Optimization reports

Please note the following when using the LLVM remarks framework options:

• If you use the --pass-remarks-output=<filename> option to save the optimization remarks to a file, make sure that you request remarks from only one of ibm-opt, ibm-llc, or LTO, or to save the remarks from each component in a separate file. For example, the following command saves the remarks from ibm-opt to file.opt.yaml and the remarks from ibm-llc to file.llc.yaml:

  xlf95 file.f -O3 \
    -Xopt --pass-remarks-output=file.opt.yaml \
    -Xllc --pass-remarks-output=file.llc.yaml

  The following command writes remarks from ibm-opt to file.yaml and then overwrites the file with the remarks from ibm-llc:

  xlf95 file.f -O3 -mllvm --pass-remarks-output=file.yaml

• You must enable debug line number generation (-g1 or -g) to get line numbers in the optimization remarks, or to use external visualization tools like opt-viewer.py.
• Many LLVM optimizations produce remarks. You can use the filtering options to control the amount of information presented.

The following example demonstrates generating optimization reports using the -qreport option as well as the LLVM remarks framework:

$ cat loop.f
module m
contains
subroutine sub(a, b, c)
  real a(160), b(160), c(160)
  integer i

  do i = 1, 160
    a(i) = b(i) + c(i)
  end do
 
  do i = 1, 8
    a(i) = a(i) * 3
  end do
end
 
subroutine sub2(a, b, c)
  real a(160), b(160), c(160)
  call sub(a, b, c)
end subroutine
end module
 
program main
  use m
  real a(160), b(160), c(160)
  read(*, *) b, c
  call sub2(a, b, c)
  print *, a
end program
$

This command generates loop.lst:

xlf95 loop.f -c -O3 -qarch=pwr9 -qreport

loop.lst contains the following optimization report:

>>>>> LOOP TRANSFORMATION SECTION <<<<<
< loop.f
 1         | module m
 2         | contains
 3         | subroutine sub(a, b, c)
 4         |   real a(160), b(160), c(160)
 5         |   integer i
 6         |
 7  U5V4,8 |   do i = 1, 160
 8         |     a(i) = b(i) + c(i)
 9         |   end do
10         |
11  U8     |   do i = 1, 8
12         |     a(i) = a(i) * 3
13         |   end do
14         | end
15         |
16         | subroutine sub2(a, b, c)
17         |   real a(160), b(160), c(160)
18 I       |   call sub(a, b, c)
19         | end subroutine
20         | end module
21         |
22         | program main
23         |   use m
24         |   real a(160), b(160), c(160)
25         |   read(*, *) b, c
26         |   call sub2(a, b, c)
27         |   print *, a
28         | end program

The letters in the left margin indicate the transformation performed. U<iterations> means the loop was unrolled with unroll factor <iterations>. V<width, count> means the loop was auto-vectorized with vector width <width> and interleaved count <count>. I means the function call was inlined.

Compile using the following command:

xlf95 loop.f -c -O3 -qarch=pwr9 -g1 \
  -Xopt --pass-remarks-output=loop.opt.yaml \
  -Xllc --pass-remarks-output=loop.llc.yaml

Set up the python3 environment for opt-viewer.py:

python3 -m virtualenv optviewer_env
. optviewer_env/bin/activate
python3 -m pip install pyyaml pygments

Use opt-viewer.py to generate the HTML report:

opt-viewer/opt-viewer.py --demangler cat -source-dir $PWD *.yaml -o report

The commands above generate a report directory containing optimization report loop.f.html. The report includes annotated source with transformation reports from ibm-opt and ibm-llc. Transformations that were performed are highlighted in green. Transformations that were missed are highlighted in red.