[llvm-opt-report] Incorrectly indicate vector length

[vfdff]

• According llvm-opt-report.html, the V is used to indicate vector length

  V: The loop is vectorized. The following numbers indicate the vector length and the interleave factor.

• demo.c :

  __attribute__((noinline))
  void demo (double * a, int N){
  for (int i = 0; i < N; i++)
  a[i] = 2.0 * a[i];
  }

• issure reproduce: It show the vector length is 1 with the following instruction in file demo.opt.lst

  step1: clang -Ofast -march=armv8.5-a+sve -mllvm -force-vector-interleave=1 -S demo.c -fsave-optimization-record step2: llvm-opt-report demo.opt.yaml -o demo.opt.lst

  [root@localhost 2.1]# cat demo.opt.lst
  < demo.c
  1      | __attribute__((noinline))
  2      | void demo (double * a, int N){
  3 V1,1 |  for (int i = 0; i < N; i++)
  4      |    a[i] = 2.0 * a[i];
  5      | }

• We can see the indicate vector length is 2 according the final output assemble file demo.s, which is different from the above report generated by llvm-opt-report.

  .LBB0_7:                                // =>This Inner Loop Header: Depth=1
  .loc    1 4 17                          // demo.c:4:17
  ld1d    { z0.d }, p0/z, [x0, x11, lsl 3]
  .loc    1 4 15 is_stmt 0                // demo.c:4:15
  fadd    z0.d, z0.d, z0.d
  .loc    1 4 9                           // demo.c:4:9
  st1d    { z0.d }, p0, [x0, x11, lsl 3]
  .loc    1 3 26 is_stmt 1                // demo.c:3:26
  add x11, x11, x10
  cmp x9, x11
  b.ne    .LBB0_7

• BTW: opt-viewer.py can report correctly

  step1: clang -Ofast -march=armv8.5-a+sve -mllvm -force-vector-interleave=1 -S demo.c -fsave-optimization-record step2: python3 /path-to-clang/share/opt-viewer/opt-viewer.py --output-dir

[fhahn]

@vfdff do you know if this works as expected for fixed-width vector factors (i.e. when dropping +sve?) . the vector factor for the loop with +sve should be vscale x 2, it is possible that the tool doesn't parse scalable VFs properly

[vfdff]

Yes, it works fine for fixed-width vector factors. so it looks like your guess is correct.

• llvm-opt-report show the vector length is 2 with the following instruction in file demo.opt.lst

  step1: clang -Ofast -march=armv8.5-a -mllvm -force-vector-interleave=1 -S demo.c -fsave-optimization-record step2: llvm-opt-report demo.opt.yaml -o demo.opt.lst

  < demo.c
  1      | __attribute__((noinline))
  2      | void demo (double * a, int N){
  3 V2,1 |  for (int i = 0; i < N; i++)
  4      |    a[i] = 2.0 * a[i];
  5      | }

• We can also get the indicate vector length is 2 according the final output assemble file demo.s

  .LBB0_7:                                // =>This Inner Loop Header: Depth=1
      .loc    1 4 17                          // demo.c:4:17
      ldr     q0, [x11]
      .loc    1 3 26                          // demo.c:3:26
      subs    x10, x10, #2
      .loc    1 4 15                          // demo.c:4:15
      fadd    v0.2d, v0.2d, v0.2d
      .loc    1 4 9 is_stmt 0                 // demo.c:4:9
      str     q0, [x11], #16
      .loc    1 3 26 is_stmt 1                // demo.c:3:26
      b.ne    .LBB0_7