v1.4.10

pxcc is experimental and subject to change without deprecation. Please check the documentation for the latest updates.
Source that compiled before this release can fail to compile after it because pxcc starts to check more conditions
- pxcc added checks that reject -Xmu -c, an explicit -j 0, a header-defined non-template kernel body, and a bad execute() argument.
- pxcc also rejects device code that needs dynamic initialization, because MU startup never runs .init_array.

PXL: Breaking change to a public enum. A successful execution now reports ExecuteStatus::Completed, where it previously reported ExecuteStatus::Idle. Code that detects completion with getExecuteStatus() == ExecuteStatus::Idle still compiles, silently stops taking its success path, and a polling loop written against that condition never exits. Migrate to == ExecuteStatus::Completed, or use the return value of synchronize(). ExecuteStatus::Idle from now on means only “has not executed yet”.
PXL: ExecuteStatus::Completed takes the value 9, which the previous header used for ExecuteStatus::Max, and Max moves to 10. A consumer compiled against the previous header must be rebuilt before linking against this library: without a rebuild it reads a successful execution as the Max sentinel or filters it out in a range check, with no diagnostic.

openSUSE Leap 16 and RHEL 9 are experimental. Verified on openSUSE Leap 16.0 and Rocky Linux 9.6; Red Hat Enterprise Linux itself is untested.

--config=<file>, an auto-discovered .pxcc.cfg, and ~/.pxcc.cfg each read compile options from a file.
PXCC_CACHE_DIR turns on an opt-in incremental compile cache, and the cache stays off without that variable.
execute() gained two argument checks, one for arithmetic width and one for device copyability.
pxcc gained a warning on a default argument on a kernel parameter.
pxcc gained an error on an anonymous-namespace type used as a template argument in device code.

PXL: Added Job::allocateMemory() for job-private device memory.
PXL: Added Job::remapMemory() and Job::unmapMemory() to place job-private memory at a user-specified device virtual address.
PXL: Added ExecuteStatus::Completed to distinguish successful completion from the initial Idle state.
PXL: Use ccma (/dev/ccma) as default allocator which make PXL support NUMA computing.

SDK: Added experimental release targets for openSUSE Leap 16 and RHEL 9 (el9 — RHEL, Rocky Linux, AlmaLinux). Ubuntu 24.04 remains the officially supported target.
SDK: Added scripts/setup_host.sh, which applies the CXL devdax host setup per distro — kernel cmdline, memory auto-online udev override, and a boot-time conversion service. Opt in with ./install.sh --with-host-setup.

Add daemon-backed xtop attach --pid, routed per-Sub capture, and passive multi-reader view.

The cache key now covers the pxcc version and both compiler binaries, so an older entry misses once.
A cache hit now restores the device object mu.<src>.o and the four intermediate files as well.
A multi-file build runs on one shared worker pool, so one slow file no longer delays a batch.
A class-level __mu_shared__ or __mu_device__ mark now reaches an out-of-line member definition too.
The same compile error prints once, because pxcc folds the device chain copy into the host chain copy.

PXL: A successfully completed Map now reports ExecuteStatus::Completed instead of ExecuteStatus::Idle. See IMPORTANT NOTES.

SDK: install.sh and install_dependencies.sh now select apt/dpkg or zypper/dnf/rpm from /etc/os-release instead of assuming Ubuntu.
Docs: Docker guide now documents the host pxl_resourced daemon as required for running hardware (single or multiple containers).

Device code extraction now handles a concept, a member initializer, a multi-declarator declaration, and a macro-expanded definition.
An out-of-line member of a class template, and the class that holds it, now reach the device source.
A host-only function that only a member initializer calls no longer enters the device build.
pxl::createModule() now returns a real module inside a global constructor.
The .d file names your source instead of host.<src>.cpp, and a device diagnostic points at the right line.

PXL: Job ids are no longer truncated to 16 bits, so Job::id() stays unique past 65536 jobs in one process.
PXL: Fixed hangs, spurious cancels, and misreported errors on the Map completion path.
PXL: Fixed cache-coherency mismatch in the Launcher.
PXL: Fixed device memory leaking when a client failed.