v1.4.9

CXL 1.1 (RCD) enumeration on SPR host — Now supported (reboot not yet supported; planned for the next release).
Mailbox payload 8-byte write — Fixed a bug where 8-byte writes to the Mailbox payload region did not behave correctly.
HDM Range Register base = 0 handling — Fixed the device to correctly treat a Range Register Base value of 0 as a valid HPA.

pxcc is experimental and subject to change without deprecation. Please check the documentation for the latest updates.

Every device allocation now rounds up to a 64B multiple: sub-64B requests consume a full cache line each, and mid-band size classes (80/96/112/160/224…) round up to the next line-multiple class. Internal fragmentation trades off against multi-thread correctness. Set PXL_DEBUG_CACHE_LINE_ALIGN=0 to restore old packing for measurement.

Added documentation for the experimental pxcc heterogeneous C/C++ compiler (available from v1.4.8) — getting-started tutorial, programming model, compiler reference, and troubleshooting.

Added opt-in incremental compilation cache, which is independent of cmake, ninja system. Set the PXCC_CACHE_DIR environment variable to reuse previously produced objects for unchanged source files.

Added reallocateMemory / Context::memRealloc to resize an existing device allocation — resized in place when possible (same size class, free forward neighbor, or shrink), otherwise relocated with a copy. Supports sub-allocated, direct, scatter, and IoMem sources.

Major improvements to device code extraction accuracy: Fixed numerous compilation errors during device code generation, including duplicate extraction related to __mu_shared__, duplicate friend functions, annotation conflicts, and variadic templates.
Build system reliability improvements: Fixed issues where the build system failed to detect failures, such as residual .o files after device compilation failures and linker incorrectly returning exit code 0.
Host/device data consistency guarantees: Fixed silent corruption bugs including struct layout mismatches caused by missing __attribute__((packed)), and runtime string corruption caused by incorrect deletion of annotation keywords inside strings/comments.
Dependency tracking and preprocessor accuracy improvements: Fixed preprocessor and dependency-related bugs including -MQ/-MT multiple flag handling, missing type tracking within template arguments, variadic macro reconstruction, -include-pch misparsing, and UTF-8 BOM handling.
#line directive and header matching improvements: Fixed #line parsing errors for file paths containing colons, and include matching issues where declarations between headers sharing the same basename were incorrectly linked.

Removed Job::subAlloc / subAlloc(numSub) / subAllocForceAsync / subFree / subFree(numSub) (and Python Job.sub_alloc / sub_free). Use Context::createJob(numSub) / createJobForceAsync(numSub, ...) instead.

Round device allocations up to a full cache-line multiple so no size class shares a 64B line, preventing host-side false sharing across contexts/threads under host cache flush. Debug override via PXL_DEBUG_CACHE_LINE_ALIGN (default: on / one cache line).