Append lld --stack-first on wasm targets

[zigazeljko]

Closes#4496

[fengb]

Comparison:

; Before
(global(;stack_start;) (muti32) (i32.const66576))
(global(;quuz;)i32 (i32.const1028))
; After
(global(;stack_start;) (muti32) (i32.const65536))
(global(;quuz;)i32 (i32.const65540))

Looks good! 👍

[fengb]

I think the CI failure is exposing an existing bug.

 0: failed to invoke `_start`
1: wasm trap: out of bounds memory access, source location: @856d24

This looks like a stack overflow. The current output "pads the stack" with globals — we accidentally consume too much stack space but it's simply stealing/corrupting memory from globals.

[kubkon]

I think the CI failure is exposing an existing bug.

 0: failed to invoke `_start`
1: wasm trap: out of bounds memory access, source location: @856d24

This looks like a stack overflow. The current output "pads the stack" with globals — we accidentally consume too much stack space but it's simply stealing/corrupting memory from globals.

Interesting. I've noticed the same buggy behaviour when troubleshooting the disabled tests in std.packed_int_array, and it definitely looks like a stack overflow. For the record, this bug can easily be replicated with increasing the quota with @setEvalBranchQuota and inlining some loop in comptime.

[kubkon]

Looks great!

I'm not sure what others reckon but IMHO skipping the failing tests in Wasm is OK for now until we figure out what the exact cause is and how to rectify it (a subject of a subsequent PR I guess?). @fengb thoughts?

[fengb]

The medium term solution is probably adding a manual stack size override: #3735 (comment). In the meantime, I think skipping it is okay.

[kubkon]

FYI, the CI failure should go away after we merge in #5529 which increases the default stack size to 1MB.

[zigazeljko]

Since #5529 got merged, can this get merged too?

[kubkon]

SGTM! Could you rebase to the current upstream and see if the tests now pass though?

[zigazeljko]

@kubkon Done. Looks like tests are now passing.