[email protected] @HaskRay

Maintainer of LLVM MC, Clang driver, lld/ELF, and libunwind maskray.me California, USA Joined November 2012

Tweets

1K
Followers

4K
Following

311
Likes

340

[email protected] @HaskRay

25 minutes ago

`# RUN: llvm-dwarfdump --debug-names %t.bolt FileCheck --check-prefix=POSTCHECK %s` It passed even if the pipe `|` was missing: llvm::cl:: parsed --check-prefix as grouped -c -h…, hit -h, printed --help, exit 0. FileCheck never ran. No "Unknown command line argument" error.

0 0 1 81 0

View Details

[email protected] @HaskRay

2 days ago

@disruptnhandlr Claude Opus 5

0 0 0 121 0

View Details

[email protected] @HaskRay

3 days ago

An llvm-project contributor whose AI-assisted PRs had all needed heavy pushback just sent a good one: small diff, tests that fail without the fix, no padding. Approved.

2 1 44 7K 9

View Details

[email protected] @HaskRay

3 days ago

Two things that bug me about Abseil flags: * They call every option a "flag" (I'm used to flags just being boolean options). * Using --noboolflag instead of --no-boolflag

0 0 10 836 1

View Details

[email protected] @HaskRay

4 days ago

maskray.me/blog/linker-io… Linker I/O tricks and their downsides

0 10 71 4K 44

View Details

[email protected] @HaskRay

4 days ago

@RepealTCPA1947 Niri (a Wayland compositor) screenshots on Claude Code generated tables.

0 0 1 30 0

View Details

[email protected] @HaskRay

5 days ago

Linking clang 21 with lld/mold/wild at 1-16 threads on Intel Core i7-14700K (8P+12E cores). Built around Sep 12 mold & wild lose ~30-40% when 2 threads share one P-core's SMT pair vs 2 distinct cores - real parallel linkers. ld.lld shows zero SMT sensitivity.

2 1 17 2K 5

View Details

[email protected] @HaskRay

5 days ago

Linux distributions really should disable SFrame. It lacks kernel support, is effectively a dead end on AArch64 and x86-64, and IBM architectures may pivot back to .eh_frame too. Designing .sframe as an assembler/linker-level protocol is a compatibility disaster from the start.

0 0 1 133 1

View Details

[email protected] @HaskRay

5 days ago

"Relying on assembler/linker output for this critical metadata format requires a level of stability that is currently concerning." -- Me, 2025-09. The .sframe damage is ongoing github.com/rui314/mold/is… zig linker didn't implement R_X86_64_PC64 - ghostty fails to build core/glibc

1 0 0 170 1

View Details

[email protected] @HaskRay

a month ago

2017年Linus拒绝内核DWARF unwinder，那句话是有作用域的：不要把带复杂状态机的unwinder用在oops代码里。 ORC的每一条理由——不能崩在崩溃处理里、简单到没 bug、快到lockdep能用，都是为那条路径论证的。 SFrame是“用户态版ORC”。它把一套为oops路径成立的结论，整体搬到了用户态上。刻舟求剑。

1 0 31 6K 16

View Details

[email protected] @HaskRay

5 days ago

I maintain lld/ELF port in my spare time. Optimizing a mature codebase is slower and less glamorous work than building a new one. New linkers arrive with a story - a paper, a name. When I cut a third off a link that people run in production, I'd like some of that attention too.

5 8 92 5K 5

View Details

[email protected] @HaskRay

5 days ago

Output in tmpfs. --no-fork. mold: use system libmimalloc.so.3.5 --no-fork lld: LLVM_ENABLE_PIC=off, CMAKE_EXE_LINKER_FLAGS=-lmimalloc Ensured there is no fallocate / transparent-huge-pages cheat.

0 0 1 334 1

View Details

[email protected] @HaskRay

5 days ago

Linking chrome-release (--gc-sections --export-dynamic, no ICF) with lld/mold/wild

1 0 2 518 1

View Details

[email protected] @HaskRay

6 days ago

ghostty bug: the zellij selection follows the pointer on every move until the next real click. Linux, GTK on Wayland (niri), with zellij running inside. . Here is the patch. "I have not studied the patch and don't have time to take it through review". github.com/ghostty-org/gh…

0 0 9 914 0

View Details

[email protected] @HaskRay

2 weeks ago

maskray.me/blog/lld-23-el… lld 23 ELF changes

0 6 20 4K 6

View Details

[email protected] @HaskRay

2 weeks ago

When github builds the default squash message from the PR title + description, it word wraps the text at 72 characters. Many agentic coding tools don't follow the 72-character rule => you get badly wrapped commit messages.

0 0 10 669 2

View Details

[email protected] @HaskRay

3 weeks ago

@rui314 I'm getting a lot of MC and lld/ELF patches right now and the quality is very low.

2 0 0 387 0

View Details

[email protected] @HaskRay

3 weeks ago

@rui314 author only half understands, you've removed the fun part and kept the exhausting part. Unless you're the BDFL, the social norm is still "mentor, don't rewrite their patch". That norm made sense when a patch represented hours of work. It doesn't nowadays, and it hasn't adapted.

1 0 9 572 0

View Details

[email protected] @HaskRay

3 weeks ago

If you're new and want to fix a recent regression: comment on the issue before you write the patch.

0 0 0 167 0

View Details

[email protected] @HaskRay

3 weeks ago

The time from "issue filed" to "plausible patch posted" is now shorter than the time a maintainer needs to understand the bug properly. a stale-analysis crash, from a change I co-authored. A patch showed up clearing the cache at one of the sites that mutates the CFG.

2 0 9 1K 1

View Details

[email protected] @HaskRay

3 weeks ago

Locally correct, tested, but wrong layer. So the review can't be a skim. You read it fully, reconstruct the design, and conclude "no." The tool didn't change who knows which layer to fix at. It changed how fast a patch can be produced by someone who doesn't.