https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128851

            Bug ID: 128851
           Summary: Fail to hoist invariant table offset when using
                    variable index
           Product: gcc
           Version: 16.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: matmal01 at gcc dot gnu.org
  Target Milestone: ---

In the below code, loop invariant motion fails to act on the invariant
table-offset part of the access in `p->S[1][x & 255]`.  When the
table-offset is manually extracted then the hoist is performed.

It seems that the loop invariant motion pass has nothing to attempt to
identify this situation so it's currently at the mercy of what IR it gets
given.  I don't know if this means that the way to catch these cases would
be by adding more cleverness to the LIM pass or performing more lowering at
an earlier pass.

With the source code of 
    struct tables
    {
      unsigned S[4][256];
    };

    unsigned
    lookup (const struct tables *p, unsigned x, unsigned int n)
    {
      for (unsigned int j = 0; j < n; ++j)
        {
    #ifdef HOIST
          const unsigned *s1 = p->S[1];
          x = s1[x & 255];
    #else
          x = p->S[1][x & 255];
    #endif
        }
      return x;
    }

We see that without the source-level isolation of the invariant component &
array access the compiler fails to hoist that dereferencing out of the loop.

    $ aarch64-none-linux-gnu-gcc -O2 -S -o - repro/repro.c
        .arch armv8-a
        .file   "repro.c"
        .text
        .align  2
        .p2align 5,,15
        .global lookup
        .type   lookup, %function
    lookup:
    .LFB0:
        .cfi_startproc
        cbz     w2, .L2
        mov     w3, 0
        .p2align 5,,15
    .L3:
        and     x1, x1, 255
        add     w3, w3, 1
        add     w1, w1, 256
        ldr     w1, [x0, x1, lsl 2]
        cmp     w2, w3
        bne     .L3
    .L2:
        mov     w0, w1
        ret
        .cfi_endproc
    .LFE0:
        .size   lookup, .-lookup

With the source-level isolation:
    $ aarch64-none-linux-gnu-gcc -O2 -S -o - repro/repro.c -DHOIST
        .arch armv8-a
        .file   "repro.c"
        .text
        .align  2
        .p2align 5,,15
        .global lookup
        .type   lookup, %function
    lookup:
    .LFB0:
        .cfi_startproc
        cbz     w2, .L4
        add     x3, x0, 1024
        mov     w0, w1
        mov     w1, 0
        .p2align 5,,15
    .L3:
        ubfiz   x0, x0, 2, 8
        add     w1, w1, 1
        ldr     w0, [x3, x0]
        cmp     w2, w1
        bne     .L3
        ret
        .p2align 2,,3
    .L4:
        mov     w0, w1
        ret
        .cfi_endproc
    .LFE0:
        .size   lookup, .-lookup

Reply via email to