[Bug target/31786] [4.1/4.2/4.3 Regression][avr] error: unable to find a register to spill in class 'BASE_POINTER_REGS'

["hutchinsonandy at netscape dot net" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #13 from hutchinsonandy at netscape dot net  2008-01-02 13:36 -------
Problem occurs because 1 base pointer is used for function pointer and the
other is used for frame pointer. (There are only 2). Backend
LEGITIMIZE_RELOAD_ADDRESS tells reload to use a base pointer for reload pass of
mem(SP+1). This fails.
Unfortunately reload does not figure out that it can use frame pointer and
gives ICE. (Ironically the frame pointer would be eliminated in this testcase
latter)

The problem potentially occurs whenever R30,R31 is tied up. (EEPROM macros are
other example)

Reload calls LEGITIMIZE_RELOAD_ADDRESS when strict address check fails
legitimate_address_p on SP+1 - since AVR has no way of indexing from stack
pointer.

Potential fix is to stop LEGITIMIZE_RELOAD_ADDRESS from telling reload
(push_reload)  to use BASE_POINTER class. POINTER CLASS is one wider and target
macros have pattern to use register 'REG_X' as index (with addition).
Alternatively don't suggest any substitution (ie not GOTO WIN) and reload finds
answer.

The latter solution has been tested and caused ICE in GCC library building as
'REG_X' is not allowable with strict address checking (legitimate_address_p ) -
so we get no pattern match ICE for MOVQI (REG_X+1),Rxx. However, backend does
contain such patterns. Changing legitimate_address_p to accept REG_X solves
problem.
In the testcase, the instruction that caused the ICE is eliminated - so the
newly allowable pattern is never used.

The build ICE caused by first change does indicate that on occasion GCC will
really use REG_X as index. Potentially producing sub-optimal code compared with
REG_Y or REG_Z. However, the incidence is clearly low and may well be offset by
other side effect improvements.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31786