I observed that HARD_REGNO_CALL_PART_CLOBBERED gets called with hard registers that HARD_REGNO_MODE_OK would reject.
Is it save to set HARD_REGNO_CALL_PART_CLOBBERED to FALSE for hard registers for which HARD_REGNO_MODE_OK is FALSE? Background is PR53593 where I see a code size increase of +10% for avr. HARD_REGNO_CALL_PART_CLOBBERED gets called with REGNO=29, MODE=HI This register is not a valid hard register because HI regs must start with even register numbers. If it /was/ valid, then HARD_REGNO_CALL_PART_CLOBBERED should return TRUE for HI:29. That is what the hook macro currently does. However, if the hook macro is defined like so, then the performance regression goes away: #define HARD_REGNO_CALL_PART_CLOBBERED(REGNO, MODE) \ (HARD_REGNO_MODE_OK (REGNO, MODE) && (more conditions)) Questions: 1: Is it save to define the macro like that? I.e. return FALSE for a register that is invalid but would be call-part-clobbered if it was valid? 2: What's the point of calling the macro with invalid hard register? I can understand this to keep the call sites simple, but I don't understand why changing the properties for an invalid hard register affects code generation that much. Thanks, Johann
