On Thu, Sep 09, 2021 at 08:16:16AM +0200, Richard Biener wrote: > But subreg _is_ bit_cast. What is odd to me is that a "disallowed" subreg > like (subreg:SF (reg:TI ..) 0) magically becomes valid (in terms of > validate_subreg) if you rewrite it as (subreg:SF (subreg:SI (reg:TI ..) 0) 0). > Of course that's nested and invalid but just push the inner subreg to a > new pseudo and the thing becomes valid.
Basically I only checked the mode of the SUBREG and the value in SUBREG_REG. I agree (subreg:SF (subreg:SI (reg:TI))) should not be allowed. When I wrote the changes we didn't see SUBREG's as much to access the bottom or top element of a structure that was held in a larger int field. If the patch is not reverted, then we have to deal with this somehow. Such as with my patch that allows (subreg:SF (reg:TI xxx)) and (subreg:SF (reg:DI xxx)). The target hook idea might be too big of a hammer. I would need to play with it to see if there are other losses. On the PowerPC, the problematical subregs only occur when you are moving a register from GPR to a vector/float register or vice versa. If you move a subreg between two GPR registers or two vector/float registers, then you can just do the copy. Part of the issue is for some moves, you need to use a temporary register to convert the results, but of course we have the long standing 'all moves for a mode must be done in a single insn that covers all of the cases'. Then you get into secondary reload territory. -- Michael Meissner, IBM IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA email: meiss...@linux.ibm.com, phone: +1 (978) 899-4797
