[Bug rtl-optimization/61605] Potential optimization: Keep unclobbered argument registers live across function calls

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61605

vries at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2014-09-29
                 CC|                            |vries at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #4 from vries at gcc dot gnu.org ---
> If a function is known to not clobber an argument register then the caller
> shouldn't have to save/reload that register across the function call.

If a function is known to not clobber an call_used_reg then the caller
can use it as a non-call_used_reg across the function call.

This diff shows the example with -fno-use-caller-save vs -fuse-caller-save:
...
 foo:
 .LFB1:
     .cfi_startproc
-    pushq    %rbx
-    .cfi_def_cfa_offset 16
-    .cfi_offset 3, -16
-    movl    %edi, %ebx
+    movl    %edi, %edx
     call    bar
-    addl    %ebx, %eax
-    popq    %rbx
-    .cfi_def_cfa_offset 8
+    addl    %edx, %eax
     ret
     .cfi_endproc
 .LFE1:
...
-fuse-caller-save removes the entry/exit save/restore pair
'pushq %rbx'/'popq %rbx'.

The 'movl %edi, %edx' is indeed non-optimal, but it's not a 'save' in the sense
of save/restore pair generated at function entry/exit or around function calls.
It's a copy at function entry of a hard reg argumument to a pseudo reg,
generated at expand, which is followed by a copy of the pseudo reg to the same
register to set the argument for the function call:
...
(insn 2 4 3 2 (set (reg/v:SI 86 [ yD.1755 ])
        (reg:SI 5 di [ yD.1755 ])) test.c:9 -1
     (nil))
(note 3 2 6 2 NOTE_INSN_FUNCTION_BEG)
(insn 6 3 7 2 (set (reg:SI 5 di)
        (reg/v:SI 86 [ yD.1755 ])) test.c:10 -1
     (nil))
...
The second insn is removed in pass_fast_rtl_dce. The reg-alloc choiche for
pseudo 86 in the first insn is dx, and the insn remains.

I think there could be two ways to address this:
1. Teach a pass after ira, like pass_cprop_hardreg or pass_gcse2 to use the
   information collected by fuse-calller-save.
2. Teach ira to prefer the dx to di in this case.

My guess would be pass_cprop_hardreg.