Link-Time Optimization on PlayStation®4 Yunzhong Gao Sony Computer Entertainment LLVM Dev Meeting, 28-29 Oct 2014
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Link-Time Optimization
on PlayStation®4
Yunzhong Gao Sony Computer Entertainment
LLVM Dev Meeting, 28-29 Oct 2014
Last Year… • LTO without linker
Combined bc file
llvm-link
opt
object file
llc
Output executable
PS4TM
linker PS4TM linker
Input bc files Input object files
• LOTS of contribution from the community
• Thank you!
• Reference: see Rafael’s talk at Euro LLVM 2014
Past work in LTO…
In the Linker… • Proprietary linker
In the Linker… • Proprietary linker
• Uses tools/LTO interface • lto_module_is_object_file()
Input bc files Input object files
In the Linker…
Input bc files Input object files
In the Linker…
Scan Phase
for each bc file: • lto_module_get_symbol_name()
• lto_module_get_symbol_attribute()
• lto_module_get_deplib()
• lto_codegen_add_module()
Input bc files Input object files
Scan Phase
In the Linker…
Input bc files Input object files
Scan Phase
In the Linker…
Compile Phase
• lto_codegen_add_must_preserve_symbol()
• lto_codegen_compile_to_file()
Input bc files Input object files
Scan Phase
Compile Phase
In the Linker…
Input bc files Input object files
Object file
Scan Phase
Compile Phase
In the Linker…
Input bc files Input object files
Object file
Scan Phase
Compile Phase
Output executable
Usual linking steps Usual linking steps
In the Linker…
Whose Bug?
Clearly, this is a bug
in the LTO library.
No, CLEARLY, it is
a linker problem.
Whose Bug?
wrapper
LTO.dll
Linker
commands
results
Whose Bug?
wrapper
Linker
commands
results
Cannot select: 0x2e329d0:
v4i32 = …
Whose Bug?
wrapper
LTO.dll
Linker
commands
results
Whose Bug?
wrapper
LTO.dll
results
A list of
commands
stdout
Evaluation
Results on some PS4TM game titles:
Game#1 Game#2 Game#3 Game#4 Game#5
Run-Time
Performance
Improved
3.39%
Improved
1.43%
Improved
9%
Improved
6.11%
Improved
6%
Evaluation
Results on some PS4TM game titles:
Game#1 Game#2 Game#3 Game#4 Game#5
Run-Time
Performance
Improved
3.39%
Improved
1.43%
Improved
9%
Improved
6.11%
Improved
6%
• Inliner
• Internalize Global Symbols
• Inter-Procedural Sparse Conditional Constant Propagation
Evaluation
Results on some PS4TM game titles:
Game#1 Game#2 Game#3 Game#4 Game#5
Executable
Size
Increased
0.242%
Increased
0.539%
Increased
2%
Decreased
11.8%
Increased
2.06%
Evaluation
Results on some PS4TM game titles:
(without debug info)
Game#1 Game#2 Game#3 Game#4 Game#5
Build Time
(LTO vs non-LTO)
4x 3x 5x 2.5x 10x
Related Work
Related Work
• LTO with debug info: memory usage and build time
Related Work
• LTO with debug info: memory usage and build time
• Versioning of bc files
Related Work
• LTO with debug info: memory usage and build time
• Versioning of bc files
• Further performance improvement in LTO: • Move some optimization passes from compiler to LTO?
(e.g. loop-unroll, vectorizer)
Related Documents
