Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

HIFI3 patches for LX6 CPU family rebased to Clang-16 (LLVM-295) #85

Conversation

maciej-czekaj
Copy link

This PR extends support for Xtensa processors to support HIFI3 instruction set from LX6 DSPs used by Intel HD Audio. It is a refined and rebased version of #80

Main features:

  • VALIGN register support . It is used by HIFI load-aligned instructions. -Currently VALIGN is backed by v8i8 type but it will be changed to target extension type in LLVM 16.
  • BR4 and BR2 boolean vector registers - used by HIFI conditional instructions.
  • Native support for vector variables backed by AE_DR register file. v4i16, v2i32, v1i64, v1i32 vectors are supported alongside basic artithmetic operations, loads,stores, etc..
  • Support for HIFI3 intrinsic functions using vector registers. This is important for low-level optimization and using non-standard arithmetics, such as MAC, saturation, fixed point, etc.

Xtensa S3 DSP instructions are coded using explicit register allocation.
However, some instructions miss RegState:Define flag for output
registers. This leads MachineVerifier to raise errors.
This commit adds missing definitions.
Boolean Extension support consists of:
- v1i1 boolean vector type backed by BR boolean register class
- calling convection for boolean variables
- boolean instructions implementing logical operators
- truncation and zero-extension operations for conversion to scalars
- register spill and fill logic
loadImmediate is split into two functions:
- buildLoadImmediate which accepts allocated registers as params
- loadImmediate which allocates virtual registers

buildLoadImmediate is inteded to be used in post-RA passes.
BRegFixupPass implements post-RA transformations for boolean
instructions:
 - selecting the right constant for SLLI shift instruction based on
   physical register number. The number is unknown before regster
   allocation.
 - selecting the right constant for EXTUI bit extract operation based on
   physical BR register
 - emulating MOVBA (reg copy from AR to BR) operation
LLVM bitcode tests verify that Xtensa backend can:
 - convert between v1i1 and scalar
 - generate code for v1i1 vselect
 - spill & restore BR registers
 - load and store v1i1 values
Prepare a separate directory for Xtensa-specific Clang CodeGen tests.
Extend Xtensa C ABI test to include v1i1 parameters.
Adding __builtin_xtensa_movt_s and __builtin_xtensa_movf_s intrinsics.
Adding intrincic patterns to   MOVT_S anf MOVF_S definitions.
This patch adds a definition of Xtensa LX6 CPU variant present in Intel
Cannonlake and Tigerlake SOC platforms.
Some Xtensa targets may still use GAS as a default assemblwr through
-fno-integrated-as option. These changes make the assembly output
compatible with GAS by default.
  - GAS does not recognize .word but .2byte works for both
  - Dwarf CFI is not supported by GAS. Option -fdwarf-exceptions can
    still turn it on but there is no option to turn it off, so an
    opt-in approach is more portable.
Xtensa architecture uses v2i1 (BR2 reg class) and v4i1 (BR4 reg class) boolean vectors as arguments for HIFI instructions:
 - vector compare, e.g.: AE_EQ16X4
 - vector conditional move, e.g: AE_MOVT16X4
This option is passed to GNU AS and makes Xtensa compiler driver
compatible with GCC.
@github-actions github-actions bot changed the title HIFI3 patches for LX6 CPU family rebased to Clang-16 HIFI3 patches for LX6 CPU family rebased to Clang-16 (LLVM-295) Dec 1, 2023
Intrinsics: __builtin_xtensa_ae_int32x2 and builtin_xtensa_int32
are convenience functions for easy integer-vector and vector-vector
conversions that conform to Xtensa coding style.
Xtensa C dialect allows for implicit conversion between wider and
narrower vector (via shuffle) and between integer and any vector (via
broadcast). Standard Clang vectors do not support this, so these
functions provide a handicap for better portability.
@gerekon
Copy link
Collaborator

gerekon commented Dec 21, 2023

Hi @maciej-czekaj.

The patches have been merged.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
Resolution: NA Status: Done Issue is done internally
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants