[MIPS] microMIPS and MIPS16 don't use jr instruction delay slot and instead always use the version with no delay slot

[GabrielRavier]

A function such as:

#include <stdint.h>

uint8_t f(uint8_t x)
{
    return x;
}

gets compiled by GCC to:

f(unsigned char):
        jr      $31
        move    $2,$4

on MIPS with either -mmicromips or -mips16 in use, whereas in those cases LLVM emits:

f(unsigned char):                                  # @f(unsigned char)
        move    $2, $4
        jrc     $ra

I've searched around and could not find anything contradicting the idea that using the version with a delay slot when the delay slot can be filled is faster.

[llvmbot]

@llvm/issue-subscribers-backend-mips

Author: Gabriel Ravier (GabrielRavier)

A function such as:

#include <stdint.h>

uint8_t f(uint8_t x)
{
    return x;
}

gets compiled by GCC to:

f(unsigned char):
        jr      $31
        move    $2,$4

on MIPS with either -mmicromips or -mips16 in use, whereas in those cases LLVM emits:

f(unsigned char):                                  # @<!-- -->f(unsigned char)
        move    $2, $4
        jrc     $ra

I've searched around and could not find anything contradicting the idea that using the version with a delay slot when the delay slot can be filled is faster.