clang runs loop-idiom-vectorize at -O0 for AArch64, RISCV, and Hexagon

[arsenm]

clang is apparently running the loop vectorizer at -O0, but only when targeting aarch64.

clang++ -O0 -target aarch64-pc-linux-gnu -c -o /dev/null -emit-llvm  -mllvm -print-pipeline-passes /tmp/test.cpp
verify,function(ee-instrument<>),always-inline,function(loop(loop-idiom-vectorize)),coro-cond(coro-early,cgscc(coro-split),coro-cleanup,globaldce),function(annotation-remarks),verify,BitcodeWriterPass

x86 is not running this pass:

clang++ -O0 -target x86_64-pc-linux-gnu -c -o /dev/null -emit-llvm  -mllvm -print-pipeline-passes /tmp/test.cpp
verify,function(ee-instrument<>),always-inline,coro-cond(coro-early,cgscc(coro-split),coro-cleanup,globaldce),function(annotation-remarks),verify,BitcodeWriterPass

Same for RISCV. Hexagon also runs extra passes, function(loop(hexagon-loop-idiom)),function(loop(hexagon-vlcr))

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Matt Arsenault (arsenm)

clang is apparently running the loop vectorizer at -O0, but only when targeting aarch64.

clang++ -O0 -target aarch64-pc-linux-gnu -c -o /dev/null -emit-llvm  -mllvm -print-pipeline-passes /tmp/test.cpp
verify,function(ee-instrument<>),always-inline,function(loop(loop-idiom-vectorize)),coro-cond(coro-early,cgscc(coro-split),coro-cleanup,globaldce),function(annotation-remarks),verify,BitcodeWriterPass

x86 is not running this pass:

clang++ -O0 -target x86_64-pc-linux-gnu -c -o /dev/null -emit-llvm  -mllvm -print-pipeline-passes /tmp/test.cpp`
verify,function(ee-instrument<>),always-inline,coro-cond(coro-early,cgscc(coro-split),coro-cleanup,globaldce),function(annotation-remarks),verify,BitcodeWriterPass

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: Matt Arsenault (arsenm)

clang is apparently running the loop vectorizer at -O0, but only when targeting aarch64.

clang++ -O0 -target aarch64-pc-linux-gnu -c -o /dev/null -emit-llvm  -mllvm -print-pipeline-passes /tmp/test.cpp
verify,function(ee-instrument<>),always-inline,function(loop(loop-idiom-vectorize)),coro-cond(coro-early,cgscc(coro-split),coro-cleanup,globaldce),function(annotation-remarks),verify,BitcodeWriterPass

x86 is not running this pass:

clang++ -O0 -target x86_64-pc-linux-gnu -c -o /dev/null -emit-llvm  -mllvm -print-pipeline-passes /tmp/test.cpp`
verify,function(ee-instrument<>),always-inline,coro-cond(coro-early,cgscc(coro-split),coro-cleanup,globaldce),function(annotation-remarks),verify,BitcodeWriterPass

[topperc]

The code that adds it to the pipeline is here

void AArch64TargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {       
                                                                                 
  PB.registerLateLoopOptimizationsEPCallback(                                    
      [=](LoopPassManager &LPM, OptimizationLevel Level) {                       
        LPM.addPass(LoopIdiomVectorizePass());                                   
      });                                                                        
  if (getTargetTriple().isOSWindows())                                           
    PB.registerPipelineEarlySimplificationEPCallback(                            
        [](ModulePassManager &PM, OptimizationLevel, ThinOrFullLTOPhase) {       
          PM.addPass(LowerIFuncPass());                                          
        });                                                                      
}