James Karlsson

Primarily interested in compilers, formal methods, and denotational semantics.

[James Karlsson @dblejames]

at the end of the day we're all goyim who enjoy leetcode mediums

[James Karlsson @dblejames]

@ntddkh I appreciate that a lot!

[James Karlsson @dblejames]

been working on a decompiler where the recovered program is itself a typed IR

[James Karlsson @dblejames]

@pskocik tbh using an rsp clobber just to dirty the scratch space was always an awful sledgehammer. nice that prologue/epilogue gen leaves the leaf frame alone w.r.t frame pointers so CFA tracking doesnt implode just for a zero overhead ABI call

[James Karlsson @dblejames]

@dinqr0 ask them what bit in cr4 toggles smep and watch their eyes glaze over before they throw you out into the corridor

[James Karlsson @dblejames]

a girl no older than 10 came into my store and out of nowhere i heard her go "x86_64, x86_64, you can call me x86_64"

so, naturally, i walked over to her, bent down to her level and asked "are you larping, or can you tell me what prevents memory disambiguation?"

when she couldn't, i kicked her out #GateKEPT

[James Karlsson @dblejames]

@theodorvaryag infinite simulation of them explaining branch prediction

[James Karlsson @dblejames]

@theodorvaryag should be some sort of tribunal

[James Karlsson @dblejames]

@SingulCore damn dawg have you considered telling people at unknowncheats? all this just to spam sequential dword syscalls as if its 2016

[James Karlsson @dblejames]

keeping these speculative bc the same access pattern could be an array union packed object or interior pointer. the pass retains each load/store as provenance and rejects indexed addresses/phi merges unless every input resolves to the same base + displacement

[James Karlsson @dblejames]

extended on the scalar type pass with layout recovery that walks SSA def chains to normalise addresses into base + displacement then groups memory ops into field candidates keyed by offset and width. 64 bit loads reused as bases add edges between layout candidates

[James Karlsson @dblejames]

been working on a decompiler where the recovered program is itself a typed IR

[James Karlsson @dblejames]

same constraint sensitivity shows up with type boundaries too btw. canonicalise folds memref.dim to 4 through a cast while the source still carries memref<?x4xf32> but once the function argument is already memref<?x?xf32> theres no static extent left to recover

disc handles the broader problem since it carries shape constraints explicitly across lowering so later optimisation/codegen still have them https://t.co/j1M6r711Wx

[James Karlsson @dblejames]

@CaptionOfNone like for example rn I'm inferring widths from loads but I want to track accesses through the same base pointer so their offsets and widths can start to suggest a struct layout

[James Karlsson @dblejames]

@CaptionOfNone for now yeah. the idea (or at least what I'm going for) is build on scalar hints with constraints from loads/stores to recover field layouts + pointer relationships

[James Karlsson @dblejames]

@marktyrkba llvm has view-cfg though it’s basically a graphviz wrapper https://t.co/uv1XKjs1VK

[James Karlsson @dblejames]

@CaptionOfNone you'd just build SSA from the lifted IR and use operand widths + signed comparisons and shifts to infer type hints

[James Karlsson @dblejames]

@aionfork you brought silicon valley and society into it and i’m the one twisting the knife. take your medication(s) and stop interacting with me

[James Karlsson @dblejames]

@aionfork are you schizophrenic?

[James Karlsson @dblejames]

@aionfork yes i imagine thats because nobody wants to be around you (which is evident, by the way)

[James Karlsson @dblejames]

@theodorvaryag thanks chris :) i added OCaml primarily bc i wanted the recovered representation to be useful as an actual typed IR rather than just another pseudocode rendering so having something w ADTs/pattern matching makes writing transformations over it pretty natural (and less aids lmao)