qualcomm-linux · koenkooi · Jun 17, 2026 · Jun 17, 2026 · Jun 18, 2026
@@ -0,0 +1,71 @@
+From 27ed4c1275586113e84d4b79b47febcef4c084d3 Mon Sep 17 00:00:00 2001
+From: Koen Kooi <koen.kooi@oss.qualcomm.com>
+Date: Thu, 18 Jun 2026 12:00:00 +0200
+Subject: [PATCH] cmake: fix GCC 16 build (-Wmaybe-uninitialized,
+ -Wfree-nonheap-object)
+
+onnxruntime 1.26.0 builds with C++20 and sets COMPILE_WARNING_AS_ERROR
+ON (a bare -Werror), so two GCC 16 false-positive diagnostics become
+fatal:
+
+1) -Wmaybe-uninitialized fires in the optimizer, e.g.
+   matmul_nbits_fusion.cc, reading a value GCC 16 cannot prove is
+   initialized. onnxruntime already suppresses this for GCC 12; GCC 16
+   reintroduced the false positive, so extend the same guard to cover
+   GCC >= 16.
+
+2) -Wfree-nonheap-object fires on std::vector destructors, e.g.
+   vector<graph_utils::EdgeEndToMatch> in skip_layer_norm_fusion.cc:
+   GCC 16 mistakes the sized-deallocation of the vector's heap storage
+   for a delete on an offset pointer. Disable it, gated behind
+   check_cxx_compiler_flag.
+
+Upstream-Status: Pending
+
+Signed-off-by: Koen Kooi <koen.kooi@oss.qualcomm.com>
+Assisted-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
+---
+ cmake/CMakeLists.txt | 13 ++++++++++++-
+ 1 file changed, 12 insertions(+), 1 deletion(-)
+
+diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt
+index 83d1751..22ae6cf 100644
+--- a/cmake/CMakeLists.txt
++++ b/cmake/CMakeLists.txt
+@@ -632,6 +632,8 @@ else()
+   check_cxx_compiler_flag(-Wambiguous-reversed-operator HAS_AMBIGUOUS_REVERSED_OPERATOR)
+   # -Winterference-size was added in GCC 13
+   check_cxx_compiler_flag(-Winterference-size HAS_INTERFERENCE_SIZE)
++  # -Wfree-nonheap-object false-positives on std::vector dtors with GCC 16
++  check_cxx_compiler_flag(-Wfree-nonheap-object HAS_FREE_NONHEAP_OBJECT)
+   check_cxx_compiler_flag(-Warray-bounds HAS_ARRAY_BOUNDS)
+   check_cxx_compiler_flag(-Wbitwise-instead-of-logical HAS_BITWISE_INSTEAD_OF_LOGICAL)
+   check_cxx_compiler_flag(-Wcast-function-type HAS_CAST_FUNCTION_TYPE)
+@@ -721,6 +723,13 @@ else()
+     # Specifically, the code was tuned based on the build machine's CPU cache line size, which can vary a lot.
+     list(APPEND ORT_WARNING_FLAGS -Wno-interference-size)
+   endif()
++  if(HAS_FREE_NONHEAP_OBJECT)
++    # GCC 16 false-positives -Wfree-nonheap-object on std::vector
++    # destructors (e.g. vector<graph_utils::EdgeEndToMatch> in
++    # skip_layer_norm_fusion.cc): it mistakes the sized-deallocation of
++    # heap storage for a delete on an offset pointer.
++    list(APPEND ORT_WARNING_FLAGS -Wno-free-nonheap-object)
++  endif()
+   # enable warning(s) that may not be on by default
+   if (HAS_SHORTEN_64_TO_32)
+     list(APPEND ORT_WARNING_FLAGS -Wshorten-64-to-32)
+@@ -1187,7 +1196,9 @@ function(onnxruntime_set_compile_flags target_name)
+     foreach(FLAG ${ORT_WARNING_FLAGS})
+       target_compile_options(${target_name} PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:${FLAG}>")
+     endforeach()
+-    if("${CMAKE_C_COMPILER_ID}" STREQUAL "GNU" AND CMAKE_C_COMPILER_VERSION VERSION_LESS 13 AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 12)
++    # GCC 12 and GCC 16 both emit false-positive -Wmaybe-uninitialized in the
++    # optimizer (e.g. matmul_nbits_fusion.cc reads a value GCC cannot prove init).
++    if("${CMAKE_C_COMPILER_ID}" STREQUAL "GNU" AND ((CMAKE_C_COMPILER_VERSION VERSION_LESS 13 AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 12) OR CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 16))
+       target_compile_options(${target_name} PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:-Wno-maybe-uninitialized>")
+     endif()
+     if (onnxruntime_USE_CUDA)
+-- 
+2.34.1
+
@@ -18,25 +18,29 @@ DEPENDS = " \
     microsoft-gsl \
 "
 
+# Dependency versions and revisions are taken from cmake/deps.txt at the
+# rel-1.26.0 tag. They must match what onnxruntime expects, since the build
+# runs with FETCHCONTENT_FULLY_DISCONNECTED=ON.
 SRC_URI = "gitsm://github.com/microsoft/onnxruntime.git;protocol=https;branch=rel-${PV};name=ort \
-    git://github.com/HowardHinnant/date.git;protocol=https;nobranch=1;name=date;tag=v3.0.4;destsuffix=date \
-    git://github.com/boostorg/mp11.git;protocol=https;nobranch=1;name=mp11;tag=boost-1.91.0;destsuffix=mp11 \
+    file://0001-cmake-fix-GCC-16-build.patch \
+    git://github.com/HowardHinnant/date.git;protocol=https;nobranch=1;name=date;tag=v3.0.1;destsuffix=date \
+    git://github.com/boostorg/mp11.git;protocol=https;nobranch=1;name=mp11;tag=boost-1.82.0;destsuffix=mp11 \
     git://github.com/pytorch/cpuinfo.git;protocol=https;nobranch=1;name=pytorch_cpuinfo;destsuffix=pytorch_cpuinfo \
     git://github.com/dcleblanc/SafeInt.git;protocol=https;nobranch=1;name=safeint;tag=3.0.28;destsuffix=safeint \
     git://github.com/google/flatbuffers.git;protocol=https;nobranch=1;name=flatbuffers;tag=v23.5.26;destsuffix=flatbuffers \
-    git://github.com/onnx/onnx.git;protocol=https;nobranch=1;name=onnx;tag=v1.20.1;destsuffix=onnx \
+    git://github.com/onnx/onnx.git;protocol=https;nobranch=1;name=onnx;tag=v1.21.0;destsuffix=onnx \
     git://github.com/eigen-mirror/eigen.git;protocol=https;nobranch=1;name=eigen3;destsuffix=eigen3 \
 "
 
 SRCREV_FORMAT = "ort_date_mp11_pytorch_cpuinfo_safeint_flatbuffers_onnx_eigen3"
-SRCREV_ort              = "470ae16099a74fe05e31f2530489332c0525edb5"
-SRCREV_date             = "f94b8f36c6180be0021876c4a397a054fe50c6f2"
-SRCREV_mp11             = "b94b089d4ec83cd397f20958f34edf25bc3e06f4"
-SRCREV_pytorch_cpuinfo  = "ea6b9f1bb6e1001d8b21574d5bc78ddef62e499d"
+SRCREV_ort              = "8c546c37b43caaca1fa25db430dab94b901cf277"
+SRCREV_date             = "6e921e1b1d21e84a5c82416ba7ecd98e33a436d0"
+SRCREV_mp11             = "0a0b5fb001ce0233ae3a6f99d849c0649e5a7361"
+SRCREV_pytorch_cpuinfo  = "403d652dca4c1046e8145950b1c0997a9f748b57"
 SRCREV_safeint          = "4cafc9196c4da9c817992b20f5253ef967685bf8"
 SRCREV_flatbuffers      = "c20d64b8de759423af61e072fcabf916c1f7bf9f"
-SRCREV_onnx             = "d3f6b795aedb48eaecc881bf5e8f5dd6efbe25b3"
-SRCREV_eigen3           = "549bf8c75b6aae071cde2f28aa48f16ee3ae60b0"
+SRCREV_onnx             = "be2b5fde82d9c8874f3d19328bdfe3b6962dc67b"
+SRCREV_eigen3           = "1d8b82b0740839c0de7f1242a3585e3390ff5f33"
 
 # Fix buildpaths QA issue: remap TMPDIR references in both debug info and
 # string literals embedded in the compiled libraries.
@@ -49,6 +53,7 @@ OECMAKE_SOURCEPATH = "${S}/cmake"
 
 EXTRA_OECMAKE = " \
     -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+    -DCMAKE_CXX_STANDARD=20 \
     -DCMAKE_FIND_ROOT_PATH=${STAGING_DIR_TARGET} \
     -Donnxruntime_BUILD_SHARED_LIB=ON \
     -DFETCHCONTENT_FULLY_DISCONNECTED=ON \

@@ -0,0 +1,46 @@
+From 158c361a47de44003ff79a3f682319eca14db252 Mon Sep 17 00:00:00 2001
+From: Ricardo Salveti <ricardo.salveti@oss.qualcomm.com>
+Date: Tue, 16 Jun 2026 15:36:16 +0000
+Subject: [PATCH] tflite: examples: label_image: build profile_buffer.cc
+
+The label_image example uses BufferedProfiler (via profiler.h /
+buffered_profiler.h), whose EndEvent()/BeginEvent() forward to
+tflite::profiling::ProfileBuffer. ProfileBuffer is declared in
+profiling/profile_buffer.h and defined in profiling/profile_buffer.cc,
+which is deliberately excluded from libtensorflow-lite.a
+(TFLITE_PROFILER_SRCS in tensorflow/lite/CMakeLists.txt). Every tool
+that pulls in the profiler must therefore compile profile_buffer.cc
+itself - tools/benchmark/CMakeLists.txt already does, but the
+label_image example does not.
+
+This is a latent bug: older compilers inlined/eliminated the
+out-of-line reference to ProfileBuffer::EndEvent(), but GCC 16 keeps it,
+so linking label_image fails with:
+
+  undefined reference to
+  'tflite::profiling::ProfileBuffer::EndEvent(unsigned int,
+   long const*, long const*)'
+
+Add profiling/profile_buffer.cc to TFLITE_LABEL_IMAGE_SRCS, mirroring
+the benchmark_model target. Its other dependencies (time.cc,
+memory_info.cc and minimal logging) are already linked, so no further
+sources are needed.
+
+Upstream-Status: Submitted [https://github.com/tensorflow/tensorflow/pull/121367]
+
+Signed-off-by: Ricardo Salveti <ricardo.salveti@oss.qualcomm.com>
+---
+ tensorflow/lite/examples/label_image/CMakeLists.txt | 1 +
+ 1 file changed, 1 insertion(+)
+
+diff --git a/tensorflow/lite/examples/label_image/CMakeLists.txt b/tensorflow/lite/examples/label_image/CMakeLists.txt
+index ae4ab447064..cb9e0cbb310 100644
+--- a/tensorflow/lite/examples/label_image/CMakeLists.txt
++++ b/tensorflow/lite/examples/label_image/CMakeLists.txt
+@@ -25,2 +25,3 @@ list(APPEND TFLITE_LABEL_IMAGE_SRCS
+   ${TFLITE_SOURCE_DIR}/profiling/memory_info.cc
++  ${TFLITE_SOURCE_DIR}/profiling/profile_buffer.cc
+   ${TFLITE_SOURCE_DIR}/profiling/profile_summarizer.cc
+-- 
+2.43.0
+
@@ -84,6 +84,7 @@ SRC_URI = " \
     file://tflite/0007-cmake-Fix-label-image-dependencies.patch \
     file://tflite/0008-cmake-Add-install-rule-for-c-interface-shared-librar.patch \
     file://tflite/0009-tflite-Add-absl-log-dependency-for-enhanced-logging-.patch \
+    file://tflite/0013-tflite-examples-label_image-build-profile_buffer.patch \
     git://github.com/google/farmhash;name=farmhash;destsuffix=tensorflow-lite-${TF_LITE_VERSION}/third_party/farmhash/;branch=master;protocol=https \
     git://github.com/google/gemmlowp.git;name=gemmlowp;destsuffix=tensorflow-lite-${TF_LITE_VERSION}/third_party/gemmlowp/;branch=master;protocol=https \
     git://github.com/pytorch/cpuinfo.git;name=cpuinfo;destsuffix=tensorflow-lite-${TF_LITE_VERSION}/third_party/cpuinfo/;branch=main;protocol=https \