Original bug ID: 7120
Reporter: bartjacobs
Assigned to: @gasche
Status: resolved (set by @gasche on 2016-03-08T13:36:32Z)
Resolution: fixed
Priority: normal
Severity: major
Version: 4.02.3
Target version: 4.03.0+dev / +beta1
Fixed in version: 4.03.0+dev / +beta1
Category: back end (clambda to assembly)
Related to: #7118
Monitored by: @gasche @yallop

Bug description

ocamlopt (at least sometimes, perhaps always) generates an extra .cfi_adjust_cfa_offset directive between ret and .cfi_endproc at the end of an assembly routine. (.cfi_xxx directives generate DWARF stack unwinding information. This information enables debuggers such as LLDB to produce a stack backtrace even for code that does not adhere to the classical stack frame layout with base pointers. It is also used by the runtime system of certain programming languages, including C++ and Objective-C, to walk and unwind the stack for exception handling.)

This causes the .cfi_adjust_cfa_offset directives to not be pairwise balanced, i.e. this extra directive does not have a counterpart that cancels it out.

The resulting unwind info causes some programs interpreting this info (at least Apple's libunwind 35.3 on OS X 10.10.1, perhaps other consumers on other systems as well) to take the directive into account for subsequent assembly routines, causing the CFA offsets to accumulate and be incorrect for these subsequent routines.

The consequences are truncated call stacks in LLDB. But more importantly, this can lead to crashes if OCaml code calls into code that performs stack walks to find exception handlers. For example, Cocoa (the OS X GUI library) triggers such stack walks. See also bug 7118.

Steps to reproduce

Unzip the attached archive. Then:
make
lldb ./driver
break set -n camlMylib__bar_1223 # The number is non-deterministic. Check mylib.s for the right value.
run
bt # Observe that the backtrace is not correct: at this point there are really at least 10 stack frames: dylib`start -> main -> caml_main -> caml_start_program -> caml_program -> camlDriver__entry -> caml_apply11 -> camlMylib__foo1 -> caml_apply10 -> camlMylib__bar
target modules show-unwind -n camlMylib__foo1_1199
target modules show-unwind -n camlMylib__foo2_1211
target modules show-unwind -n camlMylib__bar_1223

The unwind info shown for foo1 is correct. Notice, however, that the unwind info shown for foo2 is different from that shown for foo1, even though these two OCaml functions have identical bodies and identical assembly routines. The info for foo2 is incorrect. The extra .cfi_adjust_cfa_offset value specified at the end of the assembly routine for foo1 has been accumulated onto the CFA offsets for foo2.

File attachments

• unwind-bug-repro.zip