builder: increase stack size margin for automatic stack allocation

[digitalentity]

When TinyGo calculates stack sizes automatically, it performs a static analysis of the call graph. While this analysis correctly identifies the stack requirements for Go functions, it cannot account for stack usage within assembly-level routines that are not visible to the high-level analysis.

Specifically, tinygo_swapTask pushes callee-saved registers onto the current goroutine's stack during every context switch. If this overhead is not accounted for, a goroutine might overflow its stack during a context switch even if its Go-level stack usage is within limits.

With this change in determineStackSizes, we now look up the tinygo_swapTask function in the symbol table to determine its frame size. This context switch overhead is now added to the calculated stack size for all goroutines that have a bounded stack size.

Fixes #5375

[Copilot]

Pull request overview

This PR improves TinyGo’s automatic goroutine stack size calculation by adding a safety margin for context switches, accounting for stack usage in tinygo_swapTask that is not visible to Go-level call graph analysis.

Changes:

• Look up tinygo_swapTask in the computed call graph/symbol map and capture its frame size as context-switch overhead.
• Add the context-switch overhead to all goroutine stack sizes that are compile-time Bounded.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[eliasnaur]

Add "Fixes #5375"

Otherwise, @aykevl or @dgryski should probably take a look.

[aykevl]

This is the wrong fix. The stack size of the goroutine swap functions should be taken care of already, since the assembly lists it as part of the CFA:

tinygo/src/internal/task/task_stack_cortexm.S

Line 93 in 1a1506e

.cfi_def_cfa_offset 9*4

If it isn't taken into account, that's a different bug.

[digitalentity]

@aykevl the bug is not that the .cfi_def_cfa_offset is not defined, the bug is that tinygo_swapTask is not accounted for by the stack size calculation because it's simply not part of the call tree of the goroutine. This fix is to account for it explicitly.

[digitalentity]

In #5375 (comment) I've figured out that the crashes are actually happening because of the memory corruption which sometimes gets caught by the canary, but sometimes is silent.

With the increase of the autodetected stack size by 36 bytes (tinygo_swapTask frame size) as the only change the crashes were fixed - the code that was previously crashing after 5-15 minutes was tested to run for over 12h, so clearly the autodetected stack size was incorrect.

[aykevl]

Ah sorry, I misread the PR. Will take another look later!