Summary

The callback dispatcher (both AMD64 and ARM64) calls callbackWrap directly via CALL/BL, bypassing the crosscall2 → runtime·load_g → runtime·cgocallback chain that purego and Go runtime use for proper C-to-Go transitions.

This works correctly when callbacks arrive on Go-managed threads (the primary WebGPU use case), but will crash if a C library invokes the callback from its own internally-created thread (G = nil).

Current behavior

• Callbacks on Go threads: works (goroutine pointer already loaded)
• Callbacks on C threads: crash (SIGSEGV, nil goroutine pointer)

Expected behavior

Callbacks should work regardless of which thread invokes them, matching purego's behavior.

Potential impact without crosscall2

Solution

Replace direct BL ·callbackWrap / CALL ·callbackWrap with BL crosscall2 / CALL crosscall2 in both dispatchers, following the purego pattern:

// ARM64 (purego reference):
MOVD ·callbackWrap_call(SB), R0   // fn (ABIInternal closure)
MOVD R13, R1                       // frame (&callbackArgs)
MOVD $0, R3                        // ctxt = 0
BL   crosscall2(SB)

Files to change

• ffi/callback_amd64.s — AMD64 dispatcher
• ffi/callback_arm64.s — ARM64 dispatcher
• ffi/callback.go — add callbackWrap_call closure (AMD64)
• ffi/callback_arm64.go — add callbackWrap_call closure (ARM64)

Context

Discovered during ARM64 trampoline fix (BL→MOVD+B, see #15). The direct call approach was inherited from the initial implementation and is documented in TASK-012-crosscall2-integration.md.