[Bug target/82682] [8 Regression] FAIL: gcc.target/i386/pr50038.c scan-assembler-times movzbl 2 (found 3 times) since r253958

["law at redhat dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82682

--- Comment #6 from Jeffrey A. Law <law at redhat dot com> ---
Of course we can't propagate because %esi doesn't have a QImode equivalent.

So to back up a little bit.  The costing changes twiddled the register
assignments.  In particular reg93 gets allocated into %esi rather than %ebx.

That in turn means we have to emit a reload to handle this insn:

(insn 19 18 20 4 (set (reg:SI 109 [ xr ])
        (zero_extend:SI (reg/v:QI 93 [ xr ]))) "j.c":15 140 {*zero_extendqisi2}
     (expr_list:REG_DEAD (reg/v:QI 93 [ xr ])
        (nil)))

Because we can't use %esi in QImode.  So we get:



 (insn 36 18 19 4 (set (reg/v:QI 3 bx [orig:93 xr ] [93])
        (reg/v:QI 4 si [orig:93 xr ] [93])) "j.c":15 88 {*movqi_internal}
     (nil))
(insn 19 36 20 4 (set (reg:SI 4 si [orig:109 xr ] [109])
        (zero_extend:SI (reg/v:QI 3 bx [orig:93 xr ] [93]))) "j.c":15 140
{*zero_extendqisi2}
     (nil))


The only reaching def for the source of the zero extend in insn 19 is insn 36,
so REE will try to turn insn 36 into:

(insn 36 18 19 4 (set (reg:SI 4 si)
        (zero_extend:SI (reg/v:QI 4 si [orig:93 xr ] [93]))) "j.c":15 -1
     (nil))

Which, of course fails.  Even if it succeeded, it's not really an improvement. 
We really need to chase things back up to:

(insn 15 14 16 4 (set (reg/v:QI 4 si [orig:93 xr ] [93])
        (mem:QI (reg/v/f:SI 1 dx [orig:104 in ] [104]) [0 MEM[base: in_26,
offset: 0B]+0 S1 A8])) "j.c":13 88 {*movqi_internal}
     (nil))


I don't offhand know if we can really model that in REE or not.  Though there'd
almost certainly be a general benefit to x86 in doing so.