similar to: [LLVMdev] MC-JIT Patches 2/3

Hi Jan, I would rather not work with a patch this large. Can you pull out the addition of the MCJITStreamer into its own patch, and we can iterate on getting that in as a single commit? I realize it won't work or do anything useful, but I can't deal with reviewing patches this large. The main thing I am concerned about is getting the basic design of how the streamer and the assembler and

I was delayed creating the smaller patches, but finally I had some time to put the first set together. There are three small patches, the first two are classes the MCJITStreamer uses, and the last patch is the MCJITStreamer class itself. - Jan --- On Sun, 8/1/10, Daniel Dunbar <daniel at zuster.org> wrote: > From: Daniel Dunbar <daniel at zuster.org> > Subject: Re: [LLVMdev]

I have cleaned up the code somewhat that Olivier wrote and split up the patch into three pieces. This first is to make the MCJIT not have to initialize all asm printers, but only the native one. - Jan -------------- next part -------------- A non-text attachment was scrubbed... Name: 0018_nativeasmprinterinit.patch Type: text/x-diff Size: 12993 bytes Desc: not available URL:

I cleaned up the indentation in the three patches and attached them. Please let me know if there are other issues. The patches should all apply cleanly with the latest revision. I will wait with submitting more patches until the MCJITSTreamer has been reviewed and checked in. Thanks, Jan --- On Sat, 8/21/10, Bruno Cardoso Lopes <bruno.cardoso at gmail.com> wrote: From: Bruno Cardoso

Together with Jan Sjodin (in copy of this email), we begin an implementation of the JIT with MC. The idea, suggested by Jan, is to develop a MCJIT in parallel of the current JIT and to keep the two implementations until (at least) the new MC one is mature enough. Currently code is kept on gitorious (http://gitorious.org/llvm-mc-jit/llvm-mc-jit). Following this, a boolean "bool MCJIT =

We're migrating to LLVM 3.1 and trying to use the upstream FMA patterns. Why is VFMADDSD4 defined with vector types? Is this simply because the gcc intrinsic uses vector types? It's quite unnatural if you have a compiler that generates FMAs as opposed to requiring user intrinsics. -Dave

Hi Jan, Applied with edits in r109996, thanks! I wasn't happy about the change to add "Target" everywhere -- I just added a local hack in TargetSelect.h to workaround this. We should really change the definition of LLVM_NATIVE_TARGET, but I am not in a configure hacking mood. - Daniel On Wed, Jul 28, 2010 at 10:39 AM, Jan Sjodin <jan_sjodin at yahoo.com> wrote: > I have

Better be quick! I am adding FMA4 and XOP now, and if you contribute code before I do, you can spare yourself some XOP merging. - Jan ----- Original Message ----- > From: David A. Greene <greened at obbligato.org> > To: Benjamin Kramer <benny.kra at googlemail.com> > Cc: llvmdev at cs.uiuc.edu > Sent: Thursday, December 1, 2011 12:19 PM > Subject: Re: [LLVMdev]

8th will work for me. Can we pick a place that is not overly noisy? - Jan >________________________________ > From: Rafael Espíndola <rafael.espindola at gmail.com> >To: Ehsan Akhgari <ehsan.akhgari at gmail.com> >Cc: Jeff Muizelaar <jmuizelaar at mozilla.com>; clang-dev Developers <cfe-dev at cs.uiuc.edu>; "Minard, Brian" <brian.minard at

Ah, bad example. This is a general problem for all (maybe most) SSE and AVX SS/SD patterns though, which is why I mentioned Sandybridge. You can swap out VFMADDSD in my example for VADDSD or whatever you like. I have a lion's share of such a change implemented already and performance is greatly affected. If the community is interested in this change, I would be happy to prepare a patch.

Just looked up the numbers from Agner Fog for Sandy Bridge for vmovaps/etc for loading/storing from memory. vmovaps - load takes 1 load mu op, 3 latency, with a reciprocal throughput of 0.5. vmovaps - store takes 1 store mu op, 1 load mu op for address calculation, 3 latency, with a reciprocal throughput of 1. He does not list vmovsd, but movsd has the same stats as vmovaps, so I feel it is a

Some boring style comments: - whack trailing whitespace - spaces, not tabs - the methods in MCJITStreamer.cpp should probably have blank lines between them There seems to be an ownership problem of the MCJITObjectWriter. If I understand the code correctly, the assembler Finish method takes ownership of the Writer parameter, which presumably is needed to JIT two functions. +1 for separate

Jan Sjodin <jan_sjodin at yahoo.com> writes: > Better be quick! I am adding FMA4 and XOP now, and if you contribute > code before I do, you can spare yourself some XOP merging. Go ahead. We're not going to get there soon enough. :( -Dave

On 2 May 2012 16:57, Jan Sjodin <jan_sjodin at yahoo.com> wrote: > 8th will work for me. Can we pick a place that is not overly noisy? http://www.harbordhouse.ca should be fine, but let us know if you have another suggestion. > - Jan > Cheers, Rafael

Hey Jan and Dave, It's not obvious, but there is a significant scalar performance issue following the GCC intrinsics. Let's look at the VFMADDSD pattern. We're operating on scalars with undefineds as the remaining vector elements of the operands. This sounds okay, but when one looks closer... vmovsd fp4_+1088(%rip), %xmm3 # fpppp.f:647 vmovaps %xmm3, 18560(%rsp)

Because the intrinsics uses vector types (same as gcc). - Jan ----- Original Message ----- > From: "dag at cray.com" <dag at cray.com> > To: llvmdev at cs.uiuc.edu > Cc: > Sent: Wednesday, July 25, 2012 3:26 PM > Subject: [LLVMdev] X86 FMA4 > > We're migrating to LLVM 3.1 and trying to use the upstream FMA patterns. > > Why is VFMADDSD4

Hey Michael, Thanks for the legwork! It appears that the stats you listed are for movaps [SSE], not vmovaps [AVX]. I would *assume* that vmovaps(m128) is closer to vmovaps(m256), since they are both AVX instructions. Although, yes, I agree that this is not clear from Agner's report. Please correct me if I am misunderstanding. As I am sure you are aware, we cannot use SSE (movaps)

That is too bad. :(  You can always review the patches, and if you see something that can be done better let me know. - Jan ----- Original Message ----- > From: David A. Greene <greened at obbligato.org> > To: Jan Sjodin <jan_sjodin at yahoo.com> > Cc: David A. Greene <greened at obbligato.org>; Benjamin Kramer <benny.kra at googlemail.com>; "llvmdev at

What kind of restrictions will the existing object file formats impose on the JIT? I don't know enough about the JIT and object file format interaction to know if this will be an issue. It seems clear that it would be worse to try to encode "extra things" in some obscure way than to create the FOO format initially. If FOO is truly a superset of everything this could even be the

On Sat, Feb 9, 2019 at 4:44 PM Jan Sjodin <jan_sjodin at yahoo.com> wrote: > > The reason I'm looking for solutions that can work without "scanning the > > code" or "spooky action at a distance" is that we should have a solution > > that's easily digestible by folks who are not aware of GPU execution > models. > > > > The fallback