[avr-gcc-list] invalid use of reg_unused_after in avr.c

[Tristan Gingold]

Hi,

while trying to understand why gcc generates buggy code, I wondered if  
it is safe to use reg_unused_after
(and more particularly dead_or_set_p) while emitting asm.
It looks like liveness information is not maintained after reload.

Here is the issue:

At pass 188.ce3:

BB 18:
...
(insn 166 165 167 18 c730a01.adb:62 (set (cc0)
        (reg:HI 16 r16 [ D.1758+2 ])) 87 {tsthi} (expr_list:REG_DEAD  
(reg:HI 16 r16 [ D.1758+2 ])
        (nil)))
...

BB 20 (its only predecessor is BB 18)
...
(insn 693 176 177 20 c730a01.adb:62 (set (reg:HI 30 r30)
        (reg/f:HI 12 r12 [orig:68 SR.20 ] [68])) 8 {*movhi} (nil))
...

But after 190.cprop_hardreg:
(insn 166 165 167 18 c730a01.adb:62 (set (cc0)
        (reg:HI 16 r16 [ D.1758+2 ])) 87 {tsthi} (expr_list:REG_DEAD  
(reg:HI 16 r16 [ D.1758+2 ])
        (nil)))

(insn 693 176 177 20 c730a01.adb:62 (set (reg:HI 30 r30)
        (reg:HI 16 r16 [orig:68 SR.20 ] [68])) 8 {*movhi} (nil))

Clearly the use of r16 in insn 693 is wrong as the tsthi (which  
generates or r16, r17) kills r16.
As a consequence, the code generated is wrong.

AVR seems to be the only target to use dead_or_set_p so late.  Other  
targets use peephole optimizations
to address such cases.

Thoughts ?

Tristan.