[Bug rtl-optimization/123652] [13/14/15/16 Regression] Unnecessary x64 code generation when returning zero-initialized 7-byte types

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123652

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
The reason for the missed CSE from the "bad" initial RTL is likely the overlap
trick we're using for the initialization of the stack temporary:

;; D.18588 = {};

(insn 5 4 6 (set (mem/c:SI (plus:DI (reg/f:DI 93 virtual-stack-vars)
                (const_int -7 [0xfffffffffffffff9])) [0 D.18588+0 S4 A8])
        (const_int 0 [0])) "t.C":5:13 -1
     (nil))

(insn 6 5 0 (set (mem/c:SI (plus:DI (reg/f:DI 93 virtual-stack-vars)
                (const_int -4 [0xfffffffffffffffc])) [0 D.18588+3 S4 A8])
        (const_int 0 [0])) "t.C":5:13 -1
     (nil))

CSE fails to handle

(insn 8 7 9 (set (reg:DI 100)
        (zero_extend:DI (mem/c:SI (plus:DI (reg/f:DI 93 virtual-stack-vars)
                    (const_int -7 [0xfffffffffffffff9])) [0 D.18588+0 S4 A8])))
"t.C":5:13 discrim 1 -1
     (nil))

with the partial clobber from the 2nd store. That trick is "new" and thus
likely the bisect point.