bswap: Recognize (int) __builtin_bswap64 (arg) idioms or __builtin_bswap?? (arg)...

bswap: Recognize (int) __builtin_bswap64 (arg) idioms or __builtin_bswap?? (arg) & mask [PR86723]

The following patch recognizes in the bswap pass (only there for now,
haven't done it for store merging pass yet) code sequences that can
be handled by (int32) __builtin_bswap64 (arg), i.e. where we have
0x05060708 n->n with 64-bit non-memory argument (if it is memory, we
can just load the 32-bit at 4 bytes into the address and n->n would
be 0x01020304; and only 64 -> 32 bit, because 64 -> 16 bit or 32 -> 16 bit
would mean only two bytes in the result and probably not worth it),
and furthermore the case where we have in the 0x0102030405060708 etc.
numbers some bytes 0 (i.e. known to contain zeros rather than source bytes),
as long as we have at least two original bytes in the right
positions (and no unknown bytes).  This can be handled by
__builtin_bswap64 (arg) & 0xff0000ffffff00ffULL etc.
The latter change is the reason why counting the bswap messages doesn't work
too well in optimize-bswap* tests anymore, while the pass iterates from end
of basic block towards start, it will often match both the bswap at the end
and some of the earlier bswaps with some masks (not a problem generally,
we'll just DCE it away whenever possible).  The pass right now doesn't
handle __builtin_bswap* calls in the pattern matching (which is the reason
why it operates backwards), but it uses FOR_EACH_BB_FN (bb, fun) order
of handling blocks and matched sequences can span multiple blocks, so I was
worried about cases like:
void bar (unsigned long long);
unsigned long long
foo (unsigned long long value, int x)
{
  unsigned long long tmp = (((value & 0x00000000000000ffull) << 56)
          | ((value & 0x000000000000ff00ull) << 40)
          | ((value & 0x00000000ff000000ull) << 8));
  if (x)
    bar (tmp);
  return (tmp
          | ((value & 0x000000ff00000000ull) >> 8)
          | ((value & 0x0000ff0000000000ull) >> 24)
          | ((value & 0x0000000000ff0000ull) << 24)
          | ((value & 0x00ff000000000000ull) >> 40)
          | ((value & 0xff00000000000000ull) >> 56));
}
but it seems we handle even that fine, while bb2 ending in GIMPLE_COND
is processed first, we recognize there a __builtin_bswap64 (value) & mask1,
in the last bb we recognize tmp | (__builtin_bswap64 (value) & mask2) and
PRE optimizes that into t = __builtin_bswap64 (value); tmp = t & mask1;
in the first bb and return t; in the last one.

2021-08-23  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/86723
	* gimple-ssa-store-merging.c (find_bswap_or_nop_finalize): Add
	cast64_to_32 argument, set *cast64_to_32 to false, unless n is
	non-memory permutation of 64-bit src which only has bytes of
	0 or [5..8] and n->range is 4.
	(find_bswap_or_nop): Add cast64_to_32 and mask arguments, adjust
	find_bswap_or_nop_finalize caller, support bswap with some bytes
	zeroed, as long as at least two bytes are not zeroed.
	(bswap_replace): Add mask argument and handle masking of bswap
	result.
	(maybe_optimize_vector_constructor): Adjust find_bswap_or_nop
	caller, punt if cast64_to_32 or mask is not all ones.
	(pass_optimize_bswap::execute): Adjust find_bswap_or_nop_finalize
	caller, for now punt if cast64_to_32.

	* gcc.dg/pr86723.c: New test.
	* gcc.target/i386/pr86723.c: New test.
	* gcc.dg/optimize-bswapdi-1.c: Use -fdump-tree-optimized instead of
	-fdump-tree-bswap and scan for number of __builtin_bswap64 calls.
	* gcc.dg/optimize-bswapdi-2.c: Likewise.
	* gcc.dg/optimize-bswapsi-1.c: Use -fdump-tree-optimized instead of
	-fdump-tree-bswap and scan for number of __builtin_bswap32 calls.
	* gcc.dg/optimize-bswapsi-5.c: Likewise.
	* gcc.dg/optimize-bswapsi-3.c: Likewise.  Expect one __builtin_bswap32
	call instead of zero.