Options

-debug

Turns off compiler optimizations and enables backtraces.

Use this when debugging compiler output or runtime failures.

This is the default optimization mode.

-release

Turns on compiler optimizations and disables debug information.

Use this for production builds or performance benchmarking.

-disable-exceptions

Disables exception handling. Attempting to raise an exception will trap. Enables further optimization by eliminating exception code paths.

Default: false

-fast-math

Enables fast-math optimizations.

This may improve performance but can change floating-point behavior.

Default: false

-disable-native

Disables architecture-specific optimizations.

By default, Codon may optimize for the host CPU architecture.

Default: false

-auto-python

Automatically falls back to Python when importing modules that Codon does not support natively.

Default: false

-D<name>=<value>

Adds a literal variable definition. For example -Dfoo=42 will include variable foo with value 42 as a Literal[int] during compilation.

-disable-opt <name>

Disables a specific Codon IR optimization pass. Can be specified multiple times.

-plugin <path>

Loads the specified Codon plugin. The provided path should refer to the directory containing the plugin's plugin.toml configuration file. Can be specified multiple times.

-log <streams>

Enables the specified compiler log streams. See Logging for more information.

-numerics=<py|c>

• py: Uses Python-style numerical semantics. Closely matches CPython behavior, but may disable certain optimizations such as vectorization.
• c: Uses C-style numerical semantics. Provides the best performance, but behavior may differ from Python in areas such as integer division and division by zero.

Note that Codon ints are 64-bit signed integers regardless of this setting.

Default: py

-libdevice <path>

Path to the NVIDIA libdevice bitcode library used when compiling GPU kernels.

Default: /usr/local/cuda/nvvm/libdevice/libdevice.10.bc

-gpu-name <architecture>

Target GPU architecture or compute capability.

Default: sm_30

-gpu-features <features>

Additional GPU feature flags passed to the backend.

Default: +ptx42

-ptx <file>

Writes generated PTX assembly to the specified file.

-unordered-dict

Uses the unordered dictionary implementation.

Default: false

-global-ctor=<yes|no|auto>

• yes: insert a call to top-level program code in a global constructor
• no: do not generate global constructors
• auto: yes if generating a shared library, no otherwise

Default: auto