[Bug target/128443] New: Reassociate shift-add-add chain with an out of range constant for RISC-V

[law at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128443

            Bug ID: 128443
           Summary: Reassociate shift-add-add chain with an out of range
                    constant for RISC-V
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: law at gcc dot gnu.org
            Blocks: 26163
  Target Milestone: ---
            Target: riscv

This is a hacked testcase from x264.

typedef unsigned char pixel;

typedef struct
{
  signed char ref[2][40];
  char pad[22992 - 80];
  pixel *p_fref[2][32][6];
} x264_t;

pixel *
f (x264_t *h, int i8)
{
  int i_ref0 = h->ref[0][i8];
  return h->p_fref[0][i_ref0][0];
}

        li      a4,24576                # 17    [c=4 l=4]  *movdi_64bit/1
        sh1add  t0,a5,a5        # 13    [c=4 l=4]  *shNadd
        slli    t1,t0,4 # 14    [c=4 l=4]  ashldi3
        add     a0,a0,t1        # 16    [c=4 l=4]  *adddi3/0
        add     t2,a0,a4        # 18    [c=4 l=4]  *adddi3/0

So we could add 1536 to t0 before the shift by 4.  That would exchange an
li+add for an addi.  Profitable.  But very small on x264 AFAICT.  And with the
li issuing speculatively with no input dependencies, it's even less profitable
in practice.  Very low priority.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=26163
[Bug 26163] [meta-bug] missed optimization in SPEC (2026, 2k17, 2k and 2k6 and
95)