forwprop: Canonicalize atomic fetch_op op x to op_fetch or vice versa [PR98737]

forwprop: Canonicalize atomic fetch_op op x to op_fetch or vice versa [PR98737]

When writing the PR98737 fix, I've handled just the case where people
use __atomic_op_fetch (p, x, y) etc.
But some people actually use the other builtins, like
__atomic_fetch_op (p, x, y) op x.
The following patch canonicalizes the latter to the former and vice versa
when possible if the result of the builtin is a single use and if
that use is a cast with same precision, also that cast's lhs has a single
use.
For all ops of +, -, &, | and ^ we can do those
__atomic_fetch_op (p, x, y) op x -> __atomic_op_fetch (p, x, y)
(and __sync too) opts, but cases of INTEGER_CST and SSA_NAME x
behave differently.  For INTEGER_CST, typically - x is
canonicalized to + (-x), while for SSA_NAME we need to handle various
casts, which sometimes happen on the second argument of the builtin
(there can be even two subsequent casts for char/short due to the
promotions we do) and there can be a cast on the argument of op too.
And all ops but - are commutative.
For the other direction, i.e.
__atomic_op_fetch (p, x, y) rop x -> __atomic_fetch_op (p, x, y)
we can't handle op of & and |, those aren't reversible, for
op + rop is -, for - rop is + and for ^ rop is ^, otherwise the same
stuff as above applies.
And, there is another case, we canonicalize
x - y == 0 (or != 0) and x ^ y == 0 (or != 0) to x == y (or x != y)
and for constant y x + y == 0 (or != 0) to x == -y (or != -y),
so the patch also virtually undoes those canonicalizations, because
e.g. for the earlier PR98737 patch but even generally, it is better
if a result of atomic op fetch is compared against 0 than doing
atomic fetch op and compare it to some variable or non-zero constant.
As for debug info, for non-reversible operations (& and |) the patch
resets debug stmts if there are any, for -fnon-call-exceptions too
(didn't want to include debug temps right before all uses), but
otherwise it emits (on richi's request) the reverse operation from
the result as a new setter of the old lhs, so that later DCE fixes
up the debug info.

On the emitted assembly for the testcases which are fairly large,
I see substantial decreases of the *.s size:
-rw-rw-r--. 1 jakub jakub 116897 Jan 13 09:58 pr98737-1.svanilla
-rw-rw-r--. 1 jakub jakub  93861 Jan 13 09:57 pr98737-1.spatched
-rw-rw-r--. 1 jakub jakub  70257 Jan 13 09:57 pr98737-2.svanilla
-rw-rw-r--. 1 jakub jakub  67537 Jan 13 09:57 pr98737-2.spatched
There are some functions where due to RA we get one more instruction
than previously, but most of them are smaller even when not hitting
the PR98737 previous patch's optimizations.

2022-01-14  Jakub Jelinek  <jakub@redhat.com>

	PR target/98737
	* tree-ssa-forwprop.c (simplify_builtin_call): Canonicalize
	__atomic_fetch_op (p, x, y) op x into __atomic_op_fetch (p, x, y)
	and __atomic_op_fetch (p, x, y) iop x into
	__atomic_fetch_op (p, x, y).

	* gcc.dg/tree-ssa/pr98737-1.c: New test.
	* gcc.dg/tree-ssa/pr98737-2.c: New test.