Need Advice on AVX

Project Infrastructure LLVM Dev List Archives
1

Ok, I am tracking down some bugs in our AVX stuff and came upon
an interesting conundrum.

The MOVQ instruction (MOVPQIto64rr in X86Instr64bit.td) only takes
xmm registers. There is no ymm version since the xxm's are subregisters.

I need to be able to match a vector element extract of element 0 on a
v4i64 vector. Obviously this is not a legal operation even with AVX
because MOVQ only operates on xmms.

So I can mark it as not legal but how should it be lowered? I can't
bitcast to a v2i64 because the vector sizes are different. I could
do an extract_subreg and then write a pattern to match that to MOVQ.

Anyone have other thoughts?

                              -Dave

2

What does this mean?

def MOVSX16rr8 : I<0xBE, MRMSrcReg, (outs GR16:$dst), (ins GR8 :$src),
                   "movs{bl|x}\t{$src, ${dst:subreg32}|${dst:subreg32},
$src}",
                   [(set GR16:$dst, (sext GR8:$src))]>, TB;

How does ${dst:subreg32} work? Can one do the same for sources? I
don't see an references to "subreg" in the TableGen sources. I see
it handled in the X86 AsmPrinter. Looks like it should work for
sources too.

Is it preferable to use the source modifier or write an EXTRACT_SUBREG
pattern explicitly?

                              -Dave

3

Hello, David

How does ${dst:subreg32} work?

This is just modifier provided to asmprinting code. Here, it seems, 16
bit register is passed to asmprinter, but it sees modifier and grabs
32-bit superreg.

Can one do the same for sources?

Yes, this is just modifier for printing, nothing more...

Is it preferable to use the source modifier or write an EXTRACT_SUBREG
pattern explicitly?

It depends what you're want to do. But I guess you need to model
subreg access properly...

4

Modeling subregisters isn't hard. Do you have some guidance as to when
one method is preferable? I am leaning toward using the modifier since
conceptually, a vector_extract of element zero on a v4i64 makes sense with AVX
(so it is "legal"). You just have to emit the register name as "xmm" rather
than "ymm." Why write an additional complicated pattern for this case?

                                   -Dave

5

Please don't use asmprinter modifiers. I'm trying to remove them as part of the MCization code. Please use the SUBREG machine instr and (subreg) tblgen node as appropriate.

-Chris

6

Ok.

                            -Dave