Ensure the personality does not panic

In a cdylib that uses std and is free from panics in the code, the panic machinery will *still* be pulled in because of the personality function when using LLD. The personality function is used by unwinding to figure out what to do when unwinding through a function. Each function that participates in unwind has an associated FDE (frame descriptor entries) in `.eh_frame`. This FDE points to a CIE (common information entry), which can reference a language-specific personality function, like `rust_eh_personality` in our case. As long as there is a CIE that references the personality, the function cannot be removed. If all references to a CIE get removed (because the functions and their associated FDEs have been removed), LLD will not remove the CIE, likely due to the ordering of its passes). Binutils ld will remove it. In the case where the CIE is still around (despite not being used), it will still reference the personality function, so that will still be around. This is not great since it's a bunch of code, but also not _that_ much. But this is where panicking comes in. Before this change, the personality function internally made use of `dyn Fn`. This caused an indirect call that LLVM was not able to analyze as guaranteed free of unwinding, even during fat LTO. This meant that an `invoke` was used, with a landing pad. In an `extern "C"` function, which the personality function is, all landing pads call `panic_cannot_unwind`, which is a `panic_nounwind`, which is, obviously, a panic. And as a panic, it pulls in *all* the panic machinery, which is very big and sad. It is also completely unnecessary, because these indirect functions do not panic, as they are just a convenient abstraction provided from the outside. By restructuring the code to remove these indirect calls, LLVM is able to fully analyze everything and see that rust_eh_personality cannot panic, and therefore remove its landing pad. With this change, exporting a panic-free function from a cdylib will only contain the function and the personality (when linked with LLD at least, with binutils ld it will only contain the function), with no panic code being present at all, which is great.

Author: Noratrieb

library/std/src/sys/personality/gcc.rs

@@ -335,8 +335,7 @@ unsafe fn find_eh_action(context: *mut uw::_Unwind_Context) -> Result<EHAction,
  // `ip = -1` has special meaning, so use wrapping sub to allow for that
  ip: if ip_before_instr != 0 { ip } else { ip.wrapping_sub(1) },
  func_start: uw::_Unwind_GetRegionStart(context),
- get_text_start: &|| uw::_Unwind_GetTextRelBase(context),
- get_data_start: &|| uw::_Unwind_GetDataRelBase(context),
+ raw_context: context,
  };
  eh::find_eh_action(lsda, &eh_context)
  }

library/std/src/sys/personality/gcc/eh.rs

@@ -14,6 +14,8 @@
 
 use core::ptr;
 
+use unwind as uw;
+
 use super::dwarf::DwarfReader;
 
 pub const DW_EH_PE_omit: u8 = 0xFF;
@@ -37,11 +39,10 @@ pub const DW_EH_PE_aligned: u8 = 0x50;
 pub const DW_EH_PE_indirect: u8 = 0x80;
 
 #[derive(Copy, Clone)]
-pub struct EHContext<'a> {
- pub ip: *const u8, // Current instruction pointer
- pub func_start: *const u8, // Pointer to the current function
- pub get_text_start: &'a dyn Fn() -> *const u8, // Get pointer to the code section
- pub get_data_start: &'a dyn Fn() -> *const u8, // Get pointer to the data section
+pub struct EHContext {
+ pub(crate) ip: *const u8, // Current instruction pointer
+ pub(crate) func_start: *const u8, // Pointer to the current function
+ pub(crate) raw_context: *mut uw::_Unwind_Context,
 }
 
 /// Landing pad.
@@ -63,7 +64,7 @@ pub enum EHAction {
 pub const USING_SJLJ_EXCEPTIONS: bool =
  cfg!(all(target_vendor = "apple", not(target_os = "watchos"), target_arch = "arm"));
 
-pub unsafe fn find_eh_action(lsda: *const u8, context: &EHContext<'_>) -> Result<EHAction, ()> {
+pub unsafe fn find_eh_action(lsda: *const u8, context: &EHContext) -> Result<EHAction, ()> {
  if lsda.is_null() {
  return Ok(EHAction::None);
  }
@@ -224,7 +225,7 @@ unsafe fn read_encoded_offset(reader: &mut DwarfReader, encoding: u8) -> Result<
 /// [LSB-dwarf-ext]: https://refspecs.linuxfoundation.org/LSB_5.0.0/LSB-Core-generic/LSB-Core-generic/dwarfext.html
 unsafe fn read_encoded_pointer(
  reader: &mut DwarfReader,
- context: &EHContext<'_>,
+ context: &EHContext,
  encoding: u8,
 ) -> Result<*const u8, ()> {
  if encoding == DW_EH_PE_omit {
@@ -241,8 +242,8 @@ unsafe fn read_encoded_pointer(
  }
  context.func_start
  }
- DW_EH_PE_textrel => (*context.get_text_start)(),
- DW_EH_PE_datarel => (*context.get_data_start)(),
+ DW_EH_PE_textrel => unsafe { uw::_Unwind_GetTextRelBase(context.raw_context) },
+ DW_EH_PE_datarel => unsafe { uw::_Unwind_GetDataRelBase(context.raw_context) },
  // aligned means the value is aligned to the size of a pointer
  DW_EH_PE_aligned => {
  reader.ptr = reader.ptr.with_addr(round_up(reader.ptr.addr(), size_of::<*const u8>())?);

tests/run-make/panic-free-cdylib/add.rs

@@ -0,0 +1,5 @@
+#![crate_type = "cdylib"]
+
+pub extern "C" fn add(a: u64, b: u64) -> u64 {
+ a + b
+}

tests/run-make/panic-free-cdylib/rmake.rs

@@ -0,0 +1,25 @@
+// This ensures that a cdylib that uses std and panic=unwind but does not
+// have any panics itself will not have *any* panic-related code in the final
+// binary, at least when using fat LTO
+// (since all the necessary nounwind propagation requires fat LTO).
+//
+// This code used to be pulled in via a landing pad in the personality function,
+// (since that is `extern "C"` and therefore panics if something unwinds), so
+// if this failed because you modified the personality function, ensure it contains
+// no potentially unwinding calls.
+
+use run_make_support::{llvm_nm, rustc};
+
+fn main() {
+ rustc().input("add.rs").edition("2024").lto("fat").opt_level("3").run();
+
+ llvm_nm()
+ .input("libadd.so")
+ .run()
+ // a collection of panic-related strings. if this appears in the output
+ // for other reasons than having panic symbols, I am sorry.
+ .assert_stdout_not_contains("panic")
+ .assert_stdout_not_contains("addr2line")
+ .assert_stdout_not_contains("backtrace")
+ .assert_stdout_not_contains("gimli");
+}