Leandro Santiago
2013-Aug-08 20:26 UTC
[LLVMdev] Inserting C++ code on tablegen to improve code generation statistics
Hello everyone. I'm working in a project and my task right now is improving LLVM code generation's statistics for X86 target. Currently you can show statistics by passing the command-line parameter -stats to any llvm program (I'm using llc to compile .bc to native code) and it'll show something like this: ===-------------------------------------------------------------------------== ... Statistics Collected ... ===-------------------------------------------------------------------------== 416 asm-printer - Number of machine instrs printed 1 codegen-dce - Number of dead instructions deleted 4 codegenprepare - Number of GEPs converted to casts 20 dagcombine - Number of dag nodes combined 11 isel - Number of blocks selected using DAG 11 isel - Number of entry blocks encountered 487 isel - Number of times dag isel has to try another path 544 pei - Number of bytes used for stack in all functions 5 regalloc - Number of cross class joins performed 28 regalloc - Number of identity moves eliminated after rewriting 8 regalloc - Number of instructions deleted by DCE 10 regalloc - Number of instructions re-materialized 7 regalloc - Number of interval joins performed 36 regalloc - Number of registers assigned 2 twoaddrinstr - Number of two-address instructions 6 x86-codegen - Number of floating point instructions 2 x86-isel - Number of loads moved below TokenFactor I aim to improve the x86-codegen and asm-printer sections, detailing and adding more informations on it. I'll need to add the number of every kind of generated assembly instructions (x add, y movl, etc.). Reading briefly the X86 target code, I realized these instructions are defined on tablegen .td files, so I have two ways: write c++ code on .td files, which I don't believe it's possible (otherwise, I really'd like to know how to do that); or changing the tablegen program allowing it to generate c++ code with the statistics enabled. After the contextualizarion, I'd like to know your opinion about the viability of each above alternative. Perhaps (and probably) there's another way to to this, so I'm open to suggestions :-) Thanks in advance! -- Sent from my mind -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130808/865276ee/attachment.html>
Seemingly Similar Threads
- [LLVMdev] Very slow performance of lli on x86
- [LLVMdev] Doubts about register interferences in register allocators
- [LLVMdev] Very slow performance of lli on x86
- [LLVMdev] Inserting a branch in PPCTargetLowering::LowerFormalArguments_SVR4
- [LLVMdev] Poor register allocation (constants causing spilling)