fix: keep large arrays off the stack on Windows GNU builds

CHANGELOG.md

@@ -24,6 +24,16 @@ All notable changes to CableDyn are recorded in this file. The format follows
   the displaced mass) no longer depends on the sign of the round-off weight: a line whose
   total weight is below 1e-12 of its axial stiffness is seeded as a straight, uniformly
   stretched line. Previously the catenary seed could fail to close on some platforms.
+- Windows builds with gfortran: the `modal` test could stop with a stack overflow on some
+  CPUs, because it kept 1.7 MB of dense matrices on the stack. Large test arrays and the
+  initialisation workspaces of the coupled aggregate and the C API are now allocated on the
+  heap, every executable of such a build reserves a 64 MiB stack, and the new `stack` tests
+  run the library on a 1 MiB thread stack. Results are unchanged.
+- The static Windows `openfast.exe` compiles the CableDyn adapter with `/heap-arrays:1024`,
+  like the CableDyn core, so its run-time-sized arrays no longer use the
+  `openfast.exe` stack.
+- Builds with gfortran 16 are free of `-Wuninitialized` warnings; the reported values (the
+  bounds of unallocated components) were never read.
 
 ## [0.1.0] - 2026-10-01
 

CMakeLists.txt

@@ -99,6 +99,22 @@ if(CMAKE_Fortran_COMPILER_ID STREQUAL "GNU")
                       $<$<COMPILE_LANGUAGE:C>:-Wall>
                       $<$<COMPILE_LANGUAGE:C>:-Wextra>)
 endif()
+# GNU builds on Windows (MinGW-w64): every executable gets a 64 MiB stack reserve instead of
+# the 2 MiB MinGW default. With OpenMP, gfortran places fixed-size local arrays on the stack
+# (-fopenmp implies -frecursive), and the stack depth of the OpenBLAS kernels varies with the
+# CPU, so a small reserve can overflow on one machine and pass on another. A reserve is address
+# space, not committed memory. The library itself must run within a small stack, since a host
+# such as python.exe sets the stack of the threads that call it: the stack_* tests check that.
+# The static IFX release sets its own reserve on the driver.
+set(CABLEDYN_GNU_WINDOWS_STACK OFF)
+if(MINGW AND CMAKE_Fortran_COMPILER_ID STREQUAL "GNU")
+  set(CABLEDYN_GNU_WINDOWS_STACK ON)
+  set(CABLEDYN_EXE_STACK_RESERVE 67108864)
+  set(CABLEDYN_SMALL_STACK_RESERVE 1048576)
+  add_link_options(
+    "$<$<STREQUAL:$<TARGET_PROPERTY:TYPE>,EXECUTABLE>:-Wl,--stack,${CABLEDYN_EXE_STACK_RESERVE}>")
+endif()
+
 if(CABLEDYN_ENABLE_COVERAGE AND NOT CMAKE_Fortran_COMPILER_ID STREQUAL "GNU")
   message(FATAL_ERROR "CABLEDYN_ENABLE_COVERAGE requires GNU Fortran/gcov instrumentation")
 endif()
@@ -1074,6 +1090,25 @@ else()
                  "the portable C API lifecycle gate remains enabled")
 endif()
 
+# Portable (Win32 or POSIX threads) stack bound of the shared library: each deck is
+# initialised, stepped, and evaluated on a thread with a 1 MiB stack (a host program, not
+# CableDyn, chooses the stack of the threads that call the library). The decks are a
+# 2048-element finite-EI cable (12294 DOFs), an EI=0 + finite-EI aggregate, and an EI=0
+# mooring; a stack need that grew with the model would overflow here.
+add_executable(test_c_api_stack tests/test_c_api_stack.c)
+target_include_directories(test_c_api_stack PRIVATE "${CMAKE_CURRENT_SOURCE_DIR}/src")
+target_link_libraries(test_c_api_stack PRIVATE $<TARGET_LINKER_FILE:cabledyn_shared>)
+add_dependencies(test_c_api_stack cabledyn_shared)
+if(NOT WIN32)
+  target_link_libraries(test_c_api_stack PRIVATE Threads::Threads m)
+endif()
+add_test(NAME stack_c_api
+         COMMAND test_c_api_stack 1024
+                 0.05 ${CMAKE_CURRENT_SOURCE_DIR}/tests/data/stack_humboldt800_2048_elements.dat
+                 0.025 ${CMAKE_CURRENT_SOURCE_DIR}/examples/iea15mw_umaine_mixed_cabledyn.dat
+                 0.1 ${CMAKE_CURRENT_SOURCE_DIR}/examples/iea15mw_volturnus_mooring.dat)
+set_tests_properties(stack_c_api PROPERTIES LABELS "c;unit;capi;stack" TIMEOUT 900)
+
 # Portable (Win32 or POSIX threads) gate: concurrent InitDeck of one deck and of decks
 # sharing a Syrope OWC table in the user's locale, concurrent Step on distinct handles,
 # the missing-file diagnostic, and the BLAS thread/memory policy of CableDyn_Create.
@@ -1630,6 +1665,53 @@ target_link_libraries(test_modal PRIVATE cabledyn_core)
 add_test(NAME modal COMMAND test_modal)
 set_tests_properties(modal PROPERTIES LABELS "fortran;unit;driver;static;validation")
 
+# Stack bound (GNU Windows builds): the modal test and the driver relinked with a 1 MiB stack
+# reserve, half the MinGW default, instead of the 64 MiB reserve of the other executables. A
+# large local array (as the dense matrices test_modal once kept on the stack) or a library
+# frame that grows with the model overflows here on every machine. The driver runs a
+# 2048-element finite-EI cable. Linux and macOS executables take their main-thread stack
+# from the environment (ulimit -s), so these links are Windows-only; test_c_api_stack covers
+# the library on every platform.
+if(CABLEDYN_GNU_WINDOWS_STACK)
+  file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/stack_small")
+  add_executable(test_modal_small_stack tests/test_modal.f90)
+  target_link_libraries(test_modal_small_stack PRIVATE cabledyn_core)
+  add_executable(cabledyn_small_stack app/cabledyn.f90)
+  target_link_libraries(cabledyn_small_stack PRIVATE cabledyn_core)
+  foreach(_cabledyn_small IN ITEMS test_modal_small_stack cabledyn_small_stack)
+    # Appended after the 64 MiB directory-wide option; the last --stack wins (checked below).
+    target_link_options(${_cabledyn_small} PRIVATE "-Wl,--stack,${CABLEDYN_SMALL_STACK_RESERVE}")
+  endforeach()
+  add_test(NAME stack_modal COMMAND test_modal_small_stack
+           WORKING_DIRECTORY "${CMAKE_BINARY_DIR}/stack_small")
+  set_tests_properties(stack_modal PROPERTIES LABELS "fortran;unit;driver;static;stack")
+  add_test(NAME stack_driver_large_deck
+           COMMAND cabledyn_small_stack
+                   "${CMAKE_CURRENT_SOURCE_DIR}/tests/data/stack_humboldt800_2048_elements.dat"
+                   "${CMAKE_BINARY_DIR}/stack_small/stack_driver_large_deck")
+  set_tests_properties(stack_driver_large_deck PROPERTIES
+                       LABELS "fortran;driver;integration;dynamic;stack" TIMEOUT 900)
+  # The reserve each link actually received, read from the PE headers: <binary>=<bytes>.
+  string(JOIN "|" _cabledyn_stack_checks
+         "$<TARGET_FILE:cabledyn>=${CABLEDYN_EXE_STACK_RESERVE}"
+         "$<TARGET_FILE:test_modal>=${CABLEDYN_EXE_STACK_RESERVE}"
+         "$<TARGET_FILE:test_c_api_stack>=${CABLEDYN_EXE_STACK_RESERVE}"
+         "$<TARGET_FILE:test_modal_small_stack>=${CABLEDYN_SMALL_STACK_RESERVE}"
+         "$<TARGET_FILE:cabledyn_small_stack>=${CABLEDYN_SMALL_STACK_RESERVE}")
+  add_test(NAME stack_reserve
+           COMMAND ${CMAKE_COMMAND} "-DCHECKS=${_cabledyn_stack_checks}"
+                   -P ${CMAKE_CURRENT_SOURCE_DIR}/tests/check_stack_reserve.cmake)
+  set_tests_properties(stack_reserve PROPERTIES LABELS "unit;driver;stack")
+  # The check itself must see a reserve that differs from the expected one.
+  add_test(NAME stack_reserve_mismatch
+           COMMAND ${CMAKE_COMMAND}
+                   "-DCHECKS=$<TARGET_FILE:test_modal_small_stack>=${CABLEDYN_EXE_STACK_RESERVE}"
+                   -P ${CMAKE_CURRENT_SOURCE_DIR}/tests/check_stack_reserve.cmake)
+  set_tests_properties(stack_reserve_mismatch PROPERTIES LABELS "unit;driver;stack"
+                       PASS_REGULAR_EXPRESSION
+                       "stack reserve ${CABLEDYN_SMALL_STACK_RESERVE} bytes, expected ${CABLEDYN_EXE_STACK_RESERVE}")
+endif()
+
 add_executable(test_seabed_friction_aniso tests/test_seabed_friction_aniso.f90)
 target_link_libraries(test_seabed_friction_aniso PRIVATE cabledyn_core)
 add_test(NAME seabed_friction_aniso COMMAND test_seabed_friction_aniso)

DEVELOPMENT.md

@@ -44,6 +44,16 @@ The Debug configuration (`-DCMAKE_BUILD_TYPE=Debug`) adds `-fcheck=all
 conda OpenBLAS through `CONDA_PREFIX`. `cmake --install build --prefix stage`
 installs the driver, the shared library, and `CableDyn_CAPI.h`.
 
+A Windows build with gfortran links every executable with a 64 MiB stack reserve
+(the MinGW default is 2 MiB). With OpenMP enabled, gfortran keeps fixed-size local
+arrays on the stack, and the stack depth of the OpenBLAS kernels differs between
+CPUs, so a small reserve can pass on one machine and overflow on another. The
+shared library cannot rely on this reserve: the host program (for example
+`python.exe`, which reserves about 2 MB) sets the stack of the threads that call
+it. The `-L stack` tests check the library on a thread with a 1 MiB stack and, in
+a Windows gfortran build, the driver and the modal test relinked with a 1 MiB
+reserve and the reserve written into each executable.
+
 | Selection | Tests |
 |---|---|
 | `-L fortran` | All Fortran tests |
@@ -56,6 +66,7 @@ installs the driver, the shared library, and `CableDyn_CAPI.h`.
 | `-L external_ref` | IEA-15MW VolturnUS-S mooring comparisons |
 | `-L cosserat` | Secondary Cosserat path |
 | `-L python` | Python package, deck-style, and documentation checks |
+| `-L stack` | Library, driver, and modal test on a 1 MiB stack |
 
 A test program can also be run directly from the build tree (`build/`, or
 `build\bin\` with the conda toolchain on Windows); it prints `PASS:` or
@@ -124,7 +135,14 @@ Fortran lines longer than 120 columns must be split with `&` before formatting.
   `ErrMsg (CHARACTER(*), INTENT(OUT))`.
 - LAPACK: banded solvers (`DGBSV`, and `DPBTRF`/`DPBTRS` for symmetric positive
   definite bands); a dense assembly is packed into band storage before the solve.
-- Warning-clean under `-std=f2018 -Wall -Wextra -fimplicit-none`.
+- Warning-clean under `-std=f2018 -Wall -Wextra -fimplicit-none` with gfortran 15
+  and 16. gfortran 16 can report `-Wuninitialized` for a default-initialised local
+  derived-type variable with allocatable components: the optimiser copies the
+  unset bounds of its unallocated arrays, which the program never reads. Such a
+  local is `ALLOCATABLE` where this occurs.
+- Keep large data off the stack: a local array or derived-type variable that can
+  exceed about 64 KB is `ALLOCATABLE`, except on the per-step path, which uses the
+  preallocated workspaces and does not allocate.
 
 ### SPDX headers
 

app/cabledyn.f90

@@ -395,7 +395,7 @@ SUBROUTINE run_mixed_held_deck(deck, root, dt, has_dt, run_time, has_run_time, h
     INTEGER, INTENT(OUT) :: run_stat
     CHARACTER(*), INTENT(OUT) :: run_msg
 
-    TYPE(CD_AGG_ModuleType) :: aggregate
+    TYPE(CD_AGG_ModuleType), ALLOCATABLE :: aggregate   ! about 37 KB: on the heap, not the stack
     REAL(wp), ALLOCATABLE :: position(:, :), velocity(:, :), acceleration(:, :), load(:, :)
     REAL(wp) :: time, first_miss_time, last_miss_time, step_dt
     INTEGER :: es, n_moving, nstep, step, n_iter, miss_count, consecutive_miss, max_consecutive_miss
@@ -447,6 +447,12 @@ SUBROUTINE run_mixed_held_deck(deck, root, dt, has_dt, run_time, has_run_time, h
     END IF
 
     mixed_run: BLOCK
+      ALLOCATE (aggregate, STAT=ios)
+      IF (ios /= 0) THEN
+        run_stat = CD_DECKDRV_SOLVEFAIL
+        run_msg = 'CableDyn_driver: cannot allocate the mixed aggregate'
+        EXIT mixed_run
+      END IF
       ! The aggregate is caller-driven and cannot refresh standalone wave/current
       ! OPTIONS. Reject such decks by name instead of silently holding zero fluid.
       CALL CD_AGG_Init_From_Deck(aggregate, deck, step_dt, es, em, forbid_deck_ambient=.TRUE., &
@@ -574,7 +580,7 @@ SUBROUTINE run_mixed_held_deck(deck, root, dt, has_dt, run_time, has_run_time, h
     END BLOCK mixed_run
 
     IF (file_open) CLOSE (unit)
-    CALL CD_AGG_End(aggregate, es, em)
+    IF (ALLOCATED(aggregate)) CALL CD_AGG_End(aggregate, es, em)
   END SUBROUTINE run_mixed_held_deck
 
   LOGICAL FUNCTION deck_requests_modes(deck) RESULT(requested)

integration/openfast/README.md

@@ -45,7 +45,9 @@ the system code page cannot spell; the static release build embeds it.
 
 For the static Windows release, pass a clean, unpatched checkout to
 `release/build_static_windows.ps1`; it applies the series itself (see
-[Installation](../../doc/installation.rst)).
+[Installation](../../doc/installation.rst)). Its Visual Studio project compiles the CableDyn
+adapter with `/heap-arrays:1024`, as the static CMake build compiles the CableDyn core, so arrays
+whose size is known only at run time are placed on the heap rather than on the `openfast.exe` stack.
 
 ## License of modified files
 

integration/openfast/vs-build/CableDyn.vfproj.in

@@ -5,35 +5,35 @@
 	</Platforms>
 	<Configurations>
 		<Configuration Name="Debug|x64" UseCompiler="ifxCompiler" OutputDirectory="..\..\build\lib" IntermediateDirectory="..\..\build\$(Configuration)_$(Platform)\$(ProjectName)\" ConfigurationType="typeStaticLibrary">
-			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" DebugInformationFormat="debugEnabled" MultiProcessorCompilation="true" Optimization="optimizeDisabled" Preprocess="preprocessYes" WarnInterfaces="true" Traceback="true" BoundsCheck="true" StackFrameCheck="true" RuntimeLibrary="rtMultiThreadedDebug" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
+			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" DebugInformationFormat="debugEnabled" MultiProcessorCompilation="true" Optimization="optimizeDisabled" Preprocess="preprocessYes" WarnInterfaces="true" Traceback="true" BoundsCheck="true" StackFrameCheck="true" RuntimeLibrary="rtMultiThreadedDebug" HeapArrays="1024" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
 			<Tool Name="VFLibrarianTool"/>
 		</Configuration>
 		<Configuration Name="Double_Debug|x64" UseCompiler="ifxCompiler" OutputDirectory="..\..\build\lib" IntermediateDirectory="..\..\build\$(Configuration)_$(Platform)\$(ProjectName)\" ConfigurationType="typeStaticLibrary">
-			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" DebugInformationFormat="debugEnabled" MultiProcessorCompilation="true" Optimization="optimizeDisabled" Preprocess="preprocessYes" WarnInterfaces="true" RealKIND="realKIND8" Traceback="true" BoundsCheck="true" StackFrameCheck="true" RuntimeLibrary="rtMultiThreadedDebug" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
+			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" DebugInformationFormat="debugEnabled" MultiProcessorCompilation="true" Optimization="optimizeDisabled" Preprocess="preprocessYes" WarnInterfaces="true" RealKIND="realKIND8" Traceback="true" BoundsCheck="true" StackFrameCheck="true" RuntimeLibrary="rtMultiThreadedDebug" HeapArrays="1024" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
 			<Tool Name="VFLibrarianTool"/>
 		</Configuration>
 		<Configuration Name="Release|x64" UseCompiler="ifxCompiler" OutputDirectory="..\..\build\lib" IntermediateDirectory="..\..\build\$(Configuration)_$(Platform)\$(ProjectName)\" ConfigurationType="typeStaticLibrary">
-			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" MultiProcessorCompilation="true" Preprocess="preprocessYes" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
+			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" MultiProcessorCompilation="true" Preprocess="preprocessYes" HeapArrays="1024" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
 			<Tool Name="VFLibrarianTool"/>
 		</Configuration>
 		<Configuration Name="Double_Release|x64" UseCompiler="ifxCompiler" OutputDirectory="..\..\build\lib" IntermediateDirectory="..\..\build\$(Configuration)_$(Platform)\$(ProjectName)\" ConfigurationType="typeStaticLibrary">
-			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" MultiProcessorCompilation="true" Preprocess="preprocessYes" RealKIND="realKIND8" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
+			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" MultiProcessorCompilation="true" Preprocess="preprocessYes" RealKIND="realKIND8" HeapArrays="1024" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
 			<Tool Name="VFLibrarianTool"/>
 		</Configuration>
 		<Configuration Name="OpenMP_Release|x64" UseCompiler="ifxCompiler" OutputDirectory="..\..\build\lib" IntermediateDirectory="..\..\build\$(Configuration)_$(Platform)\$(ProjectName)\" ConfigurationType="typeStaticLibrary">
-			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" MultiProcessorCompilation="true" Preprocess="preprocessYes" OpenMP="OpenMPParallelCode" EnableOpenMPSupport="OpenMPParallelCodeIFX" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
+			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" MultiProcessorCompilation="true" Preprocess="preprocessYes" OpenMP="OpenMPParallelCode" EnableOpenMPSupport="OpenMPParallelCodeIFX" HeapArrays="1024" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
 			<Tool Name="VFLibrarianTool"/>
 		</Configuration>
 		<Configuration Name="Double_OpenMP_Release|x64" UseCompiler="ifxCompiler" OutputDirectory="..\..\build\lib" IntermediateDirectory="..\..\build\$(Configuration)_$(Platform)\$(ProjectName)\" ConfigurationType="typeStaticLibrary">
-			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" MultiProcessorCompilation="true" Preprocess="preprocessYes" OpenMP="OpenMPParallelCode" EnableOpenMPSupport="OpenMPParallelCodeIFX" RealKIND="realKIND8" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
+			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" MultiProcessorCompilation="true" Preprocess="preprocessYes" OpenMP="OpenMPParallelCode" EnableOpenMPSupport="OpenMPParallelCodeIFX" RealKIND="realKIND8" HeapArrays="1024" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
 			<Tool Name="VFLibrarianTool"/>
 		</Configuration>
 		<Configuration Name="Matlab_Debug|x64" UseCompiler="ifxCompiler" OutputDirectory="..\..\build\lib" IntermediateDirectory="..\..\build\$(Configuration)_$(Platform)\$(ProjectName)\" ConfigurationType="typeStaticLibrary">
-			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" DebugInformationFormat="debugEnabled" Optimization="optimizeDisabled" Preprocess="preprocessYes" WarnInterfaces="true" Traceback="true" BoundsCheck="true" StackFrameCheck="true" RuntimeLibrary="rtMultiThreadedDebug" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
+			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" DebugInformationFormat="debugEnabled" Optimization="optimizeDisabled" Preprocess="preprocessYes" WarnInterfaces="true" Traceback="true" BoundsCheck="true" StackFrameCheck="true" RuntimeLibrary="rtMultiThreadedDebug" HeapArrays="1024" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
 			<Tool Name="VFLibrarianTool"/>
 		</Configuration>
 		<Configuration Name="Matlab_Release|x64" UseCompiler="ifxCompiler" OutputDirectory="..\..\build\lib" IntermediateDirectory="..\..\build\$(Configuration)_$(Platform)\$(ProjectName)\" ConfigurationType="typeStaticLibrary">
-			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" Preprocess="preprocessYes" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
+			<Tool Name="VFFortranCompilerTool" SuppressStartupBanner="true" Preprocess="preprocessYes" HeapArrays="1024" UseMkl="mklSequential" AdditionalIncludeDirectories="@CABLEDYN_MODULE_DIR@"/>
 			<Tool Name="VFLibrarianTool"/>
 		</Configuration>
 	</Configurations>