Richard Sandiford - Re: [RTL, ColdFire 24/63] Add support for a MODE_INDEX_REG_CLASS macro (original) (raw)

This is the mail archive of the gcc-patches@gcc.gnu.orgmailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Jeffrey Law law@redhat.com writes:

My concern about efficiency still stands though. Concretely, your patch is going to mean that the MIPS port does (29 modes x 188 regs =) 5452 legitimate address checks at startup, all of which will be false. The figure for PowerPC is (38 modes x 114 regs =) 4332 checks and the figure for ia64 is (40 modes * 334 regs =) 13360 checks. (I realise that ia64 is one of those targets where you expect the compiler to be slower, but still. ;)) As the patch stands, it will also create new garbage addresses for each check. I think at the very least you should:

  • create one address and use it for all checks, changing parts of it as appropriate Definitely. Please go ahead and do this.

Sorry, but I still strongly believe you're leading us in the wrong direction here. I suggest we leave it for someone else with less strong feelings. And...

I was merely showing a quick proof of concept, it wasn't necessarily meant to go into the tree as-is, but instead show you the way to getting this problem cleanly resolved.

...with respect Jeff, I didn't need you to tell me how the double_reg_address_ok thing should be implemented. I was resisting doing it because I disagreed with it, not because I didn't understand the concept.

Richard

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]