Remove potential unitialized memory in the GC stack

[gbaraldi]

Fixes #60622

This was a very messy thing. But apparently the code we generate for value_to_pointer is liable to being transformed by LLVM, and under some transformations specifically LICM we can materialize a root that didn't exist at codegen time and that root is liable to being being undef which can cause segfaults
This conservativaly initializes the gc pointers to zero.

[gbaraldi]

Only issue was. Writing a test for this was almost impossible.

[topolarity]

Should be as simple as:

    jl_datatype_t *dt = (jl_datatype_t *)v.typ;
    bool hasptr = dt->layout->first_ptr >= 0;
    size_t npointers = hasptr ? dt->layout->npointers : 0;
    auto InsertPoint = ctx.builder.saveIP();
    ctx.builder.SetInsertPoint(ctx.topalloca->getParent(), ++ctx.topalloca->getIterator());
    for (size_t i = 0; i < npointers; i++) {
        // make sure these are nullptr early from LLVM's perspective, in case it decides to SROA it
        Value *ptr_field = emit_ptrgep(ctx, loc, jl_ptr_offset(dt, i) * sizeof(void *));
        StoreInst *store = ctx.builder.CreateAlignedStore(
            Constant::getNullValue(ctx.types().T_prjlvalue), ptr_field, Align(sizeof(void *)));
    }
    ctx.builder.restoreIP(InsertPoint);

although you can get fancier if you want to coalesce adjacent roots to a single memset.

This logic should probably be moved to cgutils.cpp and implemented more similar to split_value_into / undef_derived_strct

[topolarity]

Writing a test for this was almost impossible.

It might be a reasonable addition to the GCinvariantPass?
Any alloca containing a ptr(10) should set that memory to 0 in the first basic block IIUC.

edit: Gabriel pointed out that DCE and (post-)dominating stores / loads will break that check, so that won't work sadly.

[gbaraldi]

Addressed the review feedback - simplified the code to use Julia's datatype layout information (jl_ptr_offset) instead of walking LLVM types with TrackCompositeType, as suggested.