Immutable expression trees in an arena, and derivative fixes

.github/workflows/ci.yml

@@ -6,7 +6,7 @@ on:
 
 jobs:
   test:
-    name: ${{ matrix.os }} / ${{ matrix.cc }} / C${{ matrix.std }}
+    name: ${{ matrix.os }} / ${{ matrix.cc }} / C${{ matrix.std }}${{ matrix.sanitize && ' / sanitizers' || '' }}
     runs-on: ${{ matrix.os }}
     strategy:
       fail-fast: false
@@ -17,6 +17,7 @@ jobs:
           - { os: ubuntu-latest, cc: gcc, std: 23 }
           - { os: macos-latest, cc: clang, std: 17 }
           - { os: macos-latest, cc: gcc-14, std: 17 }
+          - { os: ubuntu-latest, cc: clang, std: 17, sanitize: ON }
     env:
       CC: ${{ matrix.cc }}
     steps:
@@ -25,7 +26,7 @@ jobs:
         if: runner.os == 'Linux'
         run: sudo apt-get update && sudo apt-get install -y flex bison
       - name: Configure
-        run: cmake -B build -DCMAKE_BUILD_TYPE=Release -DCMAKE_C_STANDARD=${{ matrix.std }} -DCMAKE_C_FLAGS=-Werror
+        run: cmake -B build -DCMAKE_BUILD_TYPE=Release -DCMAKE_C_STANDARD=${{ matrix.std }} -DCMAKE_C_FLAGS=-Werror -DKC_SANITIZE=${{ matrix.sanitize || 'OFF' }}
       - name: Build
         run: cmake --build build -j4
       - name: Test
@@ -34,5 +35,5 @@ jobs:
         if: failure()
         uses: actions/upload-artifact@v4
         with:
-          name: test-work-${{ matrix.os }}-${{ matrix.cc }}-c${{ matrix.std }}
+          name: test-work-${{ matrix.os }}-${{ matrix.cc }}-c${{ matrix.std }}${{ matrix.sanitize && '-sanitize' || '' }}
           path: build/test-work

CHANGES.md

@@ -21,6 +21,13 @@ current C compilers, and an automated test suite has been added.
   depend on the compiler or CPU.
 * Legacy material (retired generators, DOS copies, platform Makefiles,
   site-specific scripts) has been removed. It remains in git history.
+* Expression trees are immutable and allocated from an arena: every
+  operation returns a new tree that shares its arguments instead of
+  copying them. This removes the memory leaks of the old copying API
+  and shrinks `symbmath.c` by about 600 lines. The generated code is
+  unchanged.
+* The CMake option `KC_SANITIZE` builds kc with AddressSanitizer and
+  UndefinedBehaviorSanitizer.
 
 ### Tests
 
@@ -29,6 +36,10 @@ current C compilers, and an automated test suite has been added.
   mode 3 simulations. Models that kc must reject are tested too.
 * GitHub Actions runs the suite on Linux (GCC, clang, C23) and macOS
   (clang, GCC 14).
+  One more Linux job runs it with sanitizers.
+* New test models cover the derivatives of all functions
+  (`derivatives1`, `derivatives2`) and second and third derivatives by
+  dynamical variables in mode 5 (`hessian`).
 
 ### Changed results
 
@@ -37,6 +48,11 @@ current C compilers, and an automated test suite has been added.
   reported a failure in `hqr2`, and D3 printed all-zero eigenvectors.
   The eigenvector output (mostly normalization and sign) changes for
   16 models.
+* Reducing `f+(-g)` and `f-(-g)` kept the negation, so the sign of `g`
+  flipped every time the expression was reduced. Single Jacobian
+  entries of gtf and shn90 had the wrong value, which changes their
+  mode 1 eigenvalues and mode 3 end states. In other models only the
+  printed form changes, e.g. `x+(-(y))` is now `x-y`.
 
 ### Bug fixes
 
@@ -64,3 +80,14 @@ current C compilers, and an automated test suite has been added.
   result of `TreeEval` and table lookups of unknown names.
 * The warning for an unknown command-line option printed the wrong
   character.
+* Derivatives of functions: ln gave 0, tan and tanh gave `u'/(u*2)`,
+  asinh gave `u'`, acosh was not implemented and acos crashed kc.
+* `asinh`, `acosh` and `atanh` of constants evaluated to their
+  argument.
+* Mode 2: the initial values in `in.dat` were shifted by one species;
+  the first one was read out of bounds.
+* Mode 5: second and third derivatives by dynamical variables were
+  stored in the wrong slot, which made kc crash on models with two
+  parameters and a dynamical variable. The `djacobian` procedure
+  assigned the wrong variables, with a division by zero in models
+  without constraints.

CMakeLists.txt

@@ -48,6 +48,15 @@ add_executable(kc
 target_include_directories(kc PRIVATE src ${CMAKE_CURRENT_BINARY_DIR})
 target_link_libraries(kc PRIVATE kcsolvers)
 
+# Developer option: check kc itself for memory errors, leaks and undefined
+# behaviour. Only kc is instrumented, so generated simulators link as usual.
+option(KC_SANITIZE "Build kc with AddressSanitizer and UndefinedBehaviorSanitizer" OFF)
+if(KC_SANITIZE)
+  set(sanitize -fsanitize=address,undefined -fno-sanitize-recover=undefined -fno-omit-frame-pointer)
+  target_compile_options(kc PRIVATE ${sanitize})
+  target_link_options(kc PRIVATE ${sanitize})
+endif()
+
 configure_file(kci.in ${CMAKE_CURRENT_BINARY_DIR}/kci @ONLY)
 
 install(TARGETS kc kcsolvers)

README.md

@@ -187,6 +187,15 @@ compared as well, so error handling is tested too.
 intended change of the output, rewrite the expected files with
 `KC_UPDATE=1 ctest --test-dir build` and review the diff.
 
+To check kc for memory errors, leaks and undefined behaviour, build it with
+sanitizers (leaks are only detected on Linux):
+
+```sh
+cmake -B build-sanitize -DKC_SANITIZE=ON
+cmake --build build-sanitize
+ctest --test-dir build-sanitize -j8
+```
+
 ## Legal issues
 
 The program is copyrighted in the sense of GNU General Public License

TODO.md

@@ -5,7 +5,6 @@ following list shows with which priority different subject should be
 done:
 
 * Better reduction of expressions.
-* Better memory allocation (especially in `symbmath.{c,h}`).
 * Better documentation, both user's and programmer's manual.
 * Reorganise the table manager.
 * Rewrite the grammar, i.e. better error recovery and many more consistency checks.