llvm-ve-rv-2.2.0

• Fix VSC codegen.
• RV+isel support for strided VLD/VST (requires a constant stride).

llvm-ve-rv-2.1.0

Bug fix release

• Fixed sotoc clang tool (incorrect dylib linkage).

Otherwise identical to https://github.com/sx-aurora-dev/llvm-project/releases/llvm-ve-rv-v2.0.0

llvm-ve-rv-1.9.0

Tag of hpce/stable branch with important fixes for omp target (both, native and sotoc).

Offloading options with omp target:

VH -> VE (llvm-ve compiled device code)
clang -march=native -fopenmp -fopenmp-targets=ve-linux ...

VH -> VE (ncc compiled device code)
clang -march=native -fopenmp -fopenmp-targets=aurora-nec-veort ...

VE -> VH
clang -target=ve-linux -fopenmp -fopenmp-targets=x86_64-pc-linux-gnu ...

Other news:

RegionVectorizer diagnostics: enable with -Rpass=rv. This only applies to -target ve-linux compilations.

The limitations from release 1.8.0 are gone, no need to use -mvepacked for instrinsics any more, also no need for -fplugin=libRV.so.

llvm-ve-1.20.0

llvm version is 12.

Change:

• Fix cost of vector store. This may fix the issue that loop unrolling is not worked as expected.

llvm-ve-1.19.0

llvm version is 12.

Changes:

• Add intrinsics for lcr,scr,tscr,fidcr,fence instructions.

llvm-ve-1.18.0

llvm version is 12.0.0.

Change:

• Improve the stability of generated program by -O3.

llvm-ve-1.17.0

llvm version is 12.0.0

Changes

• Change to use internal assembler by default
• Change to compile profiler library
• Support stores of the address of a label
• Update intrinsics
• Fix initializer visibility
• Support bcr %sy, 0, dist format in assembly
• Change to rematerialize instructions retrieve consant immediate values.
  This allows users to use more vector registeres without spilling.
• Optimize generated code using a fold immediate in a peephole optimization
  pass.
• Support a vector register constraint in inline assembler.
• Support jump tables.
• Optimize prologue/epilogue coeds.
• Fix a problem that vgt intrinsic ignores the value of 3rd immediate value
  operand and vsc intrinsic instruction ignores the value of 4th immediate
  value operand.
• Correct a problem that LVL instruction is not generated correctly after
  function calls.
• Change the default CPU name to "generic".
• Remove musl support.
• Merge hpce/develop branch.
• Add new vbrv/vldz/vpcnt intrinsics.

llvm-ve-rv-2.0.0

Changes

• Compiler resource paths brought in line with develop branch.
• [rv] Fixed vectorization of certain loops with irregular loop structure in IR ("long" iteration variable bug).
• [rv] RV integrated as an LLVM component.
• Added script for setting environment variables. (llvmvevars.sh)
• LLVM build as a monolithing shared library (DYLIB build).

Offloading options with omp target:

VH -> VE (llvm-ve compiled device code)
clang -march=native -fopenmp -fopenmp-targets=ve-linux ...

VH -> VE (ncc compiled device code)
clang -march=native -fopenmp -fopenmp-targets=aurora-nec-veort ...

VE -> VH
clang -target=ve-linux -fopenmp -fopenmp-targets=x86_64-pc-linux-gnu ...

Getting llvm-ve-rv on the command line

To add the binaries of this release to your $PATH, run:
. /usr/local/ve/llvm-ve-rv-2/bin/llvmvevars.sh

llvm-ve-rv-1.8.0

Merges develop at 'feature/merge-upstream-20201124-3'.

• To use VL intrinsics, run clang/clang++ with the flag -mvepacked.
• To compile c++ executables, run clang++ -static - there are known issues with the shared libc++ on VE.
• To run clang/clang++ with the Region Vectorizer pass the option -fplugin=libRV.so and make sure that libRV.so is in your LD_LIBRARY_PATH.

llvm-ve-1.16.0

• Optimize leaf functions
• Change calling convention to expand all values to 64 bits width in order to follow ABI
• Optimize 1 byte atomic swap instructionsve_vl_pvandlo
• Remove some intrinsics

vel_pvandup
vel_pveqvlo
vel_pveqvup
vel_pvorlo
vel_pvorup
vel_pvslalo
vel_pvslaup
vel_pvslllo
vel_pvsllup
vel_pvsralo
vel_pvsraup
vel_pvsrllo
vel_pvsrlup
vel_pvxorlo
vel_pvxorup