Load relocation table entries as words

[maleadt]

Since #960, Metal kernels that compare a Julia object against a relocated one give the wrong answer when Metal's shader validation is enabled. Metal.jl's relocations test shows it (JuliaGPU/Metal.jl#997, the :mag: API validation job):

function kernel(out, sym::Symbol)
    @inbounds out[1] = sym === :foo ? Int32(1) : (sym === :bar ? Int32(2) : Int32(-1))
    return
end

With MTL_SHADER_VALIDATION=1 this returns -1 for both :foo and :bar; without validation it works. The bisect points at 9d808bc ("Lower relocation slots by substituting table addresses"): GPUCompiler 2.9 passes, 2.10 and 2.11 fail. This isn't related to LLVM.jl 10.

Why

With the :table lowering, which Metal uses, relocations are words that Metal.jl's loader writes into a device buffer. #960 replaced each relocation slot by the address of its table entry, but the load from that address kept the type Julia gave it. For a Symbol, that's a pointer. So the kernel ended up reading a host address out of device memory as a pointer:

; 2.10, AS 0 is thread memory in AIR
%foo = load ptr, ptr addrspace(1) %entry, align 8, !nonnull !0
%arg = inttoptr i64 %sym to ptr
%eq  = icmp eq ptr %foo, %arg

On Metal, the default address space is thread memory, and shader validation instruments loads of thread pointers. A host address read that way doesn't keep its bits. Comparing the values as integers afterwards doesn't help: it's the pointer-typed load itself. Before #960, the lowering loaded an i64, which is what the table actually holds:

; this PR (and 2.9)
%foo = load i64, ptr addrspace(1) %entry, align 8
%eq  = icmp eq i64 %foo, %sym

What this changes

The :table lowering now loads every table entry as a word, and converts it back with inttoptr where the code used it as a pointer. That puts the table's contract (it holds the words resolved_relocation_table returns) where the representation changes, instead of special-casing Metal. Identity comparisons then become integer comparisons. The rest of #960 stays:

• the address PHIs and selects, and address-space inference
• the :patch and :bake lowerings, and cglobal collection

Loads keep their alignment, volatility, atomic ordering and metadata, except for metadata that only makes sense for pointers (!nonnull, !dereferenceable, ...). The comment in src/metal.jl that called address space 0 generic now says what it is in AIR: thread memory, modeled as the flat space only for address-space inference.

We also considered fixing this in the Metal back-end, by rewriting address-space-0 loads late, or in Metal.jl. Neither knows which pointers are really host identities, so they would have to guess.

Testing

• Metal IR tests: the merged-slot test previously required load ptr, ptr addrspace(1), i.e. the bug. It now requires word loads, and a new Symbol test checks for word loads and integer comparisons (the latter on LLVM 20+, which folds icmp of two inttoptrs).
• Native tests: new ones run pointer-valued relocations under :bake, :patch and :table.
• Full suite: passes on Julia 1.12 and 1.13. The new tests fail without the fix.
• Hardware: on an M1, Metal.jl's relocations test passes with and without shader validation on 1.12 and 1.13. Metal's full suite has no new failures.

This also bumps the version to 2.11.1, which Metal.jl#997 needs to go green.

[codecov]

Codecov Report

❌ Patch coverage is 90.00000% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.73%. Comparing base (a47c2e4) to head (6cb69b9).

Files with missing lines	Patch %	Lines
src/relocation.jl	90.00%	2 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #975   +/-   ##
=======================================
  Coverage   87.72%   87.73%           
=======================================
  Files          30       30           
  Lines        6078     6097   +19     
=======================================
+ Hits         5332     5349   +17     
- Misses        746      748    +2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.