[ELD] Fix --wrap archive extraction for __real_<sym> references by vijayvarghese · Pull Request #1187 · qualcomm/eld

vijayvarghese · 2026-05-17T03:40:20Z

Problem

Fixes #936

When --wrap=<sym> is active and an object file references __real_<sym>,
the archive containing <sym> was not being extracted, producing:

Error: wrap.o(.text+0x15): undefined reference to `add'

Root Cause

IRBuilder::addSymbol renames __real_<sym> to <sym> in the namepool
before archive scanning begins. The existing check in shouldIncludeSymbol
only looked for __real_<sym> in the namepool — which no longer exists
under that name — so the archive member was never pulled.

Fix

In shouldIncludeSymbol, added || info->isUndef() as a second condition,
covering the case where the __real_ reference was renamed before archive
scanning.

Test

Added test/Common/standalone/WrapSymbolArchivePull/ reproducing the exact
scenario from the issue: main.o and wrap.o as direct inputs, add.o
inside a static archive, linked with --wrap=add.

quic-seaswara · 2026-05-18T15:54:26Z

  std::string realSymbol = ("__real_" + SymName).str();
  ResolveInfo *RealSymbol = m_Module.getNamePool().findInfo(realSymbol);
-  if (RealSymbol && RealSymbol->isUndef())
+  if ((RealSymbol && RealSymbol->isUndef()) || info->isUndef())


can we update tracing for wrap symbols with --trace=wrap-symbols to show this information ?

Added --trace=wrap-symbols tracing in ArchiveParser::shouldIncludeSymbol() to log archive member extraction triggered during wrapped symbol resolution. The trace now reports the wrapped symbol along with, from which archive it being pulled.

quic-seaswara · 2026-05-18T15:54:38Z

  std::string realSymbol = ("__real_" + SymName).str();
  ResolveInfo *RealSymbol = m_Module.getNamePool().findInfo(realSymbol);
-  if (RealSymbol && RealSymbol->isUndef())
+  if ((RealSymbol && RealSymbol->isUndef()) || info->isUndef())


Please update the docs too.

Added additional documentation for --wrap usage covering archive symbol extraction behavior, how __real_<sym> references can trigger archive member inclusion, and how to use --trace=wrap-symbols for debugging these cases.

quic-seaswara · 2026-05-18T15:55:16Z

+
+# Compile inputs.
+RUN: %clang %clangopts -c %p/Inputs/1.c -o %t1.main.o
+RUN: %clang %clangopts -c %p/Inputs/2.c -o %t1.wrap.o


check that using LTO also makes sure the behavior is consistent.

Added coverage for this in the latest update.

quic-seaswara · 2026-05-18T15:55:31Z

+
+#---WrapSymbolArchivePull.test------------------------ Executable ----------#
+#BEGIN_COMMENT
+# Regression test for GitHub issue #936:


Remove this line referencing github issue

quic-seaswara · 2026-05-18T15:56:15Z

  std::string realSymbol = ("__real_" + SymName).str();
  ResolveInfo *RealSymbol = m_Module.getNamePool().findInfo(realSymbol);
-  if (RealSymbol && RealSymbol->isUndef())
+  if ((RealSymbol && RealSymbol->isUndef()) || info->isUndef())


Can we make sure that _wrap is also handled accordingly ?

Added coverage for this.
Verified __wrap_<sym> handling with an additional test where __wrap_<sym> is defined inside an archive member. In this case it is included through the existing archive inclusion flow (renameMap lookup path), and the test confirms correct resolution.

quic-seaswara · 2026-05-18T15:57:17Z

  std::string realSymbol = ("__real_" + SymName).str();
  ResolveInfo *RealSymbol = m_Module.getNamePool().findInfo(realSymbol);
-  if (RealSymbol && RealSymbol->isUndef())
+  if ((RealSymbol && RealSymbol->isUndef()) || info->isUndef())


Can we also check if __real or __wrap if not defined and what the error is ?

Added test coverage for both cases.
Validated linker diagnostics for:
missing __wrap_<sym> → unresolved __wrap_<sym>
missing __real_<sym> → No archive pull needed, linking finished with no errors.

Also by default behavior verified if <sym> is not defined and there is a call to __real_<sym> will result in a Error: build/wrap.o: undefined reference to `<sym>'

quic-seaswara · 2026-05-26T21:09:38Z

+# Test that --wrap archive extraction works correctly under LTO.
+# When --wrap=add is active and wrap.o references __real_add,
+# IRBuilder renames __real_add to add in the name pool. Later,
+# when the archive is scanned, shouldIncludeSymbol() must still


remove comments not relevant for the test, test should contain only test specific comments.

quic-seaswara · 2026-05-26T21:09:56Z

+# Verify the archive member was pulled in.
+#CHECK: {{.*}}libfoo.a{{.*}}add.o
+
+#UNSUPPORTED: windows, ndk-build, riscv32, riscv64


Why they are not supported ?

quic-seaswara · 2026-05-26T21:11:31Z

  std::string realSymbol = ("__real_" + SymName).str();
  ResolveInfo *RealSymbol = m_Module.getNamePool().findInfo(realSymbol);
-  if ((RealSymbol && RealSymbol->isUndef()) || info->isUndef())
+  if ((RealSymbol && RealSymbol->isUndef()) || info->isUndef()){


should this be _real or _wrap ?

When --wrap=<sym> is active and an object file references __real_<sym>, the archive containing <sym> was not being extracted from the archive. Root cause: IRBuilder::addSymbol renames __real_<sym> to <sym> before archive scanning begins, so the archive check for '__real_<sym>' in the namepool always fails. The original symbol <sym> remains undefined in the namepool but was never checked. Fix: In shouldIncludeSymbol, also return Include when info->isUndef() is true for the archive symbol itself, covering the case where the __real_ reference was renamed before archive scanning. Added a test case that reproduces the exact scenario from the issue: main.o and wrap.o as direct inputs, add.o inside a static archive, linked with --wrap=add. Fixes qualcomm#936 Signed-off-by: Vijay Varghese <33589052+vijayvarghese@users.noreply.github.com>

- add --trace=wrap-symbols diagnostics for wrapped archive member extraction in ArchiveParser::shouldIncludeSymbol - add diagnostic template for archive pull trace output for wrapped symbols - add test coverage for undefined __real_<sym> and __wrap_<sym> cases - add test coverage to verify __wrap_<sym> resolution when defined inside an archive member - add LTO test to validate archive pulling behavior remains consistent during post-LTO symbol resolution - remove GitHub issue reference from test comments - update wrap documentation with additional details on archive symbol extraction and trace-based debugging Signed-off-by: Vijay Varghese <33589052+vijayvarghese@users.noreply.github.com>

quic-seaswara requested changes May 18, 2026

View reviewed changes

quic-seaswara requested review from parth-07 and quic-areg May 18, 2026 15:57

vijayvarghese force-pushed the fix/wrap-archive-real-sym-936 branch from 2b9fd5a to 4c8a4bc Compare May 25, 2026 08:24

vijayvarghese requested a review from quic-seaswara May 25, 2026 08:50

quic-seaswara requested changes May 26, 2026

View reviewed changes

vijayvarghese added 2 commits May 26, 2026 17:15

quic-seaswara force-pushed the fix/wrap-archive-real-sym-936 branch from 4c8a4bc to 0e375ba Compare May 26, 2026 22:15

Conversation

vijayvarghese commented May 17, 2026

Problem

Root Cause

Fix

Test

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants