[lld] 22eba47 - [ELF] Prevent .eh_frame_hdr size oscillation in address assignment (#216248)

[via llvm-commits]

Author: Sriraman Tallam
Date: 2026-08-15T04:36:08Z
New Revision: 22eba47a1c4533e9383a422edd29d46d903a43a4

URL: https://github.com/llvm/llvm-project/commit/22eba47a1c4533e9383a422edd29d46d903a43a4
DIFF: https://github.com/llvm/llvm-project/commit/22eba47a1c4533e9383a422edd29d46d903a43a4.diff

LOG: [ELF] Prevent .eh_frame_hdr size oscillation in address assignment (#216248)

When .eh_frame_hdr size shifts, section addresses can change whether
FDEs evaluate to identical PC-relative offsets, causing FDE
deduplication count to toggle and section size to oscillate infinitely
between passes.

Prevent .eh_frame_hdr from shrinking during iterative address assignment
passes (matching .relr.dyn) to ensure monotonic convergence.

Added: 
    lld/test/ELF/eh-frame-hdr-oscillation.s

Modified: 
    lld/ELF/SyntheticSections.cpp

Removed: 
    


################################################################################
diff  --git a/lld/ELF/SyntheticSections.cpp b/lld/ELF/SyntheticSections.cpp
index 6bac881446fd0..d2891ac781617 100644
--- a/lld/ELF/SyntheticSections.cpp
+++ b/lld/ELF/SyntheticSections.cpp
@@ -491,6 +491,12 @@ bool EhFrameHeader::updateAllocSize(Ctx &ctx) {
   // Compute size.
   size_t oldSize = size;
   finalizeContents();
+
+  // Don't allow the section to shrink; otherwise the size of the section can
+  // oscillate infinitely.
+  if (size < oldSize)
+    size = oldSize;
+
   return size != oldSize;
 }
 

diff  --git a/lld/test/ELF/eh-frame-hdr-oscillation.s b/lld/test/ELF/eh-frame-hdr-oscillation.s
new file mode 100644
index 0000000000000..7fe8858dba98b
--- /dev/null
+++ b/lld/test/ELF/eh-frame-hdr-oscillation.s
@@ -0,0 +1,51 @@
+# REQUIRES: x86
+## .eh_frame_hdr's size feeds back into its own contents: it shifts downstream
+## addresses, which decides whether two FDEs share an initial location and get
+## deduplicated, which changes the FDE count and therefore the size again. Any
+## output section boundary whose alignment is comparable to the size delta can
+## close the loop.
+##
+## fz is zero-length at the end of .text.a. With a 36-byte .eh_frame_hdr, .text.a
+## ends 16-byte aligned, and fz and fb share an address. Deduplicating their
+## FDEs gives a 28-byte header, which shifts # .text.b by 8 and separates
+## them again. Clamping the size converges.
+
+# RUN: rm -rf %t && split-file %s %t && cd %t
+# RUN: llvm-mc -filetype=obj -triple=x86_64 a.s -o a.o
+# RUN: ld.lld -T lds --eh-frame-hdr a.o -o a
+# RUN: llvm-readelf -S -x .eh_frame_hdr a | FileCheck %s
+
+## The header keeps room for 3 FDEs while fde_count is 2. The trailing entry is
+## unwritten; consumers use fde_count.
+# CHECK:      .eh_frame_hdr PROGBITS 0000000000010000 001000 000024
+# CHECK:      Hex dump of section '.eh_frame_hdr':
+# CHECK-NEXT: 0x00010000 011b033b {{[0-9a-f]+}} 02000000 {{[0-9a-f]+}}
+# CHECK-NEXT: 0x00010010 {{[0-9a-f]+}} {{[0-9a-f]+}} {{[0-9a-f]+}} 00000000
+# CHECK-NEXT: 0x00010020 00000000
+
+#--- lds
+SECTIONS {
+  . = 0x10000;
+  .eh_frame_hdr : { *(.eh_frame_hdr) }
+  .eh_frame : { *(.eh_frame) }
+  .text.a : { *(.text.a) }
+  .text.b : ALIGN(16) { *(.text.b) }
+}
+
+#--- a.s
+.globl _start
+.section .text.a, "ax", @progbits
+_start:
+  .cfi_startproc
+## Sized so that .text.a ends 16-byte aligned in the 36-byte layout.
+  .space 12, 0x90
+  .cfi_endproc
+fz:
+  .cfi_startproc
+  .cfi_endproc
+
+.section .text.b, "ax", @progbits
+fb:
+  .cfi_startproc
+  nop
+  .cfi_endproc