llvm-project

caa18a80 - [llvm-dwp] Replace MCStreamer with direct ELF writer for zero-copy output (#192112)

[llvm-dwp] Replace MCStreamer with direct ELF writer for zero-copy output (#192112) Replace the MCStreamer-based output pipeline with a lightweight direct ELF writer (DWPWriter). Section data is stored as zero-copy StringRef chunks pointing to the mmap'd input files, and written as a minimal ELF64 relocatable object directly to disk. ## Rationale The MCStreamer pipeline copies all section data into 16KB MCDataFragment blocks, accumulates them in memory, then writes everything out during MCAssembler::Finish(). This can be cause lots of memory pressure and slow down llvm-dwp. For instance, on a 3.3GB DWP file, this translates to rougly ~3.3GB of heap allocation and two full copies of the data. The new DWPWriter avoids this via: - emitBytes() stores a StringRef chunk (zero-copy, no allocation) - emitIntValue() writes to a small per-section buffer (index tables) - writeELF() streams chunks directly from input mmap to output file - for single-input DWP files, string deduplication is also skipped since all strings are already unique. (minor optimization) Bonus: this also removes all MC library dependencies from llvm-dwp (AllTargetsCodeGens, AllTargetsDescs, AllTargetsInfos, MC, TargetParser), reducing the binary size. ## Benchmark I benchmarked on a 3.3GB production DWP file (8638 CUs, ~981MB .debug_str.dwo): Results: Before: 23.6s wall (19.6s user, 3.9s sys) After: 6.0s wall (3.0s user, 2.9s sys) **3.9x** wall time improvement, 6x fewer page faults (178K vs ~1M).