[LLVMdev] Strange spill behaviour

[Lee Hammerton]

Hi,

 While working some more on the backend, I've added callee register saving
and have come across something odd (I assume its because i've not
implemented something), this is with optimisations which makes it even more
odd :

      MOV.L   [R7+ 12], R1            ; 4-byte Folded Spill
      MOV.L   [R7+ 8], R2             ; 4-byte Folded Spill
      ADD.L   R0,R0 + R1
      ADD.L   R0,R0 + #7
      SUB.L   R0,R0 - R2
      MOV.L   R2, [R7+ 8]             ; 4-byte Folded Reload
      MOV.L   R1, [R7+ 12]            ; 4-byte Folded Reload

 I've isolated the epilogue and prologue away from the function (they don't
make use of R1 or R2, Calling convention states R0-R3 are used as inputs to
a function. R1-R4 are marked as calleesaved. From the above code I cannot
see any reason why R1 & R2 are being spilled, I assume the fault lies in my
tablegen definitions (relevant ones reproduced below) :

let isCommutable=1 in
{
def ADD32rrr : M819Inst<(outs GR32:$dst),(ins GR32:$src1,
GR32:$src2),"ADD.L\t{$dst,$src1 + $src2}",[(set GR32:$dst, (add GR32:$src1,
GR32:$src2))]>;
}
def ADD32rri : M819Inst<(outs GR32:$dst),(ins GR32:$src1,
i32imm:$src2),"ADD.L\t{$dst,$src1 + $src2}",[(set GR32:$dst, (add
GR32:$src1, imm:$src2))]>;

def SUB32rrr : M819Inst<(outs GR32:$dst),(ins GR32:$src1,
GR32:$src2),"SUB.L\t{$dst,$src1 - $src2}",[(set GR32:$dst, (sub GR32:$src1,
GR32:$src2))]>;

def MOV32rm : M819Inst<(outs GR32:$dst), (ins memsrc:$src),"MOV.L\t{$dst,
[$src]}",[(set GR32:$dst,(load addr:$src))]>;

def MOV32mr : M819Inst<(outs), (ins memdst:$dst,
GR32:$src),"MOV.L\t{[$dst], $src}",[(store GR32:$src, addr:$dst)]>;


 Am I misunderstanding something here?



 Lee
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130320/5b61a24a/attachment.html>