如何用LLVM工具复现clang的-O2优化？从foo.ll生成optimized.s的正确步骤

Great question — the reason opt -S -O2 foo.ll; llc optimized.ll doesn't match the output of clang -O2 is that clang does far more than just run opt and llc with basic -O2 flags. It includes frontend-specific optimizations, target-tailored parameters, and a curated optimization pipeline that plain opt/llc calls miss entirely. Here's how to properly replicate the exact result:

Step 1: Generate a Truly Unoptimized LLVM IR #

First, ensure you start with the same unoptimized baseline IR that clang uses for -O2. The default clang -S -emit-llvm foo.c might include minor implicit optimizations, so explicitly use -O0 to disable all optimizations:

clang -O0 -S -emit-llvm foo.c -o foo.ll

Step 2: Capture clang's Exact Toolchain Arguments #

Clang passes a ton of critical target-specific flags, pass manager settings, and architecture-specific optimization options to opt and llc under the hood. To see these, run clang in verbose mode when generating optimized assembly:

clang -O2 -S foo.c -o optimized_ref.s -v

This will print the full commands clang uses to call its frontend (cc1), opt, and llc. Focus on the lines starting with the path to opt and llc — these will contain all the parameters you need to replicate the pipeline.

Step 3: Manually Run opt with clang's Exact Parameters #

From the verbose output, copy all the arguments clang passes to opt, then run it on your unoptimized foo.ll. For example, if clang's opt command looks like:

/path/to/opt -O2 -mtriple=x86_64-pc-linux-gnu -mcpu=skylake -enable-new-pm=1 -passes=... foo.bc -o optimized.bc

Adjust it to use your foo.ll as input:

opt -O2 -mtriple=x86_64-pc-linux-gnu -mcpu=skylake -enable-new-pm=1 -passes=... foo.ll -o optimized.ll

This ensures you're running the exact same optimization pipeline clang uses, including target-specific passes and pass manager configurations.

Step 4: Manually Run llc with clang's Exact Parameters #

Next, repeat the process for llc. Grab the full set of arguments from the verbose output (like -mtriple, -mcpu, -mattr, and codegen flags) and run it on your optimized IR:

llc -mtriple=x86_64-pc-linux-gnu -mcpu=skylake -filetype=asm optimized.ll -o optimized.s

Why the Basic opt/llc Call Fails #

The key missing pieces in a plain opt -O2 call are:

• Target-specific flags: -mtriple, -mcpu, and -mattr tell LLVM about the target architecture, enabling architecture-specific optimizations (like AVX for x86) that clang automatically includes.
• Pass manager settings: Newer LLVM versions use the new pass manager, which clang enables with specific flags — plain opt -O2 might default to the legacy manager, leading to different optimization results.
• Pipeline alignment: clang coordinates the order and configuration of passes to work seamlessly with its frontend optimizations, which a generic opt call doesn't replicate.

Just make sure you use the same version of opt and llc that comes with your clang — mixing versions can lead to mismatches in pass behavior.

内容的提问来源于stack exchange，提问作者hsivonen