[LLD] Make lld define __eh_frame_* symbols

[quic-k]

This patch makes LLD explicitly define the following symbols during linking:

• __eh_frame_start
• __eh_frame_end
• __eh_frame_hdr_start
• __eh_frame_hdr_end

These are expected by libunwind for exception handling in baremetal builds
Right now, these symbols need to be explicitly defined by the linker script.
I think its better that the linker defines them implicitly, so we don't have to create a linker script only because we needed to provide these symbols

I am interested in feedback on this change

[llvmbot]

@llvm/pr-subscribers-lld

@llvm/pr-subscribers-lld-elf

Author: None (quic-k)

Changes

This patch makes LLD explicitly define the following symbols during linking:

• __eh_frame_start
• __eh_frame_end
• __eh_frame_hdr_start
• __eh_frame_hdr_end

These are expected by libunwind for exception handling in baremetal builds
Right now, these symbols need to be explicitly defined by the linker script.
I think its better that the linker defines them implicitly, so we don't have to create a linker script only because we needed to provide these symbols

I am interested in feedback on this change

---

Full diff: https://github.com/llvm/llvm-project/pull/162638.diff

2 Files Affected:

• (modified) lld/ELF/Writer.cpp (+6)
• (added) lld/test/ELF/eh-frame-syms.s (+25)

diff --git a/lld/ELF/Writer.cpp b/lld/ELF/Writer.cpp
index 4fa80397cbfa7..145bff5007466 100644
--- a/lld/ELF/Writer.cpp
+++ b/lld/ELF/Writer.cpp
@@ -2269,6 +2269,12 @@ template <class ELFT> void Writer<ELFT>::addStartEndSymbols() {
   define("__preinit_array_start", "__preinit_array_end", ctx.out.preinitArray);
   define("__init_array_start", "__init_array_end", ctx.out.initArray);
   define("__fini_array_start", "__fini_array_end", ctx.out.finiArray);
+  // Define __eh_frame_* symbols, libunwind needs these symbols.
+  // Define them only if the corresponding output section exists.
+  if (OutputSection *sec = findSection(ctx, ".eh_frame"))
+    define("__eh_frame_start", "__eh_frame_end", sec);
+  if (OutputSection *sec = findSection(ctx, ".eh_frame_hdr"))
+    define("__eh_frame_hdr_start", "__eh_frame_hdr_end", sec);
 
   // As a special case, don't unnecessarily retain .ARM.exidx, which would
   // create an empty PT_ARM_EXIDX.
diff --git a/lld/test/ELF/eh-frame-syms.s b/lld/test/ELF/eh-frame-syms.s
new file mode 100644
index 0000000000000..074221ade4fb6
--- /dev/null
+++ b/lld/test/ELF/eh-frame-syms.s
@@ -0,0 +1,25 @@
+## Verify that lld defines __eh_frame_start/end and __eh_frame_hdr_start/end symbols
+
+# REQUIRES: x86
+
+# RUN: llvm-mc -filetype=obj -triple=x86_64-unknown-linux %s -o %t.o
+# RUN: ld.lld %t.o -eh-frame-hdr -o %t
+# RUN: llvm-readelf -s %t | FileCheck %s --check-prefix=SYM
+
+# SYM: __eh_frame_start
+# SYM: __eh_frame_end
+# SYM: __eh_frame_hdr_start
+# SYM: __eh_frame_hdr_end
+
+.text
+.globl _start
+.type _start, @function
+_start:
+  nop
+
+# Emit references so the link will fail if these are not defined
+check_symbol:
+  .quad __eh_frame_start
+  .quad __eh_frame_end
+  .quad __eh_frame_hdr_start
+  .quad __eh_frame_hdr_end

[smithp35]

The __eh_frame_* symbols are equivalent to __exidx_start and __exidx_end so this seems reasonable to me.

llvm libunwind reference with suggested linker script fragment (https://github.com/llvm/llvm-project/blob/main/libunwind/src/AddressSpace.hpp#L75)

[quic-k]

yes, with this patch we won't have to use that fragment just to provide these symbols

IIUC __exidx_start/end are ARM's version of __eh_frame_start/end, right?
does lld define them?

[smithp35]

Yes, just below on line 2281. I think I added that before we had good linker script support, could also be that some unwinders even on Linux might use those symbols.

[MaskRay]

These symbols not even in GNU ld's internal linker script. I think you should explicitly define them in linker scripts:

.eh_frame : {
  PROVIDE(__eh_frame_start = .);
  *(.eh_frame)
  PROVIDE(__eh_frame_end = .);
}

[MaskRay]

.