| Issue |
63788
|
| Summary |
Unable to replicate clang optimization pipeline using opt with LLVM 16
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
lucasreis1
|
I've been experimenting with optimization via Clang and I'm in need of a custom optimization pipeline for IR. I figured that I would start by replicating `clang -O3` using opt and remove passes I don't need.
A few quick tests showed me that for a specific application, Clang produced a binary that is 10% faster than when using opt. I've tried mimicking everything Clang does up until codegen except for optimizations, which I apply at opt. The result is still visible.
Here's a summary of my pipeline:
```
clang -c -emit-llvm -O3 a.cpp
clang -c -emit-llvm -O3 b.cpp
clang -c -emit-llvm -O3 c.cpp
llvm-link a.bc b.bc c.bc -o from_clang.bc
```
```
clang -Xclang -mframe-pointer=all -Xclang -vectorize-loops -Xclang vectorize-slp -Xclang -disableO0-optnone -c -emit-llvm a.bc
opt a.bc -passes="default<O3>" -o a.bc
clang -Xclang -mframe-pointer=all -Xclang -vectorize-loops -Xclang vectorize-slp -Xclang -disableO0-optnone -c -emit-llvm b.cpp
opt b.bc -passes="default<O3>" -o b.bc
clang -Xclang -mframe-pointer=all -Xclang -vectorize-loops -Xclang vectorize-slp -Xclang -disableO0-optnone -c -emit-llvm c.cpp
opt c.bc -passes="default<O3>" -o c.bc
llvm-link a.bc b.bc c.bc from_opt.bc
```
When generating binaries, `from_clang.bc` was 10% faster than `from_opt.bc`.
To my understanding, [Clang's optimization pipeline](https://github.com/llvm/llvm-project/blob/release/16.x/clang/lib/CodeGen/BackendUtil.cpp#L992C1-L992C1) is [identical to opt's](https://github.com/llvm/llvm-project/blob/release/16.x/llvm/lib/Passes/PassBuilder.cpp#L1207) when running the new pass manager.
So, is there anything I'm missing here? How come the results are so different between these two binaries with alleged identical optimization pipelines?
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
