

Open llvmbot opened 7 years ago

llvmbot commented 7 years ago
Bugzilla Link 33321
Version 4.0
OS MacOS X
Attachments Bitcode file
Reporter LLVM Bugzilla Contributor
CC @fhahn

Extended Description

My Common Lisp compiler (github.com/drmeister/clasp) generates a RUN-ALL function for generating Common Lisp objects at startup. The time it takes clang to compile this function scales roughly as N^3 where N is the number of instructions.

Use llc -O3 biglist20000.bc

This builds a 20,000 entry list It takes more than 2 minutes on my machine. With 7,000 entries it takes 7 seconds.

fhahn commented 5 years ago

Majority of the time is spent in Local Splitting

===-------------------------------------------------------------------------=== ... Pass execution timing report ... ===-------------------------------------------------------------------------=== Total Execution Time: 2085.7452 seconds (2189.4576 wall clock)

---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name --- 1972.5722 ( 95.3%) 12.1684 ( 80.6%) 1984.7406 ( 95.2%) 2087.3912 ( 95.3%) Greedy Register Allocator 82.7289 ( 4.0%) 0.8171 ( 5.4%) 83.5460 ( 4.0%) 84.4669 ( 3.9%) Machine Instruction Scheduler 5.4599 ( 0.3%) 1.8749 ( 12.4%) 7.3348 ( 0.4%) 7.4621 ( 0.3%) X86 Assembly Printer 5.2984 ( 0.3%) 0.0151 ( 0.1%) 5.3134 ( 0.3%) 5.3259 ( 0.2%) Stack Slot Coloring 2.0682 ( 0.1%) 0.1150 ( 0.8%) 2.1833 ( 0.1%) 2.1835 ( 0.1%) X86 DAG->DAG Instruction Selection 0.4669 ( 0.0%) 0.0103 ( 0.1%) 0.4772 ( 0.0%) 0.4773 ( 0.0%) Live Variable Analysis 0.1772 ( 0.0%) 0.0024 ( 0.0%) 0.1796 ( 0.0%) 0.1796 ( 0.0%) Simple Register Coalescing

===-------------------------------------------------------------------------=== Register Allocation ===-------------------------------------------------------------------------=== Total Execution Time: 1873.7156 seconds (1970.9389 wall clock)

---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name --- 1861.3112 ( 99.9%) 11.1987 ( 98.1%) 1872.5099 ( 99.9%) 1969.5353 ( 99.9%) Local Splitting 0.6046 ( 0.0%) 0.1823 ( 1.6%) 0.7869 ( 0.0%) 0.9582 ( 0.0%) Evict 0.3730 ( 0.0%) 0.0369 ( 0.3%) 0.4099 ( 0.0%) 0.4366 ( 0.0%) Spiller 0.0087 ( 0.0%) 0.0001 ( 0.0%) 0.0088 ( 0.0%) 0.0088 ( 0.0%) Seed Live Regs 1862.2976 (100.0%) 11.4180 (100.0%) 1873.7156 (100.0%) 1970.9389 (100.0%) Total

llvmbot commented 7 years ago

Are you able to profile and see where the time is spent? I recommend passing -time-passes to llc and see what tells you.

llvmbot commented 7 years ago

Bug llvm/llvm-project#32667 has been marked as a duplicate of this bug.

arsenm commented 1 year ago

Seems improved:

$ llc -O3 -time-passes  -o /tmp/biglists.s  biglists20000.bc
===-------------------------------------------------------------------------===
                          Pass execution timing report
===-------------------------------------------------------------------------===
  Total Execution Time: 28.6779 seconds (28.6949 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
  22.6764 ( 81.1%)   0.0597 (  8.2%)  22.7362 ( 79.3%)  22.7434 ( 79.3%)  Greedy Register Allocator #2
   1.5435 (  5.5%)   0.0617 (  8.5%)   1.6053 (  5.6%)   1.6120 (  5.6%)  X86 DAG->DAG Instruction Selection
   0.4389 (  1.6%)   0.5915 ( 81.3%)   1.0304 (  3.6%)   1.0313 (  3.6%)  X86 Assembly Printer