Use framehop for stack unwinding.

This produces much higher quality stacks than simple frame pointer
stack walking:

 - On arm64, it correctly unwinds "frameless" leaf functions, which
   don't create a frame entry. Fixes #2.
 - When sampling happens while the current instruction is in a function
   prologue or epilogue, this is detected by disassembling the
   instructions in the vicinity, and a correct stack is built.
 - For binaries without frame pointers, DWARF-based stackwalking (using
   the information in the `__eh_frame` section) is used.

Author: mstange

Readme.md

@@ -83,4 +83,8 @@ It can enumerate threads, pause them at will, and read process memory.
 
 We use these primitives to walk the stack and enumerate shared libraries.
 
-At the moment, only frame pointer stack walking is implemented. This is usually fine because keeping frame pointers is the default on macOS.
+Stack unwinding uses the [`framehop` crate](https://github.com/mstange/framehop/), which
+emits high quality stacks on both x86_64 and arm64. It supports Apple's compact unwind
+info format and DWARF CFI, and has heuristics for function prologues and epilogues. As
+a result, stacks should always be available, even for binaries that were compiled without
+frame pointers.

perfrecord/Cargo.lock

@@ -25,6 +25,7 @@ crossbeam-channel = "0.5.1"
 signal-hook = "0.3.9"
 serde_json = "1.0.53"
 gecko_profile = { version = "0.2.0", path = "../gecko_profile" }
+framehop = "0.2.0"
 
 [profile.release]
 debug = true