Help with PPC64 JIT

Project Infrastructure LLVM Dev List Archives
#1

Hello,

I am currently working with PPC64 JIT support for LLVM. So far I could make function calls
work by adding function descriptors in 'lib/Target/PowerPC/PPCJITInfo.h' and adding a
virtual method at 'LLVM::TargetJITInfo' that is called within 'JITEmitter::finishFunction'
just after 'sys::Memory::InvalidateInstructionCache' to update the Global Mapping with
function descriptor instead of the function address. The JIT function descriptor is
loaded correctly in 'JIT::runFunction', instead of assuming the JIT function code is
an ODP.

Now I'm trying to make the relocation work properly. Using the testcase '2003-01-04-ArgumentBug'
the assembly generated for main functions is the following:

.L.main:
# BB#0:
        mflr 0
        std 0, 16(1)
        stdu 1, -112(1)
        lis 3, .LCPI1_0@ha
        li 4, 1
        lfs 1, .LCPI1_0@l(3)
        li 3, 0
        bl foo
        nop
        addi 1, 1, 112
        ld 0, 16(1)
        mtlr 0
        blr

Which is correct, however for the JIT one generated in memory the relocations generate some issues.

First the 'lis 3, .LCPI1_0@ha' can possible overflow which will generate an wrong relocation.
Since the const data will be place just before the function code in JIT generation, my first
approach was to point the functions descriptor TOC to the JIT function base, so the 'lis' relocation
could be rewritten as 'addis 3,2,.LCPI1@TOC@ha'.

And there where I could use some help: is it the best approach? Where is the best place to make this
kind of analysis? Is there another way to make code adjustments for JIT?

At fist I though to put the logic at 'lib/Target/PowerPC/PPCJITInfo.cpp', but from what I could
understand the 'relocate' function method is indeed just to operate on the relocation
addresses, not to change the upcode.

Any advices/tips/suggestion would be appreciated.

#2

Hi Adhemerval Zanella, the old JIT infrastructure is going away, to be replaced
by "MC-JIT" (try passing -use-mcjit to lli). It sounds like you are working on
the old JIT, so I suggest you work instead on getting MC-JIT working on powerpc.

Ciao, Duncan.

#3

Hi Adhemerval Zanella, the old JIT infrastructure is going away, to be replaced
by "MC-JIT" (try passing -use-mcjit to lli). It sounds like you are working on
the old JIT, so I suggest you work instead on getting MC-JIT working on powerpc.

Hi Duncan,
   Thanks for the pointers. We hadn't stumbled across the MC-JIT refs
in our digging so far. (LLVM GettingStarted.html doesn't mention it,
and (FIXME: T.B.D.) under llvm/lib/mc, so I suppose the MC jit is fairly
recent?)
Are we safe to assume the old JIT infrastructure will be completely
replaced, or are there likely to be 'users' that may still need
functionality?
I imagine since the old jit is where Adhemerval was focused, the
build/test steps must still be referencing it.

Thanks,
-Will

#4

Thanks for the reply Duncan, I shifted my focus fo MCJIT. So far I could check there are basically
two issues with MCJIT relating to PPC64 support:

1. Looks like LLVM does not generate PPC64 PIC code with TOC references. Using the testcase
'2003-01-04-ArgumentBug.ll' I tried to invoke the llc with '-march=ppc64 -mcpu=ppc64 -mattr=64bit'
and still uses the 32-bits relocations (@ha/@l(3)) where TOC would be expected (@toc). I did some
adjustments to JIT code and function calling is correct (the function ODP is created and called),
but BSS access is still not right since the 64-bits relocations are not done correctly.
Could someone point where should I focus my efforts to add TOC support on PowerPC backend?

2. Relocations: I could add some support for PPC32 relocations by implementing some functions at
'RuntimeDyldELF.cpp', but then I fall back again in 1. (the PPC64 support is incomplete).

Is my analysis correct or I am issuing llc/lli wrongly and not activating the PPC64 PIC support?
Anyone to advice me on this one? Any help would be appreciated.

#5

The modifications I have done so far are in attachments. Basically it adds the PPC64 machine
name in Target, add some relocation support in Elf Runtime class. I'm still figuring out a
lot of internal llvm and the patch is still WIP.

mcjit-ppc64.patch (6.25 KB)

#6

>>> Hi Adhemerval Zanella, the old JIT infrastructure is going away,
>>> to be replaced by "MC-JIT" (try passing -use-mcjit to lli). It
>>> sounds like you are working on the old JIT, so I suggest you work
>>> instead on getting MC-JIT working on powerpc.
>> Hi Duncan,
>> Thanks for the pointers. We hadn't stumbled across the MC-JIT
>> refs in our digging so far. (LLVM GettingStarted.html doesn't
>> mention it, and (FIXME: T.B.D.) under llvm/lib/mc, so I suppose
>> the MC jit is fairly recent?)
>> Are we safe to assume the old JIT infrastructure will be completely
>> replaced, or are there likely to be 'users' that may still need
>> functionality?
>> I imagine since the old jit is where Adhemerval was focused, the
>> build/test steps must still be referencing it.
>>
>> Thanks,
>> -Will
> Thanks for the reply Duncan, I shifted my focus fo MCJIT. So far I
> could check there are basically two issues with MCJIT relating to
> PPC64 support:
>
> 1. Looks like LLVM does not generate PPC64 PIC code with TOC
> references. Using the testcase '2003-01-04-ArgumentBug.ll' I tried
> to invoke the llc with '-march=ppc64 -mcpu=ppc64 -mattr=64bit' and
> still uses the 32-bits relocations (@ha/@l(3)) where TOC would be
> expected (@toc). I did some adjustments to JIT code and function
> calling is correct (the function ODP is created and called), but
> BSS access is still not right since the 64-bits relocations are not
> done correctly. Could someone point where should I focus my efforts
> to add TOC support on PowerPC backend?
>
> 2. Relocations: I could add some support for PPC32 relocations by
> implementing some functions at 'RuntimeDyldELF.cpp', but then I
> fall back again in 1. (the PPC64 support is incomplete).
>
> Is my analysis correct or I am issuing llc/lli wrongly and not
> activating the PPC64 PIC support? Anyone to advice me on this one?
> Any help would be appreciated.
>
The modifications I have done so far are in attachments. Basically it
adds the PPC64 machine name in Target, add some relocation support in
Elf Runtime class. I'm still figuring out a lot of internal llvm and
the patch is still WIP.

Great! Thanks for sharing this. I'll send comments on the patch to the
commits list.

-Hal

#7

Hi Adhemerval,

It looks like the default relocation model for PowerPC is "DynamicNoPIC" for the Darwin OS and "Static" elsewhere. There is a command line option ("relocation-model") to override this default in both lli and llc. However, I'm not sure using this will reduce your troubles. You may end up going down code paths that have not been sufficiently exercised in the past.

I don't really know anything specifically about PowerPC, but I think that you may be able to glean some guidance from a patch which Petar Jovanovic recently submitted to add MCJIT support for MIPS. The patch (which hasn't yet been committed) is here:

http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20120723/146977.html

The big difference I see between Petar's patch and yours is that Petar added explicit ELFWriterInfo support for his platform. This is outside the area of the code that I know well, but it looks like that might be the right direction for you to investigate to fix your relocation difficulties.

-Andy