CI: Updated dependency for build-docs updated to insure arm thumb is included in the docs

Feature: Adding support for ARM thumb

.github/workflows/ci.yml

@@ -151,7 +151,7 @@ jobs:
     - name: Use ARM64 container
       run: |
         docker run --rm -v $PWD:/app -w /app --platform linux/arm64 ghcr.io/nonannet/arm64_test:1 \
-        bash -lc "pip install . && \
+        bash -lc "pip install .[mindev] && \
                   mkdir -p build/runner && \
                   gcc -O3 -static -DENABLE_LOGGING -o build/runner/coparun src/coparun/runmem.c \
                   src/coparun/coparun.c src/coparun/mem_man.c && \
@@ -180,7 +180,8 @@ jobs:
     - name: Use ARMv6 container
       run: |
         docker run --rm -v $PWD:/app -w /app --platform linux/arm/v6 ghcr.io/nonannet/armv6_test:1 \
-        bash -lc "pip install . && \
+        bash -lc "set -x && \
+                  pip install .[mindev] && \
                   mkdir -p build/runner && \
                   gcc -O3 -static -DENABLE_LOGGING -o build/runner/coparun src/coparun/runmem.c \
                   src/coparun/coparun.c src/coparun/mem_man.c && \
@@ -209,9 +210,14 @@ jobs:
     - name: Use ARMv7 container
       run: |
         docker run --rm -v $PWD:/app -w /app --platform linux/arm/v7 ghcr.io/nonannet/armv7_test:1 \
-        bash -lc "pip install . && \
+        bash -lc "set -x && \
+                  pip install .[mindev] && \
                   mkdir -p build/runner && \
-                  gcc -O3 -static -DENABLE_LOGGING -o build/runner/coparun src/coparun/runmem.c \
+                  gcc -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -marm -static \
+                  -Wall -Wextra -Wconversion -Wsign-conversion \
+                  -Wshadow -Wstrict-overflow -O3 \
+                  -DENABLE_LOGGING \
+                  -o build/runner/coparun src/coparun/runmem.c \
                   src/coparun/coparun.c src/coparun/mem_man.c && \
                   pytest && \
                   bash tools/create_asm.sh"
@@ -221,6 +227,76 @@ jobs:
         name: runner-linux-armv7
         path: build/runner/*
 
+  build-armv7thumb:
+    needs: [build_stencils]
+    runs-on: ubuntu-latest
+    continue-on-error: true
+    steps:
+    - uses: actions/checkout@v4
+    - uses: actions/download-artifact@v4
+      with:
+        name: stencil-object-files
+        path: src/copapy/obj
+    - name: Set up QEMU for ARMv7
+      uses: docker/setup-qemu-action@v3
+      with:
+        platforms: linux/arm/v7
+    - name: Use ARMv7 container
+      run: |
+        docker run --rm -v $PWD:/app -w /app --platform linux/arm/v7 ghcr.io/nonannet/armv7_test:1 \
+        bash -lc "set -x && \
+                  pip install .[mindev] && \
+                  mkdir -p build/runner && \
+                  gcc -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -marm -static \
+                  -Wall -Wextra -Wconversion -Wsign-conversion \
+                  -Wshadow -Wstrict-overflow -O3 \
+                  -DENABLE_LOGGING \
+                  -o build/runner/coparun src/coparun/runmem.c \
+                  src/coparun/coparun.c src/coparun/mem_man.c && \
+                  export CP_TARGET_ARCH=armv7thumb && \
+                  pytest && \
+                  bash tools/create_asm.sh"
+
+    - uses: actions/upload-artifact@v4
+      with:
+        name: runner-linux-armv7thumb
+        path: build/runner/*
+
+  build-armv7mthumb:
+    needs: [build_stencils]
+    runs-on: ubuntu-latest
+    continue-on-error: true
+    steps:
+    - uses: actions/checkout@v4
+    - uses: actions/download-artifact@v4
+      with:
+        name: stencil-object-files
+        path: src/copapy/obj
+    - name: Set up QEMU for ARMv7
+      uses: docker/setup-qemu-action@v3
+      with:
+        platforms: linux/arm/v7
+    - name: Use ARMv7 container
+      run: |
+        docker run --rm -v $PWD:/app -w /app --platform linux/arm/v7 ghcr.io/nonannet/armv7_test:1 \
+        bash -lc "set -x && \
+                  pip install .[mindev] && \
+                  mkdir -p build/runner && \
+                  gcc -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -marm -static \
+                  -Wall -Wextra -Wconversion -Wsign-conversion \
+                  -Wshadow -Wstrict-overflow -O3 \
+                  -DENABLE_LOGGING \
+                  -o build/runner/coparun src/coparun/runmem.c \
+                  src/coparun/coparun.c src/coparun/mem_man.c && \
+                  export CP_TARGET_ARCH=armv7mthumb && \
+                  pytest && \
+                  bash tools/create_asm.sh"
+
+    - uses: actions/upload-artifact@v4
+      with:
+        name: runner-linux-armv7mthumb
+        path: build/runner/*
+
   build-windows:
     needs: [build_stencils]
     runs-on: windows-latest
@@ -244,7 +320,7 @@ jobs:
         python-version: ${{ matrix.python-version }}
 
     - name: Install Python dependencies
-      run: python -m pip install .[dev]
+      run: python -m pip install .[mindev]
 
     - name: Set up MSVC environment
       uses: microsoft/setup-msbuild@v2
@@ -270,7 +346,7 @@ jobs:
         path: build/runner/*
 
   release-stencils:
-    needs: [build_stencils, build-ubuntu, build-windows, build-arm64, build-armv6, build-armv7]
+    needs: [build_stencils, build-ubuntu, build-windows, build-arm64, build-armv6, build-armv7, build-armv7thumb, build-armv7mthumb]
     runs-on: ubuntu-latest
     if: github.ref == 'refs/heads/main' && github.event_name == 'push'
     permissions:
@@ -319,7 +395,7 @@ jobs:
           fi
 
   build-docs:
-    needs: [build_stencils, build-ubuntu, build-windows, build-arm64, build-armv6, build-armv7]
+    needs: [build_stencils, build-ubuntu, build-windows, build-arm64, build-armv6, build-armv7, build-armv7thumb, build-armv7mthumb]
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4

.gitignore

@@ -30,3 +30,4 @@ core
 *.log
 docs/source/start.md
 /src/copapy/_version.py
+sketch*.py

README.md

@@ -13,7 +13,7 @@ The main features can be summarized as:
 - Memory and type safety with a minimal set of runtime errors
 - Deterministic execution
 - Automatic differentiation for efficient realtime optimization (reverse-mode)
-- Optimized machine code for x86_64, ARMv6, ARMv7 and AArch64
+- Optimized machine code for x86_64, 32 Bit ARM (Cortex-A and Cortex-M) and AArch64
 - Highly portable to new architectures
 - Small Python package with minimal dependencies and no cross-compile toolchain required
 
@@ -31,7 +31,6 @@ While hardware I/O is obviously a core aspect of the project, it is not yet avai
 
 Currently in development:
 - Array stencils for handling very large arrays and generating SIMD-optimized code - e.g., for machine vision and neural network applications
-- Support for Thumb instructions required by ARM*-M targets (for MCUs)
 - Constant regrouping for further symbolic optimization of the computation graph
 
 Despite missing SIMD-optimization, benchmark performance shows promising numbers. The following chart plots the results in comparison to NumPy 2.3.5:
@@ -253,4 +252,4 @@ This project is licensed under the MIT license - see the [LICENSE](LICENSE) file
 
 [^2]: The compiler must support tail-call optimization (TCO). Currently, GCC is supported. Porting to a new architecture requires implementing a subset of relocation types used by that architecture.
 
-[^3]: Supported architectures: x86_64, AArch64, ARMv6 and 7 (non-Thumb). ARMv6/7-M (Thumb) support is in development. Code for x86 32-bit exists but has unresolved issues and a low priority.
+[^3]: Supported architectures: x86_64, AArch64, ARMv6/7 (non-Thumb) and ARMv7 Thumb for Cortex-A and Cortex-M. Code for x86 32-bit exists but has unresolved issues and a low priority.

pyproject.toml

@@ -45,14 +45,18 @@ dev = [
     "ruff",
     "mypy",
     "pytest",
-    "pelfy>=1.0.7"
+    "pelfy>=1.0.8"
+]
+mindev = [
+    "pytest",
+    "pelfy>=1.0.8"
 ]
 doc_build = [
     "sphinx",
     "pydata_sphinx_theme",
     "sphinx-autodoc-typehints",
     "myst-parser",
-  "pelfy>=1.0.7"
+    "pelfy>=1.0.8"
 ]
 
 [tool.mypy]

src/copapy/_binwrite.py

@@ -6,11 +6,14 @@ ByteOrder = Literal['little', 'big']
 
 Command = Enum('Command', [('ALLOCATE_DATA', 1), ('COPY_DATA', 2),
                            ('ALLOCATE_CODE', 3), ('COPY_CODE', 4),
-                           ('PATCH_FUNC', 0x1000), ('PATCH_OBJECT', 0x2000),
+                           ('PATCH_FUNC', 0x1000),
+                           ('PATCH_FUNC_ARM32_THM', 0x1005),
+                           ('PATCH_OBJECT', 0x2000),
                            ('PATCH_OBJECT_HI21', 0x2001),
                            ('PATCH_OBJECT_ABS', 0x2002),
                            ('PATCH_OBJECT_REL', 0x2003),
                            ('PATCH_OBJECT_ARM32_ABS', 0x2004),
+                           ('PATCH_OBJECT_ARM32_ABS_THM', 0x2006),
                            ('ENTRY_POINT', 7),
                            ('RUN_PROG', 64), ('READ_DATA', 65),
                            ('END_COM', 256), ('FREE_MEMORY', 257), ('DUMP_CODE', 258)])
@@ -22,6 +25,11 @@ class data_writer():
         self._data: list[tuple[str, bytes, int]] = []
         self.byteorder: ByteOrder = byteorder
 
+    def copy(self) -> 'data_writer':
+        cp = data_writer(self.byteorder)
+        cp._data = self._data.copy()
+        return cp
+
     def write_int(self, value: int, num_bytes: int = 4, signed: bool = False) -> None:
         self._data.append((f"INT {value}", value.to_bytes(length=num_bytes, byteorder=self.byteorder, signed=signed), 0))
 

src/copapy/_compiler.py

@@ -393,6 +393,7 @@ def compile_to_dag(node_list: Iterable[Node], sdb: stencil_database) -> tuple[bi
     # assemble stencils to main program and patch stencils
     data = sdb.get_function_code('entry_function_shell', 'start')
     data_list.append(data)
+    #print(f"* entry_function_shell (0) " + ' '.join(f'{d:02X}' for d in data))
     offset = aux_func_len + len(data)
 
     for associated_net, node in extended_output_ops:
@@ -451,10 +452,8 @@ def compile_to_dag(node_list: Iterable[Node], sdb: stencil_database) -> tuple[bi
         #print('--> ', name, list(sdb.get_relocations(name)))
         for reloc in sdb.get_relocations(name):
 
-            #assert reloc.target_symbol_info != 'STT_FUNC', "Not tested yet!"
-
             if not reloc.target_section_index:
-                assert reloc.pelfy_reloc.type == 'R_ARM_V4BX'
+                assert reloc.pelfy_reloc.type == 'R_ARM_V4BX', (reloc.pelfy_reloc.type, name, reloc.pelfy_reloc.symbol.name)
 
             elif reloc.target_symbol_info in {'STT_OBJECT', 'STT_NOTYPE', 'STT_SECTION'}:
                 # Patch constants/variable addresses on heap
@@ -489,6 +488,6 @@ def compile_to_dag(node_list: Iterable[Node], sdb: stencil_database) -> tuple[bi
         dw.write_int(patch.value, signed=True)
 
     dw.write_com(binw.Command.ENTRY_POINT)
-    dw.write_int(aux_func_len)
+    dw.write_int(aux_func_len + sdb.thumb_mode)
 
     return dw, variables

src/copapy/_math.py

@@ -310,7 +310,7 @@ def get_42(x: value[Any]) -> value[float]: ...
 def get_42(x: NumLike) -> value[float] | float:
     """Returns the value representing the constant 42"""
     if isinstance(x, value):
-        return add_op('get_42', [x, x])
+        return add_op('get_42', [x])
     return float((int(x) * 3.0 + 42.0) * 5.0 + 21.0)
 
 

src/copapy/_stencils.py

@@ -2,6 +2,7 @@ from dataclasses import dataclass
 from typing import Generator, Literal, Iterable, TYPE_CHECKING
 import struct
 import platform
+import os
 
 if TYPE_CHECKING:
     import pelfy
@@ -49,11 +50,14 @@ class patch_entry:
 
 
 def detect_process_arch() -> str:
-    """
+    """For running the code locally in the python module
-    For running the code locally in the python module
     the architecture of the current process is detected
     by this function to load the correct stencil database.
     """
+    cp_target_arch = os.environ.get("CP_TARGET_ARCH")
+    if cp_target_arch:
+        return cp_target_arch
+
     bits = struct.calcsize("P") * 8
     arch = platform.machine().lower()
 
@@ -88,11 +92,20 @@ def get_return_function_type(symbol: pelfy.elf_symbol) -> str:
 
 def get_stencil_position(func: pelfy.elf_symbol) -> tuple[int, int]:
     start_index = 0  # There must be no prolog
+
     # Find last relocation in function
     last_instr = get_last_call_in_function(func)
-    function_size = func.fields['st_size']
+
-    if last_instr + 5 >= function_size:  # Check if jump is last instruction
+    assert func.section, f"No code section specified for symbol {func.name}"
-        end_index = last_instr  # Jump can be striped
+
+    # func.section.fields['sh_size'] is equivalent to func.fields['st_size']
+    # expect for ARM thumb, here nop padding at the end for 4-byte alignment
+    # is not included in st_size
+    function_size = func.section.fields['sh_size']
+
+    # Check if jump is the last instruction and can be striped
+    if last_instr + 5 >= function_size:
+        end_index = last_instr
     else:
         end_index = function_size
 
@@ -106,11 +119,12 @@ def get_last_call_in_function(func: pelfy.elf_symbol) -> int:
     if reloc.symbol.name.startswith('dummy_'):
         return -0xFFFF  # Last relocation is not a jump
     else:
-        # Assume the call instruction is 4 bytes long for relocations with less than 32 bit and 5 bytes otherwise
+        # Assume the jump/call instruction is 4 bytes long for relocations
+        # with less than 32 bit and 5 bytes otherwise
         instruction_lengths = 4 if reloc.bits < 32 else 5
         address_field_length = 4
         #print(f"-> {[r.fields['r_offset'] - func.fields['st_value'] for r in func.relocations]}")
-        return reloc.fields['r_offset'] - func.fields['st_value'] + address_field_length - instruction_lengths
+        return reloc.fields['r_offset'] - func.offset_in_section + address_field_length - instruction_lengths
 
 
 def get_op_after_last_call_in_function(func: pelfy.elf_symbol) -> int:
@@ -118,7 +132,12 @@ def get_op_after_last_call_in_function(func: pelfy.elf_symbol) -> int:
     assert func.relocations, f'No call function in stencil function {func.name}.'
     reloc = func.relocations[-1]
     assert reloc.bits <= 32, "Relocation segment might be larger then 32 bit"
-    return reloc.fields['r_offset'] - func.fields['st_value'] + 4
+    return reloc.fields['r_offset'] - func.offset_in_section + 4
+
+
+def add_sign_int32(value: int) -> int:
+    """Convert a 32-bit unsigned integer to a signed integer."""
+    return value - 0x100000000 if value > 0x7FFFFFFF else value
 
 
 class stencil_database():
@@ -129,6 +148,7 @@ class stencil_database():
         var_size (dict[str, int]): dictionary of object names and their sizes
         byteorder (ByteOrder): byte order of the ELF file
         elf (elf_file): the loaded ELF file
+        thumb_mode (bool): entry_function_shell in ARM thumb mode
     """
 
     def __init__(self, obj_file: str | bytes):
@@ -155,6 +175,8 @@ class stencil_database():
         #                 if s.info == 'STT_OBJECT'}
         self.byteorder: ByteOrder = self.elf.byteorder
 
+        self.thumb_mode = self.elf.symbols['entry_function_shell'].thumb_mode
+
         #for name in self.function_definitions.keys():
         #    sym = self.elf.symbols[name]
         #    sym.relocations
@@ -196,19 +218,20 @@ class stencil_database():
         for reloc in symbol.relocations:
 
             # address to fist byte to patch relative to the start of the symbol
-            patch_offset = reloc.fields['r_offset'] - symbol.fields['st_value'] - start_index
+            patch_offset = reloc.fields['r_offset'] - symbol.offset_in_section - start_index
 
             if patch_offset < end_index - start_index:  # Exclude the call to the result_* function
                 reloc_entry = relocation_entry(reloc.symbol.name,
                                        reloc.symbol.info,
-                                       reloc.symbol.fields['st_value'],
+                                       reloc.symbol.fields['st_value'],  # LSB on ARM indicates thumb mode
                                        reloc.symbol.fields['st_shndx'],
-                                       symbol.fields['st_value'],
+                                       symbol.offset_in_section,
                                        start_index,
                                        reloc)
                 cache.append(reloc_entry)
                 yield reloc_entry
 
+
     def get_patch(self, relocation: relocation_entry, symbol_address: int, function_offset: int, symbol_type: int) -> patch_entry:
         """Return patch positions for a provided symbol (function or object)
 
@@ -234,12 +257,14 @@ class stencil_database():
 
         if pr.type.endswith('64_PC32') or pr.type.endswith('64_PLT32'):
             # S + A - P
-            patch_value = symbol_address + pr.fields['r_addend'] - patch_offset
+            addend = add_sign_int32(pr.fields['r_addend'])
+            patch_value = symbol_address + addend - patch_offset
             #print(f" *> {pr.type} {patch_value=} {symbol_address=} {pr.fields['r_addend']=} {pr.bits=}, {function_offset=} {patch_offset=}")
 
         elif pr.type == 'R_386_PC32':
             # S + A - P
-            patch_value = symbol_address + pr.fields['r_addend'] - patch_offset
+            addend = add_sign_int32(pr.fields['r_addend'])
+            patch_value = symbol_address + addend - patch_offset
             #print(f" *> {pr.type}     {pr.symbol.name} {patch_value=} {symbol_address=} {pr.fields['r_addend']=} {bin(pr.fields['r_addend'])} {pr.bits=}, {function_offset=} {patch_offset=}")
 
         elif pr.type == 'R_386_32':
@@ -300,28 +325,52 @@ class stencil_database():
             scale = 8
             #print(f" *> {patch_value=} {symbol_address=} {pr.fields['r_addend']=}, {function_offset=}")
 
-        elif pr.type.endswith('_MOVW_ABS_NC'):
+        elif pr.type == 'R_ARM_MOVW_ABS_NC':
-            # R_ARM_MOVW_ABS_NC
             # (S + A) & 0xFFFF
             mask = 0xFFFF
             patch_value = symbol_address + pr.fields['r_addend']
             symbol_type = symbol_type + 0x04  # Absolut value
             #print(f" *> {pr.type} {patch_value=} {symbol_address=}, {function_offset=}")
 
-        elif pr.type.endswith('_MOVT_ABS'):
+        elif pr.type =='R_ARM_MOVT_ABS':
-            # R_ARM_MOVT_ABS
             # (S + A) & 0xFFFF0000
             mask = 0xFFFF0000
             patch_value = symbol_address + pr.fields['r_addend']
             symbol_type = symbol_type + 0x04  # Absolut value
             scale = 0x10000
+            #print(f" *> {pr.type} {patch_value=} {symbol_address=}, {function_offset=}, {pr.fields['r_addend']=}")
 
         elif pr.type.endswith('_ABS32'):
             # R_ARM_ABS32
             # S + A (replaces full 32 bit)
+            assert not patch_offset % 4, 'R_ARM_ABS32 patched data like literals needs to be 4 Byte aligned'
+            # This might be caused by the call in entry_function_shell if not aligned
+
             patch_value = symbol_address + pr.fields['r_addend']
             symbol_type = symbol_type + 0x03  # Relative to data section
 
+        elif pr.type.endswith('_THM_JUMP24') or pr.type.endswith('_THM_CALL'):
+            # R_ARM_THM_JUMP24
+            # S + A - P
+            patch_value = symbol_address - patch_offset  + pr.fields['r_addend']
+            symbol_type = symbol_type + 0x05  # PATCH_FUNC_ARM32_THM
+            #print(f" *> {pr.type} {patch_value=} {symbol_address=} {pr.fields['r_addend']=} {pr.bits=}, {function_offset=} {patch_offset=}")
+
+        elif pr.type == 'R_ARM_THM_MOVW_ABS_NC':
+            # (S + A) & 0xFFFF
+            mask = 0xFFFF
+            patch_value = symbol_address + pr.fields['r_addend']
+            symbol_type = symbol_type + 0x06  # PATCH_OBJECT_ARM32_ABS_THM
+            #print(f" *> {pr.type} {patch_value=} {symbol_address=}, {function_offset=}, {pr.fields['r_addend']=}")
+        
+        elif pr.type == 'R_ARM_THM_MOVT_ABS':
+            # (S + A) & 0xFFFF0000
+            mask = 0xFFFF0000
+            patch_value = symbol_address + pr.fields['r_addend']
+            symbol_type = symbol_type + 0x06  # PATCH_OBJECT_ARM32_ABS_THM
+            scale = 0x10000
+            #print(f" *> {pr.type} {patch_value=} {symbol_address=}, {function_offset=}, {pr.fields['r_addend']=}")
+
         else:
             raise NotImplementedError(f"Relocation type {pr.type} in {relocation.pelfy_reloc.target_section.name} pointing to {relocation.pelfy_reloc.symbol.name} not implemented")
 
@@ -342,7 +391,8 @@ class stencil_database():
             func = self.elf.symbols[name]
             start_stencil, end_stencil = get_stencil_position(func)
             assert func.section
-            start_index = func.section['sh_offset'] + func['st_value'] + start_stencil
+
+            start_index = func.offset_in_file + start_stencil
             lengths = end_stencil - start_stencil
             self._stencil_cache[name] = (start_index, lengths)
 
@@ -380,7 +430,7 @@ class stencil_database():
 
     def get_symbol_offset(self, name: str) -> int:
         """Returns the offset of a specified symbol in the section."""
-        return self.elf.symbols[name].fields['st_value']
+        return self.elf.symbols[name].offset_in_section
 
     def get_symbol_section_index(self, name: str) -> int:
         """Returns the section index for a specified symbol name."""

src/coparun/runmem.c

@@ -57,6 +57,67 @@ void patch_arm32_abs(uint8_t *patch_addr, uint32_t imm16)
     *((uint32_t *)patch_addr) = instr;
 }
 
+void patch_arm_thm_abs(uint8_t *patch_addr, uint32_t imm16)
+{
+    // Thumb MOVW (T3) / MOVT (T1) encoding
+    
+    uint16_t *instr16 = (uint16_t *)patch_addr;
+    uint16_t first_half = instr16[0];
+    uint16_t second_half = instr16[1];
+
+    // Extract fields from imm16
+    uint32_t imm4 = (imm16 >> 12) & 0xF;
+    uint32_t i    = (imm16 >> 11) & 0x1;
+    uint32_t imm3 = (imm16 >> 8) & 0x7;
+    uint32_t imm8 = imm16 & 0xFF;
+
+    // Clear bits
+    first_half &= (uint16_t)(~(0x000F | (1 << 10)));
+    second_half &= (uint16_t)(~(0x00FF | (0x7 << 12)));
+
+    // Set new fields
+    first_half |= (uint16_t)((imm4 << 0) | (i << 10));
+    second_half |= (uint16_t)(imm8 | (imm3 << 12));
+
+    instr16[0] = first_half;
+    instr16[1] = second_half;
+}
+
+void patch_arm_thm_jump24(uint8_t *patch_addr, int32_t imm24)
+{
+    // Read the 32-bit instruction (two halfwords)
+    uint16_t *instr16 = (uint16_t *)patch_addr;
+    uint16_t first_half  = instr16[0];
+    uint16_t second_half = instr16[1];
+
+    // Thumb branch instructions always have LSB = 0 (halfword aligned)
+    // The imm24 offset in Thumb is shifted right by 1 when encoded
+    int32_t offset = imm24 >> 1;
+
+    // Split into S, J1, J2, imm10, imm11
+    uint32_t S     = (offset >> 23) & 0x1;
+    uint32_t I1    = (offset >> 22) & 0x1;
+    uint32_t I2    = (offset >> 21) & 0x1;
+    uint32_t imm10 = (offset >> 11) & 0x3FF;
+    uint32_t imm11 = offset & 0x7FF;
+
+    // Re-encode J1 and J2
+    uint32_t J1 = (~(I1 ^ S)) & 0x1;
+    uint32_t J2 = (~(I2 ^ S)) & 0x1;
+
+    // Clear old imm fields
+    first_half  &= 0xF800;           // Keep upper 5 bits
+    second_half &= 0xD000;           // Keep upper 5 bits
+
+    // Set new imm fields
+    first_half  |= (uint16_t)((S << 10) | imm10);
+    second_half |= (uint16_t)((J1 << 13) | (J2 << 11) | imm11);
+
+    // Write back
+    instr16[0] = first_half;
+    instr16[1] = second_half;
+}
+
 void free_memory(runmem_t *context) {
     deallocate_memory(context->executable_memory, context->executable_memory_len);
     deallocate_memory(context->data_memory, context->data_memory_len);
@@ -187,6 +248,26 @@ int parse_commands(runmem_t *context, uint8_t *bytes) {
                 patch_arm32_abs(context->executable_memory + offs, (uint32_t)((uintptr_t)(context->data_memory + value) & patch_mask) / (uint32_t)patch_scale);
                 break;
 
+            case PATCH_FUNC_ARM32_THM:
+                offs = *(uint32_t*)bytes; bytes += 4;
+                patch_mask = *(uint32_t*)bytes; bytes += 4;
+                patch_scale = *(int32_t*)bytes; bytes += 4;
+                value = *(int32_t*)bytes; bytes += 4;
+                LOG("PATCH_FUNC_ARM32_THM patch_offs=%i patch_mask=%#08x scale=%i value=%i\n",
+                    offs, patch_mask, patch_scale, value);
+                patch_arm_thm_jump24(context->executable_memory + offs, value);
+                break;
+
+            case PATCH_OBJECT_ARM32_ABS_THM:
+                offs = *(uint32_t*)bytes; bytes += 4;
+                patch_mask = *(uint32_t*)bytes; bytes += 4;
+                patch_scale = *(int32_t*)bytes; bytes += 4;
+                value = *(int32_t*)bytes; bytes += 4;
+                LOG("PATCH_OBJECT_ARM32_ABS_THM patch_offs=%i patch_mask=%#08x scale=%i value=%i imm16=%#04x\n",
+                    offs, patch_mask, patch_scale, value, (uint32_t)((uintptr_t)(context->data_memory + value) & patch_mask) / (uint32_t)patch_scale);
+                patch_arm_thm_abs(context->executable_memory + offs, (uint32_t)((uintptr_t)(context->data_memory + value) & patch_mask) / (uint32_t)patch_scale);
+                break;
+
             case ENTRY_POINT:
                 rel_entr_point = *(uint32_t*)bytes; bytes += 4;
                 context->entr_point = (entry_point_t)(context->executable_memory + rel_entr_point);

src/coparun/runmem.h

@@ -27,11 +27,13 @@
 #define ALLOCATE_CODE     3
 #define COPY_CODE         4
 #define PATCH_FUNC        0x1000
+#define PATCH_FUNC_ARM32_THM   0x1005
 #define PATCH_OBJECT      0x2000
 #define PATCH_OBJECT_HI21 0x2001
 #define PATCH_OBJECT_ABS  0x2002
 #define PATCH_OBJECT_REL  0x2003
 #define PATCH_OBJECT_ARM32_ABS 0x2004
+#define PATCH_OBJECT_ARM32_ABS_THM 0x2006
 #define ENTRY_POINT       7
 #define RUN_PROG         64
 #define READ_DATA        65

stencils/generate_stencils.py

@@ -57,8 +57,8 @@ def norm_indent(f: Callable[..., str]) -> Callable[..., str]:
 def get_entry_function_shell() -> str:
     return f"""
     {entry_func_prefix}int entry_function_shell(){{
-        volatile char stack_place_holder[{stack_size}];
+        //volatile char stack_place_holder[{stack_size}];
-        stack_place_holder[0] = 0;
+        //stack_place_holder[0] = 0;
         result_int(0);
         return 1;
     }}

tests/test_ops_armv7mthumb.py

@@ -0,0 +1,174 @@
+from copapy import NumLike, iif, value
+from copapy.backend import Store, compile_to_dag, add_read_value_remote
+import subprocess
+from copapy import _binwrite
+import copapy.backend as backend
+import os
+import warnings
+import re
+import struct
+import pytest
+import copapy as cp
+
+if os.name == 'nt':
+    # On Windows wsl and qemu-user is required:
+    # sudo apt install qemu-user
+    qemu_command = ['wsl', 'qemu-arm']
+else:
+    qemu_command = ['qemu-arm']
+
+
+def parse_results(log_text: str) -> dict[int, bytes]:
+    regex = r"^READ_DATA offs=(\d*) size=(\d*) data=(.*)$"
+    matches = re.finditer(regex, log_text, re.MULTILINE)
+    var_dict: dict[int, bytes] = {}
+
+    for match in matches:
+        value_str: list[str] = match.group(3).strip().split(' ')
+        #print('--', value_str)
+        value = bytes(int(v, base=16) for v in value_str)
+        if len(value) <= 8:
+            var_dict[int(match.group(1))] = value
+
+    return var_dict
+
+
+def run_command(command: list[str]) -> str:
+    result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, encoding='utf8', check=False)
+    assert result.returncode != 11, f"SIGSEGV (segmentation fault)\n -Error occurred: {result.stderr}\n -Output: {result.stdout}"
+    assert result.returncode == 0, f"\n -Error occurred: {result.stderr}\n -Output: {result.stdout}"
+    return result.stdout
+
+
+def check_for_qemu() -> bool:
+    command = qemu_command + ['--version']
+    try:
+        result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=False)
+    except Exception:
+        return False
+    return result.returncode == 0
+
+
+def function1(c1: NumLike) -> list[NumLike]:
+    return [c1 / 4, c1 / -4, c1 // 4, c1 // -4, (c1 * -1) // 4,
+            c1 * 4, c1 * -4,
+            c1 + 4, c1 - 4,
+            c1 > 2, c1 > 100, c1 < 4, c1 < 100]
+
+def function1ex(c1: NumLike) -> list[NumLike]:
+    return [c1 // 4]
+
+
+def function2(c1: NumLike) -> list[NumLike]:
+    return [c1 * 4.44, c1 * -4.44]
+
+
+def function3(c1: NumLike) -> list[NumLike]:
+    return [c1 / 4]
+
+
+def function4(c1: NumLike) -> list[NumLike]:
+    return [c1 == 9, c1 == 4, c1 != 9, c1 != 4]
+
+
+def function5(c1: NumLike) -> list[NumLike]:
+    return [c1 == True, c1 == False, c1 != True, c1 != False, c1 / 2, c1 + 2]
+
+
+def function6(c1: NumLike) -> list[NumLike]:
+    return [c1 == True]
+
+
+def iiftests(c1: NumLike) -> list[NumLike]:
+    return [iif(c1 > 5, 8, 9),
+            iif(c1 < 5, 8.5, 9.5),
+            iif(1 > 5, 3.3, 8.8) + c1,
+            iif(1 < 5, c1 * 3.3, 8.8),
+            iif(c1 < 5, c1 * 3.3, 8.8)]
+
+
+@pytest.mark.runner
+def test_compile():
+    c_i = value(9)
+    c_f = value(1.111)
+    c_b = value(True)
+
+    ret_test = function1(c_i) + function1(c_f) + function2(c_i) + function2(c_f) + function3(c_i) + function4(c_i) + function5(c_b) + [value(9) % 2] + iiftests(c_i) + iiftests(c_f) + [cp.asin(c_i/10)]
+    ret_ref = function1(9) + function1(1.111) + function2(9) + function2(1.111) + function3(9) + function4(9) + function5(True) + [9 % 2] + iiftests(9) + iiftests(1.111) + [cp.asin(9/10)]
+
+    out = [Store(r) for r in ret_test]
+
+    sdb = backend.stencil_db_from_package('armv7mthumb')
+    dw, variables = compile_to_dag(out, sdb)
+
+    #dw.write_com(_binwrite.Command.READ_DATA)
+    #dw.write_int(0)
+    #dw.write_int(28)
+
+    du = dw.copy()
+    dw.write_com(_binwrite.Command.RUN_PROG)
+    du.write_com(_binwrite.Command.DUMP_CODE)
+
+    for v in ret_test:
+        assert isinstance(v, value)
+        add_read_value_remote(dw, variables, v.net)
+
+    #dw.write_com(_binwrite.Command.READ_DATA)
+    #dw.write_int(0)
+    #dw.write_int(28)
+
+    dw.write_com(_binwrite.Command.END_COM)
+    du.write_com(_binwrite.Command.END_COM)
+
+    #print('* Data to runner:')
+    #dw.print()
+
+    dw.to_file('build/runner/test-armv7mthumb.copapy')
+    du.to_file('build/runner/test-armv7mthumb-dump.copapy')
+
+    if not check_for_qemu():
+        warnings.warn("qemu-armv7 not found, armv7 test skipped!", UserWarning)
+        return
+    if not os.path.isfile('build/runner/coparun-armv7'):
+        warnings.warn("armv7 runner not found, armv7 test skipped!", UserWarning)
+        return
+
+    print('----- Dump code...')
+    command = qemu_command + ['build/runner/coparun-armv7', 'build/runner/test-armv7mthumb-dump.copapy', 'build/runner/test.copapy-armv7mthumb.bin']
+    result = run_command(command)
+
+    print('----- Run code...')
+    command = qemu_command + ['build/runner/coparun-armv7', 'build/runner/test-armv7mthumb.copapy']
+    result = run_command(command)
+
+
+    print('* Output from runner:\n--')
+    print(result)
+    print('--')
+
+    assert 'Return value: 1' in result
+
+    result_data = parse_results(result)
+
+    for test, ref in zip(ret_test, ret_ref):
+        assert isinstance(test, value)
+        address = variables[test.net][0]
+        data = result_data[address]
+        if test.dtype == 'int':
+            val = int.from_bytes(data, sdb.byteorder, signed=True)
+        elif test.dtype == 'bool':
+            val = bool.from_bytes(data, sdb.byteorder)
+        elif test.dtype == 'float':
+            en = {'little': '<', 'big': '>'}[sdb.byteorder]
+            val = struct.unpack(en + 'f', data)[0]
+            assert isinstance(val, float)
+        else:
+            raise Exception(f"Unknown type: {test.dtype}")
+        print('+', val, ref, test.dtype, f"  addr={address}")
+        for t in (int, float, bool):
+            assert isinstance(val, t) == isinstance(ref, t), f"Result type does not match for {val} and {ref}"
+            assert val == pytest.approx(ref, 1e-5), f"Result does not match: {val} and reference: {ref}"  # pyright: ignore[reportUnknownMemberType]
+
+
+if __name__ == "__main__":
+    test_compile()

tests/test_ops_armv7thumb.py

@@ -0,0 +1,184 @@
+from copapy import NumLike, iif, value
+from copapy.backend import Store, compile_to_dag, add_read_value_remote
+import subprocess
+from copapy import _binwrite
+import copapy.backend as backend
+import os
+import warnings
+import re
+import struct
+import pytest
+import copapy as cp
+
+if os.name == 'nt':
+    # On Windows wsl and qemu-user is required:
+    # sudo apt install qemu-user
+    qemu_command = ['wsl', 'qemu-arm']
+else:
+    qemu_command = ['qemu-arm']
+
+
+def parse_results(log_text: str) -> dict[int, bytes]:
+    regex = r"^READ_DATA offs=(\d*) size=(\d*) data=(.*)$"
+    matches = re.finditer(regex, log_text, re.MULTILINE)
+    var_dict: dict[int, bytes] = {}
+
+    for match in matches:
+        value_str: list[str] = match.group(3).strip().split(' ')
+        #print('--', value_str)
+        value = bytes(int(v, base=16) for v in value_str)
+        if len(value) <= 8:
+            var_dict[int(match.group(1))] = value
+
+    return var_dict
+
+
+def run_command(command: list[str]) -> str:
+    result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, encoding='utf8', check=False)
+    assert result.returncode != 11, f"SIGSEGV (segmentation fault)\n -Error occurred: {result.stderr}\n -Output: {result.stdout}"
+    assert result.returncode == 0, f"\n -Error occurred: {result.stderr}\n -Output: {result.stdout}"
+    return result.stdout
+
+
+def check_for_qemu() -> bool:
+    command = qemu_command + ['--version']
+    try:
+        result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=False)
+    except Exception:
+        return False
+    return result.returncode == 0
+
+
+def function1(c1: NumLike) -> list[NumLike]:
+    return [c1 / 4, c1 / -4, c1 // 4, c1 // -4, (c1 * -1) // 4,
+            c1 * 4, c1 * -4,
+            c1 + 4, c1 - 4,
+            c1 > 2, c1 > 100, c1 < 4, c1 < 100]
+
+def function1ex(c1: NumLike) -> list[NumLike]:
+    return [c1 // 4]
+
+
+def function2(c1: NumLike) -> list[NumLike]:
+    return [c1 * 4.44, c1 * -4.44]
+
+
+def function3(c1: NumLike) -> list[NumLike]:
+    return [c1 / 4]
+
+
+def function4(c1: NumLike) -> list[NumLike]:
+    return [c1 == 9, c1 == 4, c1 != 9, c1 != 4]
+
+
+def function5(c1: NumLike) -> list[NumLike]:
+    return [c1 == True, c1 == False, c1 != True, c1 != False, c1 / 2, c1 + 2]
+
+
+def function6(c1: NumLike) -> list[NumLike]:
+    return [c1 == True]
+
+
+def iiftests(c1: NumLike) -> list[NumLike]:
+    return [iif(c1 > 5, 8, 9),
+            iif(c1 < 5, 8.5, 9.5),
+            iif(1 > 5, 3.3, 8.8) + c1,
+            iif(1 < 5, c1 * 3.3, 8.8),
+            iif(c1 < 5, c1 * 3.3, 8.8)]
+
+
+@pytest.mark.runner
+def test_compile():
+    c_i = value(9)
+    c_f = value(1.111)
+    c_b = value(True)
+
+    ret_test = function1(c_i) + function1(c_f) + function2(c_i) + function2(c_f) + function3(c_i) + function4(c_i) + function5(c_b) + [value(9) % 2] + iiftests(c_i) + iiftests(c_f) + [cp.asin(c_i/10)]
+    ret_ref = function1(9) + function1(1.111) + function2(9) + function2(1.111) + function3(9) + function4(9) + function5(True) + [9 % 2] + iiftests(9) + iiftests(1.111) + [cp.asin(9/10)]
+
+    out = [Store(r) for r in ret_test]
+
+    sdb = backend.stencil_db_from_package('armv7thumb')
+    dw, variables = compile_to_dag(out, sdb)
+
+    #dw.write_com(_binwrite.Command.READ_DATA)
+    #dw.write_int(0)
+    #dw.write_int(28)
+
+    du = dw.copy()
+    dw.write_com(_binwrite.Command.RUN_PROG)
+    du.write_com(_binwrite.Command.DUMP_CODE)
+
+    for v in ret_test:
+        assert isinstance(v, value)
+        add_read_value_remote(dw, variables, v.net)
+
+    #dw.write_com(_binwrite.Command.READ_DATA)
+    #dw.write_int(0)
+    #dw.write_int(28)
+
+    dw.write_com(_binwrite.Command.END_COM)
+    du.write_com(_binwrite.Command.END_COM)
+
+    #print('* Data to runner:')
+    #dw.print()
+
+    dw.to_file('build/runner/test-armv7thumb.copapy')
+    du.to_file('build/runner/test-armv7thumb-dump.copapy')
+
+    if not check_for_qemu():
+        warnings.warn("qemu-armv7 not found, armv7 test skipped!", UserWarning)
+        return
+    if not os.path.isfile('build/runner/coparun-armv7'):
+        warnings.warn("armv7 runner not found, armv7 test skipped!", UserWarning)
+        return
+
+    print('----- Dump code...')
+    command = qemu_command + ['build/runner/coparun-armv7', 'build/runner/test-armv7thumb-dump.copapy', 'build/runner/test.copapy-armv7thumb.bin']
+    result = run_command(command)
+
+    print('----- Run code...')
+    command = qemu_command + ['build/runner/coparun-armv7', 'build/runner/test-armv7thumb.copapy']
+    result = run_command(command)
+
+
+    print('* Output from runner:\n--')
+    print(result)
+    print('--')
+
+    assert 'Return value: 1' in result
+
+    result_data = parse_results(result)
+
+    for test, ref in zip(ret_test, ret_ref):
+        assert isinstance(test, value)
+        address = variables[test.net][0]
+        data = result_data[address]
+        if test.dtype == 'int':
+            val = int.from_bytes(data, sdb.byteorder, signed=True)
+        elif test.dtype == 'bool':
+            val = bool.from_bytes(data, sdb.byteorder)
+        elif test.dtype == 'float':
+            en = {'little': '<', 'big': '>'}[sdb.byteorder]
+            val = struct.unpack(en + 'f', data)[0]
+            assert isinstance(val, float)
+        else:
+            raise Exception(f"Unknown type: {test.dtype}")
+        print('+', val, ref, test.dtype, f"  addr={address}")
+        for t in (int, float, bool):
+            assert isinstance(val, t) == isinstance(ref, t), f"Result type does not match for {val} and {ref}"
+            assert val == pytest.approx(ref, 1e-5), f"Result does not match: {val} and reference: {ref}"  # pyright: ignore[reportUnknownMemberType]
+
+
+if __name__ == "__main__":
+    test_compile()
+
+
+"""
+qemu-arm -d in_asm,exec,cpu_reset -D qemu.log build/runner/coparun-armv7thumb build/runner/test-armv7thumb.copapy build/runner/test.copapy-armv7thumb.bin
+
+qemu-arm -d in_asm,exec -D qemu_trace.log \
+  -global driver=pl011.audiomaddr,property=addr,value=0xff7ec000 \
+  -global driver=pl011.audiomaddr,property=size,value=0x100000 \
+  your_binary
+"""

tests/test_stencil_db.py

@@ -19,11 +19,18 @@ def test_start_end_function():
 
         if symbol.relocations and symbol.relocations[-1].symbol.info == 'STT_NOTYPE':
 
-            print('-', sym_name, get_stencil_position(symbol), len(symbol.data))
+            if symbol.section and symbol.section.name == '.text':
+                print('SKIP', sym_name, '(Aux function, not a stencil)')
+                continue
+
+            if symbol.section:
+                function_size = symbol.section.fields['sh_size']  # len(symbol.data) excludes nop after the function
+
+                print('-', sym_name, get_stencil_position(symbol), function_size)
 
                 start, end = get_stencil_position(symbol)
 
-            assert start >= 0 and end >= start and end <= len(symbol.data)
+                assert (start >= 0 and end >= start and end <= function_size)
 
 
 def test_aux_functions():

tools/build.sh

@@ -1,13 +1,13 @@
 #!/bin/bash
-set -eux
+set -eu
 
 ARCH=${1:-x86_64}
 
 case "$ARCH" in
-    (x86_64|arm-v6|arm-v7|all)
+    (x86_64|arm64|arm-v6|arm-v7|arm-v7-thumb|arm-v7m-thumb|all)
         ;;
     (*)
-        echo "Usage: $0 [x86_64|arm-v6|arm-v7|all]"
+        echo "Usage: $0 [x86_64|arm64|arm-v6|arm-v7|arm-v6-thumb|arm-v7m-thumb|all]"
         exit 1
         ;;
 esac
@@ -42,13 +42,44 @@ if [[ "$ARCH" == "x86_64" || "$ARCH" == "all" ]]; then
         -o build/runner/coparun
 fi
 
+#######################################
+# ARM 64
+#######################################
+if [[ "$ARCH" == "arm64" || "$ARCH" == "all" ]]; then
+    echo "--------------arm64----------------"
+
+    LIBGCC=$(aarch64-linux-gnu-gcc -print-libgcc-file-name)
+
+    aarch64-linux-gnu-gcc -fno-pic -ffunction-sections \
+        -c $SRC -O3 -o build/stencils/stencils.o
+
+    aarch64-linux-gnu-ld -r \
+        build/stencils/stencils.o \
+        build/musl/musl_objects_arm64.o \
+        $LIBGCC \
+        -o $DEST/stencils_arm64_O3.o
+
+    aarch64-linux-gnu-objdump -d -x \
+        $DEST/stencils_arm64_O3.o \
+        > build/stencils/stencils_arm64_O3.asm
+
+    aarch64-linux-gnu-gcc \
+        -Wall -Wextra -Wconversion -Wsign-conversion -static \
+        -Wshadow -Wstrict-overflow -O3 \
+        -DENABLE_LOGGING \
+        src/coparun/runmem.c \
+        src/coparun/coparun.c \
+        src/coparun/mem_man.c \
+        -o build/runner/coparun-arm64
+fi
+
 #######################################
 # ARM v6
 #######################################
 if [[ "$ARCH" == "arm-v6" || "$ARCH" == "all" ]]; then
     echo "--------------arm-v6 32 bit----------------"
 
-    LIBGCC=$(arm-none-eabi-gcc -print-libgcc-file-name)
+    LIBGCC=$(arm-none-eabi-gcc -march=armv6 -mfpu=vfp -mfloat-abi=hard -marm -print-libgcc-file-name)
 
     arm-none-eabi-gcc -fno-pic -ffunction-sections \
         -march=armv6 -mfpu=vfp -mfloat-abi=hard -marm \
@@ -81,7 +112,7 @@ fi
 if [[ "$ARCH" == "arm-v7" || "$ARCH" == "all" ]]; then
     echo "--------------arm-v7 32 bit----------------"
 
-    LIBGCC=$(arm-none-eabi-gcc -print-libgcc-file-name)
+    LIBGCC=$(arm-none-eabi-gcc -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -marm -print-libgcc-file-name)
 
     arm-none-eabi-gcc -fno-pic -ffunction-sections \
         -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -marm \
@@ -97,6 +128,7 @@ if [[ "$ARCH" == "arm-v7" || "$ARCH" == "all" ]]; then
         $DEST/stencils_armv7_O3.o \
         > build/stencils/stencils_armv7_O3.asm
 
+    # The same runner for all ARM7
     arm-linux-gnueabihf-gcc \
         -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -marm -static \
         -Wall -Wextra -Wconversion -Wsign-conversion \
@@ -107,3 +139,71 @@ if [[ "$ARCH" == "arm-v7" || "$ARCH" == "all" ]]; then
         src/coparun/mem_man.c \
         -o build/runner/coparun-armv7
 fi
+
+#######################################
+# ARM v7 thumb Cortex-A
+#######################################
+if [[ "$ARCH" == "arm-v7-thumb" || "$ARCH" == "all" ]]; then
+    echo "--------------arm-v7a-thumb 32 bit----------------"
+
+    LIBGCC=$(arm-none-eabi-gcc -march=armv7 -mfpu=vfp3 -mthumb -print-libgcc-file-name)
+
+    arm-none-eabi-gcc -fno-pic -ffunction-sections \
+        -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -mthumb \
+        -c $SRC -O3 -o build/stencils/stencils.o
+
+    arm-none-eabi-ld -r \
+        build/stencils/stencils.o \
+        build/musl/musl_objects_armv7thumb.o \
+        $LIBGCC \
+        -o $DEST/stencils_armv7thumb_O3.o
+
+    arm-none-eabi-objdump -d -x \
+        $DEST/stencils_armv7thumb_O3.o \
+        > build/stencils/stencils_armv7thumb_O3.asm
+
+    # The same runner for all ARM7
+    arm-linux-gnueabihf-gcc \
+        -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -static \
+        -Wall -Wextra -Wconversion -Wsign-conversion \
+        -Wshadow -Wstrict-overflow -O3 \
+        -DENABLE_LOGGING \
+        src/coparun/runmem.c \
+        src/coparun/coparun.c \
+        src/coparun/mem_man.c \
+        -o build/runner/coparun-armv7thumb
+fi
+
+#######################################
+# ARM v7 thumb Cortex-M
+#######################################
+if [[ "$ARCH" == "arm-v7m-thumb" || "$ARCH" == "all" ]]; then
+    echo "--------------arm-v7m-thumb 32 bit----------------"
+
+    LIBGCC=$(arm-none-eabi-gcc -march=armv7e-m -mfpu=fpv4-sp-d16 -mfloat-abi=hard -mthumb -print-libgcc-file-name)
+
+    arm-none-eabi-gcc -fno-pic -ffunction-sections \
+        -march=armv7e-m -mfpu=fpv4-sp-d16 -mfloat-abi=hard -mthumb \
+        -c $SRC -O3 -o build/stencils/stencils.o
+
+    arm-none-eabi-ld -r \
+        build/stencils/stencils.o \
+        build/musl/musl_objects_armv7mthumb.o \
+        $LIBGCC \
+        -o $DEST/stencils_armv7mthumb_O3.o
+
+    arm-none-eabi-objdump -d -x \
+        $DEST/stencils_armv7mthumb_O3.o \
+        > build/stencils/stencils_armv7mthumb_O3.asm
+
+    # The same runner for all ARM7
+    arm-linux-gnueabihf-gcc \
+        -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -static \
+        -Wall -Wextra -Wconversion -Wsign-conversion \
+        -Wshadow -Wstrict-overflow -O3 \
+        -DENABLE_LOGGING \
+        src/coparun/runmem.c \
+        src/coparun/coparun.c \
+        src/coparun/mem_man.c \
+        -o build/runner/coparun-armv7thumb
+fi

tools/create_asm.sh

@@ -10,7 +10,6 @@ cparch=$(python3 -c "import copapy; print(copapy._stencils.detect_process_arch()
 # Disassemble stencil object file
 objdump -d -x src/copapy/obj/stencils_${cparch}_O3.o > build/runner/stencils.asm
 
-# Create example code disassembly
 python3 tools/make_example.py
 build/runner/coparun build/runner/test.copapy build/runner/test.copapy.bin
 
@@ -28,6 +27,10 @@ fi
 
 echo "Archtitecture: '$cparch'"
 
+if [[ "$cparch" == *"thumb"* ]]; then
+	objdump -D -b binary -marm -M force-thumb --adjust-vma=0x10000 build/runner/test.copapy.bin > build/runner/example.asm
+else
 	objdump -D -b binary -m $cparch --adjust-vma=0x10000 build/runner/test.copapy.bin > build/runner/example.asm
+fi
 
 rm build/runner/test.copapy.bin

tools/cross_compiler_unix/build_musl.sh

@@ -26,8 +26,11 @@ sh ../packobjs.sh arm-none-eabi-gcc arm-none-eabi-ld /object_files/musl_objects_
 # Armv7
 sh ../packobjs.sh arm-none-eabi-gcc arm-none-eabi-ld /object_files/musl_objects_armv7.o "-march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -marm"
 
+# Armv7 Thumb for Cortex-A
+sh ../packobjs.sh arm-none-eabi-gcc arm-none-eabi-ld /object_files/musl_objects_armv7thumb.o "-march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -mthumb"
+
 # Armv7 Thumb for Cortex-M3..7
-sh ../packobjs.sh arm-none-eabi-gcc arm-none-eabi-ld /object_files/musl_objects_armv7thumb.o "-march=armv7e-m -mfpu=fpv4-sp-d16 -mfloat-abi=hard -mthumb"
+sh ../packobjs.sh arm-none-eabi-gcc arm-none-eabi-ld /object_files/musl_objects_armv7mthumb.o "-march=armv7e-m -mfpu=fpv4-sp-d16 -mfloat-abi=hard -mthumb"
 
 #sh ../packobjs.sh mips mips-linux-gnu-gcc-13 mips-linux-gnu-ld
 

tools/cross_compiler_unix/packobjs.sh

@@ -25,14 +25,16 @@ ar x ../../musl/lib/libc.a sinf.o cosf.o tanf.o asinf.o acosf.o atanf.o atan2f.o
 ar x ../../musl/lib/libc.a sqrtf.o logf.o expf.o sqrt.o
 ar x ../../musl/lib/libc.a logf_data.o __tandf.o __cosdf.o __sindf.o
 ar x ../../musl/lib/libc.a fabsf.o scalbn.o floor.o floorf.o exp2f_data.o powf.o powf_data.o
-ar x ../../musl/lib/libc.a __rem_pio2f.o __math_invalidf.o __stack_chk_fail.o __math_divzerof.o __math_oflowf.o __rem_pio2_large.o __math_uflowf.o __math_xflowf.o
+ar x ../../musl/lib/libc.a __rem_pio2f.o __math_invalid.o __math_invalidf.o __stack_chk_fail.o
+ar x ../../musl/lib/libc.a __math_divzerof.o __math_oflowf.o __rem_pio2_large.o __math_uflowf.o __math_xflowf.o sqrt_data.o
 
 # Check out .lo (PIC)
 ar x ../../musl/lib/libc.a sinf.lo cosf.lo tanf.lo asinf.lo acosf.lo atanf.lo atan2f.lo
 ar x ../../musl/lib/libc.a sqrtf.lo logf.lo expf.lo sqrt.lo
 ar x ../../musl/lib/libc.a logf_data.lo __tandf.lo __cosdf.lo __sindf.lo
 ar x ../../musl/lib/libc.a fabsf.lo scalbn.lo floor.lo floorf.o exp2f_data.lo powf.lo powf_data.lo
-ar x ../../musl/lib/libc.a __rem_pio2f.lo __math_invalidf.lo __stack_chk_fail.lo __math_divzerof.lo __math_oflowf.lo __rem_pio2_large.lo __math_uflowf.lo __math_xflowf.lo
+ar x ../../musl/lib/libc.a __rem_pio2f.lo __math_invalid.lo __math_invalidf.lo __stack_chk_fail.lo
+ar x ../../musl/lib/libc.a __math_divzerof.lo __math_oflowf.lo __rem_pio2_large.lo __math_uflowf.lo __math_xflowf.lo sqrt_data.lo
 
 cd ../../musl
 

tools/crosscompile.sh

@@ -36,15 +36,20 @@ arm-none-eabi-gcc -march=armv6 -mfpu=vfp -mfloat-abi=hard -marm $FLAGS -$OPT -c
 LIBGCC=$(arm-none-eabi-gcc -print-libgcc-file-name)
 arm-none-eabi-ld -r $STMP /object_files/musl_objects_armv6.o $LIBGCC -o $DEST/stencils_armv6_$OPT.o
 
-# ARMv7 hardware fp
+# ARMv7 hardware fp for Cortex-A
 arm-none-eabi-gcc -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -marm $FLAGS -$OPT -c $SRC -o $STMP
 LIBGCC=$(arm-none-eabi-gcc -print-libgcc-file-name)
 arm-none-eabi-ld -r $STMP /object_files/musl_objects_armv7.o $LIBGCC -o $DEST/stencils_armv7_$OPT.o
 
+# ARMv7 Thumb for Cortex-A with hardware fp
+arm-none-eabi-gcc -march=armv7-a -mfpu=neon-vfpv3 -mfloat-abi=hard -mthumb $FLAGS -$OPT -c $SRC -o $STMP
+LIBGCC=$(arm-none-eabi-gcc -march=armv7 -mfpu=vfp3 -mthumb -print-libgcc-file-name)
+arm-none-eabi-ld -r $STMP /object_files/musl_objects_armv7thumb.o $LIBGCC -o $DEST/stencils_armv7thumb_$OPT.o
+
 # Armv7 Thumb for Cortex-M3..7 hardware fp
 arm-none-eabi-gcc -march=armv7e-m -mfpu=fpv4-sp-d16 -mfloat-abi=hard -mthumb $FLAGS -$OPT -c $SRC -o $STMP
-LIBGCC=$(arm-none-eabi-gcc -print-libgcc-file-name)
+LIBGCC=$(arm-none-eabi-gcc -march=armv7e-m -mfpu=fpv4-sp-d16 -mfloat-abi=hard -mthumb -print-libgcc-file-name)
-arm-none-eabi-ld -r $STMP /object_files/musl_objects_armv7thumb.o $LIBGCC -o $DEST/stencils_armv7thumb_$OPT.o
+arm-none-eabi-ld -r $STMP /object_files/musl_objects_armv7mthumb.o $LIBGCC -o $DEST/stencils_armv7mthumb_$OPT.o
 
 # PowerPC64LE
 # powerpc64le-linux-gnu-gcc-13 $FLAGS -$OPT -c $SRC -o $DEST/stencils_ppc64le_$OPT.o

tools/inspect.sh

@@ -14,3 +14,5 @@ objdump -D -b binary -m i386:x86-64 --adjust-vma=0x1000 build/runner/test.copapy
 
 build/runner/coparun-armv7 build/runner/test-armv7.copapy build/runner/test.copapy-armv7.bin
 arm-none-eabi-objdump -D -b binary -marm --adjust-vma=0x50000 build/runner/test.copapy-armv7.bin > build/runner/test.copapy-armv7.asm
+
+# arm-none-eabi-objdump -D -b binary -marm -M force-thumb --adjust-vma=0x50001 build/runner/test.copapy-armv7thumb.bin > build/runner/test.copapy-armv7thumb.asm

tools/test_example_code.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# Build arm-v7 runner and stencils
+bash tools/build.sh arm-v7
+
+# Build arm-v7-thumb stencils
+bash tools/build.sh arm-v7-thumb
+
+# Build arm-v7-thumb example code
+export CP_TARGET_ARCH=armv7thumb
+python3 tools/make_example.py
+build/runner/coparun-armv7 build/runner/test.copapy build/runner/test.copapy.bin
+
+arm-none-eabi-objdump -D -b binary -marm -M force-thumb --adjust-vma=0x1000000 build/runner/test.copapy.bin > build/runner/test.copapy-example-armv7thumb.asm
+
+# Build arm-v7-thumb example code
+export CP_TARGET_ARCH=armv7
+python3 tools/make_example.py
+build/runner/coparun-armv7 build/runner/test.copapy build/runner/test.copapy.bin
+
+arm-none-eabi-objdump -D -b binary -marm --adjust-vma=0x1000000 build/runner/test.copapy.bin > build/runner/test.copapy-example-armv7.asm

tools/test_thumb_stancils.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+
+bash tools/build.sh arm-v7-thumb
+python tests/test_ops_armv7thumb.py
+qemu-arm -d in_asm -D qemu.log build/runner/coparun-armv7thumb build/runner/test-armv7thumb.copapy build/runner/test.copapy-armv7thumb.bin
+arm-none-eabi-objdump -D -b binary -marm -M force-thumb --adjust-vma=0xff7ed000 build/runner/test.copapy-armv7thumb.bin > build/runner/test.copapy-armv7thumb.asm