[Bug target/114252] Introducing bswapsi reduces code performance

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114252

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2024-03-06
             Status|UNCONFIRMED                 |NEW
          Component|tree-optimization           |target
           Keywords|                            |missed-optimization

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  It looks like we do no cost evaluation in
maybe_optimize_vector_constructor but checking that there's an optab
for bswap with SImode.

insn-flags.h:#define HAVE_bswapsi2 1

but somehow we end up doing a libcall?

We expand as

;; bswapdst_10 = __builtin_bswap32 (load_dst_9);

(insn 6 5 7 (set (reg:SI 47)
        (mem:SI (reg/v/f:HI 46 [ buf ]) [0 MEM <long unsigned int> [(const
uint8_t *)buf_5(D)]+0 S4 A8])) -1
     (nil))

(insn 7 6 8 (set (reg:SI 22 r22)
        (reg:SI 47)) -1
     (nil))

(insn 8 7 9 (set (reg:SI 22 r22)
        (bswap:SI (reg:SI 22 r22))) -1
     (nil))

^^^

so why does that turn into a library call?

I think this is mis-communication between the middle-end and the target.