On 5/26/26 10:22 AM, Robin Dapp wrote:
From: Robin Dapp<[email protected]>
Building upon Richard's register filters, this patch series introduces filters
that can base their filtering decision on a second operand: The constrained
operand (the "dependent" operand), and the "referenced" operand.
This allows e.g. to constrain the dependent operand relative to the mode of an
output operand. For RISC-V, a typical example is when a source of a widening
vector instruction must either not overlap the destination at all or only a
specific subreg of the destination.
Your and Jeff's discussion with Richard was useful to understand the
problem riscv developers ran into.
A major difference to register filters is that dependent filters are dynamic
and can therefore be significantly heavier in compile-time. Also, by their
nature, they are order-dependent, influencing greedy coloring heuristics.
I haven't observed a compile-time hit across SPEC but so far only tested with
a single dependent filter. I wouldn't expect too many dynamic filters active
at the same time and still added a simple caching scheme.
That is good. I mostly worried about targets w/o filters. For them
additional resources are minimal (like one null pointer for
dependent_filters in lra_reg_info). As the patch is a concern now only
for RISCV maintainers you can address compilation speed later if there
is any.
I would not bother about implementation of filters in IRA fast
allocation. Fast allocation is just mostly for improving compilation
speed in -O0 mode. It generates smaller code than assigning always
memory as it was before. Therefore compilation speed is better than one
when assigning always memory as there are less insns.
It is difficult to me to say what will be faster fast allocation w/o
filters or with them because LRA will do work skipped in IRA. So I guess
using filters in fast allocation is OK as it is already implemented.
Here are some minor comments about the code:
o INVALID_REGNUM check in ira_dependent_filter (ira-color.cc) is
redundant:
if (ref_regno == INVALID_REGNUM
|| ref_regno >= FIRST_PSEUDO_REGISTER)
continue;
as INVALID_REGNUM is max(usigned int). But there is no harm,
compiler will remove it anyway.
o duplicated declaration of lra_get_dependent_filter in lra.h and
ira-int.h. I'd prefer to keep it only in lra.h (external interface of
LRA) and include "lra.h" in ira-color.cc
o lra_reset_dependent_filters () -> lra_reset_dependent_filters (void)
Definition has void but declaration does not. For C++ it is the
same. Simply it catches my eye as mostly a C programmer.
You can completely ignore my comments. They are quite minor.
So IRA/LRA patches are OK for me. Thank you for your work.
The series is organized similarly to Richard's. First, the gen plumbing, then
recog. The lra change should be sufficient for correctness and the ira patch
improves register allocation. The RA changes have been slightly
"whack-a-mole"y. I just added handling until I got the expected results.
Further input and ideas towards a more comprehensive solution appreciated.
Rather than keeping it to myself for much longer (sorry Richard for CC'ing you
again while backprop is still in flight, but I figured you're interested), I
now decided to send out the current version.
The series was regtested and bootstrapped on x86, power10, and aarch64.
Regtested on riscv64.
Robin Dapp (5):
genpreds: Dependent, dynamic register filters.
recog: Handle dependent filters.
lra: Support dependent filters.
RISC-V: Example dependent filter for widen overlap.
ira: Add dependent filter handling.
gcc/config/riscv/constraints.md | 8 +
gcc/config/riscv/riscv-protos.h | 2 +
gcc/config/riscv/riscv.cc | 35 +++
gcc/config/riscv/vector.md | 16 +-
gcc/doc/md.texi | 19 +-
gcc/doc/tm.texi | 3 +
gcc/doc/tm.texi.in | 3 +
gcc/genconfig.cc | 1 +
gcc/genpreds.cc | 166 ++++++++++++-
gcc/gensupport.cc | 37 ++-
gcc/gensupport.h | 2 +
gcc/ira-build.cc | 25 ++
gcc/ira-color.cc | 63 ++++-
gcc/ira-conflicts.cc | 52 ++++-
gcc/ira-int.h | 29 +++
gcc/ira-lives.cc | 92 ++++++++
gcc/ira.cc | 1 +
gcc/lra-assigns.cc | 23 ++
gcc/lra-constraints.cc | 221 ++++++++++++++++++
gcc/lra-int.h | 22 ++
gcc/lra.cc | 8 +-
gcc/lra.h | 3 +
gcc/recog.cc | 33 ++-
gcc/recog.h | 16 ++
gcc/rtl.def | 12 +-
.../gcc.target/riscv/rvv/base/pr112431-4.c | 2 +-
.../gcc.target/riscv/rvv/base/pr112431-5.c | 2 +-
.../gcc.target/riscv/rvv/base/pr112431-6.c | 2 +-
28 files changed, 862 insertions(+), 36 deletions(-)
