[LLVMdev] Folding nodes with more than one use during ISel
stephen.montgomery3 at btinternet.com
Wed Sep 26 04:21:48 PDT 2012
I'm working on a backend for the Freescale CPU12 family as a hobby project and I'm having difficulty getting the instruction selection pass to handle the indirect indexed addressing modes. I'd really appreciate advice on how best to handle them.
The following llvm instructions:
%arrayidx = getelementptr inbounds i8** %p, i16 3
%0 = load i8** %arrayidx, align 2, !tbaa !0
%1 = load i8* %0, align 1, !tbaa !1
%neg = xor i8 %1, -1
store i8 %neg, i8* %0, align 1, !tbaa !1
can be realised with the single CPU12 instruction which uses the indirect indexed addressing mode:
I've defined a multiclass for this type of instruction in my InstrInfo.td and the relevant entry for this addressing mode is:
def IdxI16 : IdxPB2Inst<opcode, AddrModeIdxI16, (outs), (ins memsrc:$mem),
!strconcat (str, "\t[$mem]\t; IdxI16"),
[(store (OpNode (loadi8 (loadi16 addrIAny:$mem))),
where loadi8 and loadi16 are PatFrags that match an i8 and i16 load respectively and OpNode is, in this case, the "not" PatFrag. Pointers are 16-bit by the way.
During instruction selection, this pattern fails to match and it seems to me that it's because the store uses %0 directly as an operand but also uses it by virtue of %neg (as well as via a chain). However, since all the nodes between the store via %neg down to %0 are going to be folded into the instruction, it feels like the test implemented in IsLegalToFold is too restrictive in this case.
My questions are:
- has IsLegalToFold been designed this way intentionally because there is a problem with folding in the way I want, or is it that no other targets need this behaviour so it's not been implemented yet?
- assuming it's the latter, are there any obvious pitfalls I should watch out for if I modify it
- if I do modify it successfully, would the change be helpful for anyone else
- is there a better way of doing what it want that doesn't involve modifying SelectionDagISel.cpp
I'm probably going to drop into the Cambridge social this evening so if there are any experts on this subject likely to be there, I'd be grateful for a chat.
More information about the llvm-dev