[vectorize-remark] the pass-analysis result inconsistent with the final assembly output

[vfdff]

• test: https://gcc.godbolt.org/z/WfxGE3q9r

void foo (int *restrict y, int *restrict x, int *restrict indices, int n) {
  for (int k = 0; k < 10; k++)
  for (int i = 0; i < n; i++) {
      y[i + k * M] = y[i + k* M] + 1;
      y[i + k * (M+1)] = indices[i] + 2;
  }
}

• When I add option -Rpass-analysis=loop-vectorize, it reports :8:22: remark: loop not vectorized，but in the final output assemble, we get sve assemble code as following:

.LBB0_2:                                // =>This Inner Loop Header: Depth=1
        ld1w    { z0.s }, p0/z, [x2, x9, lsl #2]
        add     z0.s, z0.s, #2                  // =0x2
        st1w    { z0.s }, p0, [x0, x9, lsl #2]
        add     x9, x9, x10
        whilelo p0.s, x9, x8
        b.mi    .LBB0_2

[fhahn]

This is related to unrolling. If you also pass -Rpass=loop-vectorize it will show we vectorized the loop once, but the outer loop has been unrolled, so there are multiple copies of the inner loop that haven't been vectorized.

Any suggestions on how this could be improved? Also tagging a few other people involved with remarks in case they have any thoughts @zjaffal @nilanjana87 @ornata @francisvm

[francisvm]

For now we don't have a way to drop optimization remarks emitted in the "past". If one optimization enables another one during a later pass, the user currently has to dig a little deeper and look through the "last" remarks for that debug location.

[ornata]

Maybe it would make sense to create groupings of remarks based off "things that tend to be related"

e.g. -Rpass=loops

Could tell us about everything that happens at the loop level. (That would mean loops + BasicBlocks that are part of loops)

Alternatively, we could add a loop mode to llvm-remarkutil and encourage users to query remarks via the tool.