[libunwind] Faster handling of frames with missed FDE records.

[vchigrin]

Some stack may have frames with missing FDE records, for example, when they implemented in assembly and author did not provide frame information manually.
In that case any attempt to unwind such stack will result in potentially long linear .eh_frame section scan.
This patch adds assumption that if .eh_frame_hdr section is present and not empty, then any frame must be present in it and we can skip long linear scan of .eh_frame section.

[llvmbot]

@llvm/pr-subscribers-libunwind

Author: Vyacheslav Chigrin (vchigrin)

Changes

Some stack may have frames with missing FDE records, for example, when they implemented in assembly and author did not provide frame information manually.
In that case any attempt to unwind such stack will result in potentially long linear .eh_frame section scan.
This patch adds assumption that if .eh_frame_hdr section is present and not empty, then any frame must be present in it and we can skip long linear scan of .eh_frame section.

---

Full diff: https://github.com/llvm/llvm-project/pull/167849.diff

1 Files Affected:

• (modified) libunwind/src/UnwindCursor.hpp (+20-3)

diff --git a/libunwind/src/UnwindCursor.hpp b/libunwind/src/UnwindCursor.hpp
index d7348254af07b..42c3cf9084219 100644
--- a/libunwind/src/UnwindCursor.hpp
+++ b/libunwind/src/UnwindCursor.hpp
@@ -1787,9 +1787,26 @@ bool UnwindCursor<A, R>::getInfoFromDwarfSection(
   }
   if (!foundFDE) {
     // Still not found, do full scan of __eh_frame section.
-    foundFDE = CFI_Parser<A>::findFDE(_addressSpace, pc, sects.dwarf_section,
-                                      sects.dwarf_section_length, 0,
-                                      &fdeInfo, &cieInfo);
+    // But only if __eh_frame_hdr is absent or empty.
+    // We assume that both sections have the same data, and don't want to waste
+    // time for long scan for absent addresses.
+    bool hasEHHeaderData = false;
+#if defined(_LIBUNWIND_SUPPORT_DWARF_INDEX)
+    if ((sects.dwarf_index_section != 0)) {
+      typename EHHeaderParser<A>::EHHeaderInfo hdrInfo;
+      const pint_t ehHdrStart = sects.dwarf_index_section;
+      const pint_t ehHdrEnd = ehHdrStart + sects.dwarf_index_section_length;
+      if (EHHeaderParser<A>::decodeEHHdr(_addressSpace, ehHdrStart, ehHdrEnd,
+                                         hdrInfo)) {
+        hasEHHeaderData = (hdrInfo.fde_count != 0);
+      }
+    }
+#endif
+    if (!hasEHHeaderData) {
+      foundFDE = CFI_Parser<A>::findFDE(_addressSpace, pc, sects.dwarf_section,
+                                        sects.dwarf_section_length, 0, &fdeInfo,
+                                        &cieInfo);
+    }
   }
   if (foundFDE) {
     if (getInfoFromFdeCie(fdeInfo, cieInfo, pc, sects.dso_base)) {

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[vchigrin]

As far as I can understand, GCC unwinding approach is similar. Check code here
https://gcc.gnu.org/git/?p=gcc.git;a=blob;f=libgcc/unwind-dw2-fde-dip.c;h=57d0c8812b13a2bfa8ff9b5e031347ad9afd7032;hb=refs/heads/releases/gcc-14#l506
seems, if their code finds suitable .eh_frame_hdr data, it uses it. And if entry not found in the table, it simply returns NULL, falling back to linear search in .eh_frame only if .eh_frame_hdr has no suitable encoding or alignment.

[vchigrin]

@kovdan01 , @atrosinenko - saw you in reviewers of one recent PR-s, touching this file. Could you either review, or redirect to appropriate reviewers for this change? Thank you in advance!

[vchigrin]

On my testing program it changed timings of construction of boost::stacktrace::stacktrace object from
4519407ns to 81666ns - that is, about 50x improvement. Of course this is depends on the many factors: how many frames has tested binary, did examined stack include frames with missed FDEs, etc.

[kovdan01]

@kovdan01 , @atrosinenko - saw you in reviewers of one recent PR-s, touching this file. Could you either review, or redirect to appropriate reviewers for this change? Thank you in advance!

@ldionne Could you please take a look at the changes and/or assign appropriate reviewers? Thanks!