Consider the Thead extensions

[koute]

• https://github.com/T-head-Semi/thead-extension-spec/releases/tag/2.2.2
• Newest LLVM has these target features:

    xtheadba                      - 'xtheadba' (T-Head address calculation instructions).
    xtheadbb                      - 'xtheadbb' (T-Head basic bit-manipulation instructions).
    xtheadbs                      - 'xtheadbs' (T-Head single-bit instructions).
    xtheadcmo                     - 'xtheadcmo' (T-Head cache management instructions).
    xtheadcondmov                 - 'xtheadcondmov' (T-Head conditional move instructions).
    xtheadfmemidx                 - 'xtheadfmemidx' (T-Head FP Indexed Memory Operations).
    xtheadmac                     - 'xtheadmac' (T-Head Multiply-Accumulate Instructions).
    xtheadmemidx                  - 'xtheadmemidx' (T-Head Indexed Memory Operations).
    xtheadmempair                 - 'xtheadmempair' (T-Head two-GPR Memory Operations).
    xtheadsync                    - 'xtheadsync' (T-Head multicore synchronization instructions).
    xtheadvdot                    - 'xtheadvdot' (T-Head Vector Extensions for Dot).

[koute]

Number of instruction from a given extension when compiling the Pinky example:

• xtheadba - 75
• xtheadbs - 54
• xtheadbb - 174
• xtheadcondmov - 179
• xtheadmemidx - 67
• xtheadmempair - 86
• all above enabled at the same time - 594

The most promising one is most likely the conditional move extension as it's relatively widely applicable, can reduce the number of basic blocks, and the whole extension is only two extra instructions.