mbox

[0/3] Dependency patches for hoist LIM code to cold loop

Message ID 20211208055416.1415283-1-luoxhu@linux.ibm.com
Headers

Message

Xionghu Luo Dec. 8, 2021, 5:54 a.m. UTC 
  This patchset is a recollect of previously sent patches.  Thanks
Richard that The "Don't move cold code out of loop by checking bb count"
is approved[1], but there are still 3 prerequesite patches to supplement
or avoid regression.

1) Patch [1/3] is the RTL part of not hoisting LIM code out of cold loop, it
could improve perlbench by 7.69% [2].
2) Patch [2/3] is a test case regression fix for pr103270.c, after enabling
gimple part of hoisting LIM code to coldest loop [1], the store
instruction in loop won't be moved out of inner loop, it is caused by a
jump-threading patch unexpectedly turning a hot inner loop to cold loop,
this patch could recover the inner loop to be hot[3].
3) As data showed in [2], besides improvement, there is also a small regression
on SPEC2017 544.nab_r (-1.55%).  After investigation, it turned out to be
the profile count and probability is not correctly adjusted in loop
split, with this patch [3/3], the only regression is also fixed. This version
slightly updates [4] to fix ICEs.

[1] https://gcc.gnu.org/pipermail/gcc-patches/2021-December/586319.html
[2] https://gcc.gnu.org/pipermail/gcc-patches/2021-September/580109.html
[3] https://gcc.gnu.org/pipermail/gcc-patches/2021-November/585195.html
[4] https://gcc.gnu.org/pipermail/gcc-patches/2021-November/585290.html

Xionghu Luo (3):
  loop-invariant: Don't move cold bb instructions to preheader in RTL
  Fix incorrect loop exit edge probability [PR103270]
  Fix loop split incorrect count and probability

 gcc/loop-invariant.c            | 10 ++--
 gcc/predict.c                   | 10 ++--
 gcc/tree-ssa-loop-split.c       | 85 +++++++++++++++++++++++++++++----
 gcc/testsuite/gcc.dg/pr103270.c | 19 ++++++++
 4 files changed, 109 insertions(+), 15 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/pr103270.c