[PATCH 0/2] match.pd: recognise abs-difference, plain and promoted

Two patches recognising the conditional spelling of absolute difference.

Patch 1 folds A > B ? A - B : B - A (and the lt/le forms) to abs (A - B)
for signed non-wrapping operands: on aarch64 the scalar becomes
subs+csneg, and the diamond no longer blocks further simplification.
With this series every testcase in the long-open
PR50856 folds to abs: both functions from the original report (the
promoted-byte one needs patch 2) and all four from comment 4 (f and f3
already fold today; f1 and f2 need patch 1).  The PR's Bugzilla
dependencies PR111910 and PR114999 concern the general
negation-detection and comparison-simplification infrastructure and
stand on their own; none of the PR's own testcases needs them once this
series is in.

Patch 2 extends it to the promoted spelling, which is what C actually
produces for sub-int types: the front end narrows the promoted
comparison back to the original type, so the compare tests A' and B'
while the arms compute (int) A' - (int) B', and patch 1's same-operand
pattern cannot match.  A strictly widening conversion to the signed
result type preserves the comparison's order and the widened difference
cannot overflow, so the fold carries over.  This is the scalar shape of
every byte sum-of-absolute-differences loop: with it, such a loop
vectorises through the abd patterns -- on aarch64 the loop body becomes
uabdl/uabdl2 where it previously was a compare, two widening-subtract
pairs and a select.  A same-width conversion is excluded, since
reinterpretation changes the comparison's order.

Testing.  Bootstrapped as a series on x86_64-pc-linux-gnu at trunk
844c665f170 with the stage2/stage3 comparison successful, and a full
make -k check run: 228036 gcc, 278459 g++ and 20095 libstdc++ expected
passes.  Every unexpected result was also present or reproducible with
the series reverted, including a pre-existing mdspan dg-error
regression (23_containers/mdspan layouts/submdspan, 14 checks over two
files) that reproduces with an unpatched compiler at the same revision.
The three new tests were additionally run with an arm-none-eabi cross
of this exact series: all eleven checks pass, including the five
ARM-specific assembler scans.  With patch 2's fold reverted and its
test kept, the ABS_EXPR scans fail (0 instead of 4).

These patches were prepared with the assistance of an AI coding tool,
recorded per-patch with Assisted-by:.  I reviewed and verified every
line, and take responsibility for the result.

Dominic P (2):
  match.pd: recognise A > B ? A - B : B - A as abs (A - B) [PR50856]
  match.pd: fold the promoted spelling of abs-difference [PR50856]

 gcc/match.pd                                  | 58 +++++++++++++++++++
 gcc/testsuite/gcc.dg/tree-ssa/absdiff-1.c     | 12 ++++
 .../gcc.dg/tree-ssa/absdiff-widen-1.c         | 20 +++++++
 gcc/testsuite/gcc.target/arm/pr50856.c        | 35 +++++++++++
 4 files changed, 125 insertions(+)
 create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/absdiff-1.c
 create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/absdiff-widen-1.c
 create mode 100644 gcc/testsuite/gcc.target/arm/pr50856.c

-- 
2.55.0