ARM64 Task Force - ghc-devs - Haskell.org

newer
Help with fusion rules and such

ARM64 Task Force

older
making ./validate run tests on all...

Luke Iannini

8 Jul 2014 8 Jul '14

2:06 a.m.

Howdy all, Would anyone like to team up on getting ARM64 support into GHC? Cheers Luke

Attachments:

attachment.html (text/html — 168 bytes)

Reply

Sign in to reply online Use email software

Show replies by date

Luke Iannini

7 Aug 7 Aug

7:58 p.m.

Hi all, An update on this -- I've made a bit of progress thanks to Karel and Colin's start at ARM64 support https://ghc.haskell.org/trac/ghc/ticket/7942 With a few tweaks*, that let me build a GHC that builds ARM64 binaries and load them onto my iPad Air, which is great! But of course they don't work yet since LLVM doesn't have the ARM64/GHC calling convention in. Happily I was able to use LLVM HEAD to do this, which means we don't need to be bound to Xcode's release schedules. I'm now studying David's patches to LLVM to learn how to add the ARM64/GHC calling convention to LLVM. *including Ben Gamari's patches to get LLVM HEAD working https://github.com/bgamari/ghc/tree/llvm-3.5-new Best Luke On Mon, Jul 7, 2014 at 11:06 PM, Luke Iannini wrote:

Howdy all,

Would anyone like to team up on getting ARM64 support into GHC?

Cheers Luke

Reply

Sign in to reply online Use email software

Karel Gardas

8 Aug 8 Aug

1:58 a.m.

On 08/ 8/14 01:58 AM, Luke Iannini wrote:

I'm now studying David's patches to LLVM to learn how to add the ARM64/GHC calling convention to LLVM.

Here is also original ARM/GHC calling convention submission. It's always good to have more examples as reference... http://lists.cs.uiuc.edu/pipermail/llvmdev/2011-October/044173.html Good luck with the ARM64/GHC porting work! Karel

Reply

Sign in to reply online Use email software

Luke Iannini

11:27 p.m.

Hi Karel, Thanks! A question: https://git.haskell.org/ghc.git/commitdiff/454b34cb3b67dec21f023339c4d53d734... adds references to s16, s17, s18, s19, d10 and d11 but I don't see those where I though to expect them in https://github.com/ghc/ghc/blob/master/includes/CodeGen.Platform.hs Am I down a wrong path? Luke On Thu, Aug 7, 2014 at 10:58 PM, Karel Gardas wrote:

On 08/ 8/14 01:58 AM, Luke Iannini wrote:

...
I'm now studying David's patches to LLVM to learn how to add the ARM64/GHC calling convention to LLVM.

Here is also original ARM/GHC calling convention submission. It's always good to have more examples as reference...

http://lists.cs.uiuc.edu/pipermail/llvmdev/2011-October/044173.html

Good luck with the ARM64/GHC porting work!

Karel

Reply

Sign in to reply online Use email software

Karel Gardas

9 Aug 9 Aug

7:22 a.m.

On 08/ 9/14 05:27 AM, Luke Iannini wrote:

Hi Karel, Thanks!

A question: https://git.haskell.org/ghc.git/commitdiff/454b34cb3b67dec21f023339c4d53d734... adds references to s16, s17, s18, s19, d10 and d11 but I don't see those

Yes, that adds FPU support for ARM.

where I though to expect them in https://github.com/ghc/ghc/blob/master/includes/CodeGen.Platform.hs

Hmm, whole ARM reg set is missing in this file. IIRC Simon Marlow were discussing this with Ben Gamari recently. I've not investigated if this is needed or not since I don't know if this is used only in NCG or in registerised build in general. If the former, ARM will not be there as there is no ARM NCG yet, if the later, then ARM should be there as ARM/LLVM/registerised build is a reality. Cheers, Karel

Reply

Sign in to reply online Use email software

Luke Iannini

10 Aug 10 Aug

9:44 p.m.

I think I've solved this particular mystery -- the registers were never defined there because that integer-representation of them is only used by the NCG. In LLVM land they were only ever stringified by the REG() macro. Except now globalRegMaybe is being used in CmmSink.hs (as Simon and Ben were discussing), and globalRegMaybe needs an integer value for each register to put into its Maybe RealReg return value. Since CmmSink.hs only checks 'isJust', it doesn't actually matter what the integer value is. So I've just gone ahead and defined them sequentially for now which seems to get me past this. Thanks! Luke On Sat, Aug 9, 2014 at 4:22 AM, Karel Gardas wrote:

On 08/ 9/14 05:27 AM, Luke Iannini wrote:

...
Hi Karel, Thanks!

A question: https://git.haskell.org/ghc.git/commitdiff/454b34cb3b67dec21f023339c4d53d 734af7605d adds references to s16, s17, s18, s19, d10 and d11 but I don't see those

Yes, that adds FPU support for ARM.

where I though to expect them in

...
https://github.com/ghc/ghc/blob/master/includes/CodeGen.Platform.hs

Hmm, whole ARM reg set is missing in this file. IIRC Simon Marlow were discussing this with Ben Gamari recently. I've not investigated if this is needed or not since I don't know if this is used only in NCG or in registerised build in general. If the former, ARM will not be there as there is no ARM NCG yet, if the later, then ARM should be there as ARM/LLVM/registerised build is a reality.

Cheers, Karel

Reply

Sign in to reply online Use email software

Luke Iannini

12 Aug 12 Aug

5:03 a.m.

I've pushed my WIP patches here: https://github.com/lukexi/llvm/commit/dfe74bb48eb05ce7847fa262f6e563d20d0b1f... https://github.com/lukexi/ghc/commit/e99b7a41e64f3ddb9bb420c0d5583f0e302e321... (they also require the latest libffi to be dropped in ftp://sourceware.org/pub/libffi/libffi-3.0.13.tar.gz due to https://ghc.haskell.org/trac/ghc/ticket/8664) These can produce an ARM64 GHC but the produced binaries aren't fully functional yet. They make it through hs_init() but crash rather opaquely when I try to call a simple fib function through the FFI. It looks like it's jumping somewhere strange; lldb tells me it's to 0x100e05110: .long 0x00000000 ; unknown opcode 0x100e05114: .long 0x00000000 ; unknown opcode 0x100e05118: .long 0x00000000 ; unknown opcode 0x100e0511c: .long 0x00000000 ; unknown opcode 0x100e05120: .long 0x00000000 ; unknown opcode 0x100e05124: .long 0x00000000 ; unknown opcode 0x100e05128: .long 0x00000000 ; unknown opcode 0x100e0512c: .long 0x00000000 ; unknown opcode If I put a breakpoint on StgRun and step by instruction, I seem to make it to about: https://github.com/lukexi/ghc/blob/e99b7a41e64f3ddb9bb420c0d5583f0e302e321e/... (give or take a line) before something goes mysteriously wrong and I'm no longer able to interact with the debugger. So I guess I'll try taking out float register support and see if that gets me anywhere. If anyone has some ideas on how to debug this I'd love to hear them! I've mostly assembled the patches by adapting the existing ARM support so it's quite possibly I'm doing something boneheaded. Cheers Luke On Sun, Aug 10, 2014 at 6:44 PM, Luke Iannini wrote:

I think I've solved this particular mystery -- the registers were never defined there because that integer-representation of them is only used by the NCG. In LLVM land they were only ever stringified by the REG() macro.

Except now globalRegMaybe is being used in CmmSink.hs (as Simon and Ben were discussing), and globalRegMaybe needs an integer value for each register to put into its Maybe RealReg return value. Since CmmSink.hs only checks 'isJust', it doesn't actually matter what the integer value is.

So I've just gone ahead and defined them sequentially for now which seems to get me past this.

Thanks! Luke

On Sat, Aug 9, 2014 at 4:22 AM, Karel Gardas wrote:

...
On 08/ 9/14 05:27 AM, Luke Iannini wrote:

...
Hi Karel, Thanks!

A question: https://git.haskell.org/ghc.git/commitdiff/ 454b34cb3b67dec21f023339c4d53d734af7605d adds references to s16, s17, s18, s19, d10 and d11 but I don't see those

Yes, that adds FPU support for ARM.

where I though to expect them in

...
https://github.com/ghc/ghc/blob/master/includes/CodeGen.Platform.hs

Hmm, whole ARM reg set is missing in this file. IIRC Simon Marlow were discussing this with Ben Gamari recently. I've not investigated if this is needed or not since I don't know if this is used only in NCG or in registerised build in general. If the former, ARM will not be there as there is no ARM NCG yet, if the later, then ARM should be there as ARM/LLVM/registerised build is a reality.

Cheers, Karel

Reply

Sign in to reply online Use email software

Karel Gardas

2:24 p.m.

On 08/12/14 11:03 AM, Luke Iannini wrote:

It looks like it's jumping somewhere strange; lldb tells me it's to 0x100e05110: .long 0x00000000 ; unknown opcode 0x100e05114: .long 0x00000000 ; unknown opcode 0x100e05118: .long 0x00000000 ; unknown opcode 0x100e0511c: .long 0x00000000 ; unknown opcode 0x100e05120: .long 0x00000000 ; unknown opcode 0x100e05124: .long 0x00000000 ; unknown opcode 0x100e05128: .long 0x00000000 ; unknown opcode 0x100e0512c: .long 0x00000000 ; unknown opcode

If I put a breakpoint on StgRun and step by instruction, I seem to make it to about: https://github.com/lukexi/ghc/blob/e99b7a41e64f3ddb9bb420c0d5583f0e302e321e/... (give or take a line)

strange that it's in the middle of the stp isns block. Anyway, this looks like a cpu exception doesn't it? You will need to find out the reg which holds the "exception reason" value and then look on it in your debugger to find out what's going wrong there. Karel

Reply

Sign in to reply online Use email software

Luke Iannini

7:47 p.m.

Hi all, Yahoo, happy news -- I think I've got it. Studying enough of the non-handwritten ASM that I was stepping through led me to make this change: https://github.com/lukexi/ghc/commit/1140e11db07817fcfc12446c74cd5a2bcdf9278... (I think disabling the floating point registers was just a red herring; I'll confirm that next) And I can now call this fib code happily via the FFI: fibs :: [Int] fibs = 1:1:zipWith (+) fibs (tail fibs) foreign export ccall fib :: Int -> Int fib :: Int -> Int fib = (fibs !!) For posterity, more detail on the crashing case earlier (this is fixed now with proper storage and updating of the frame pointer): Calling fib(1) or fib(2) worked, but calling fib(3) triggered the crash. This was the backtrace, where you can see the errant 0x0000000100f05110 frame values. (lldb) bt * thread #1: tid = 0xac6ed, 0x0000000100f05110, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=2, address=0x100f05110) frame #0: 0x0000000100f05110 frame #1: 0x0000000100f05110 * frame #2: 0x00000001000ffc9c HelloHaskell`-[SPJViewController viewDidLoad](self=0x0000000144e0cf10, _cmd=0x0000000186ae429a) + 76 at SPJViewController.m:22 frame #3: 0x00000001862f8b70 UIKit`-[UIViewController loadViewIfRequired] + 692 frame #4: 0x00000001862f8880 UIKit`-[UIViewController view] + 32 frame #5: 0x00000001862feeb0 UIKit`-[UIWindow addRootViewControllerViewIfPossible] + 72 frame #6: 0x00000001862fc6d4 UIKit`-[UIWindow _setHidden:forced:] + 296 frame #7: 0x000000018636d2bc UIKit`-[UIWindow makeKeyAndVisible] + 56 frame #8: 0x000000018657ff74 UIKit`-[UIApplication _callInitializationDelegatesForMainScene:transitionContext:] + 2804 frame #9: 0x00000001865824ec UIKit`-[UIApplication _runWithMainScene:transitionContext:completion:] + 1480 frame #10: 0x0000000186580b84 UIKit`-[UIApplication workspaceDidEndTransaction:] + 184 frame #11: 0x0000000189d846ac FrontBoardServices`__31-[FBSSerialQueue performAsync:]_block_invoke + 28 frame #12: 0x0000000181c7a360 CoreFoundation`__CFRUNLOOP_IS_CALLING_OUT_TO_A_BLOCK__ + 20 frame #13: 0x0000000181c79468 CoreFoundation`__CFRunLoopDoBlocks + 312 frame #14: 0x0000000181c77a8c CoreFoundation`__CFRunLoopRun + 1756 frame #15: 0x0000000181ba5664 CoreFoundation`CFRunLoopRunSpecific + 396 frame #16: 0x0000000186363140 UIKit`-[UIApplication _run] + 552 frame #17: 0x000000018635e164 UIKit`UIApplicationMain + 1488 frame #18: 0x0000000100100268 HelloHaskell`main(argc=1, argv=0x000000016fd07a58) + 204 at main.m:24 frame #19: 0x00000001921eea08 libdyld.dylib`start + 4 On Tue, Aug 12, 2014 at 11:24 AM, Karel Gardas wrote:

On 08/12/14 11:03 AM, Luke Iannini wrote:

...
It looks like it's jumping somewhere strange; lldb tells me it's to 0x100e05110: .long 0x00000000 ; unknown opcode 0x100e05114: .long 0x00000000 ; unknown opcode 0x100e05118: .long 0x00000000 ; unknown opcode 0x100e0511c: .long 0x00000000 ; unknown opcode 0x100e05120: .long 0x00000000 ; unknown opcode 0x100e05124: .long 0x00000000 ; unknown opcode 0x100e05128: .long 0x00000000 ; unknown opcode 0x100e0512c: .long 0x00000000 ; unknown opcode

If I put a breakpoint on StgRun and step by instruction, I seem to make it to about: https://github.com/lukexi/ghc/blob/e99b7a41e64f3ddb9bb420c0d5583f 0e302e321e/rts/StgCRun.c#L770 (give or take a line)

strange that it's in the middle of the stp isns block. Anyway, this looks like a cpu exception doesn't it? You will need to find out the reg which holds the "exception reason" value and then look on it in your debugger to find out what's going wrong there.

Karel

Reply

Sign in to reply online Use email software

Luke Iannini

13 Aug 13 Aug

8:26 p.m.

Indeed, the float register stuff was a red herring -- restoring it caused no problems and all my tests are working great. So yahoo!! We've got ARM64 support. I'll tidy up the patches and create a ticket for review and merge. Luke On Tue, Aug 12, 2014 at 4:47 PM, Luke Iannini wrote:

Hi all, Yahoo, happy news -- I think I've got it. Studying enough of the non-handwritten ASM that I was stepping through led me to make this change:

https://github.com/lukexi/ghc/commit/1140e11db07817fcfc12446c74cd5a2bcdf9278... (I think disabling the floating point registers was just a red herring; I'll confirm that next)

And I can now call this fib code happily via the FFI: fibs :: [Int] fibs = 1:1:zipWith (+) fibs (tail fibs)

foreign export ccall fib :: Int -> Int fib :: Int -> Int fib = (fibs !!)

For posterity, more detail on the crashing case earlier (this is fixed now with proper storage and updating of the frame pointer): Calling fib(1) or fib(2) worked, but calling fib(3) triggered the crash. This was the backtrace, where you can see the errant 0x0000000100f05110 frame values. (lldb) bt * thread #1: tid = 0xac6ed, 0x0000000100f05110, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=2, address=0x100f05110) frame #0: 0x0000000100f05110 frame #1: 0x0000000100f05110 * frame #2: 0x00000001000ffc9c HelloHaskell`-[SPJViewController viewDidLoad](self=0x0000000144e0cf10, _cmd=0x0000000186ae429a) + 76 at SPJViewController.m:22 frame #3: 0x00000001862f8b70 UIKit`-[UIViewController loadViewIfRequired] + 692 frame #4: 0x00000001862f8880 UIKit`-[UIViewController view] + 32 frame #5: 0x00000001862feeb0 UIKit`-[UIWindow addRootViewControllerViewIfPossible] + 72 frame #6: 0x00000001862fc6d4 UIKit`-[UIWindow _setHidden:forced:] + 296 frame #7: 0x000000018636d2bc UIKit`-[UIWindow makeKeyAndVisible] + 56 frame #8: 0x000000018657ff74 UIKit`-[UIApplication _callInitializationDelegatesForMainScene:transitionContext:] + 2804 frame #9: 0x00000001865824ec UIKit`-[UIApplication _runWithMainScene:transitionContext:completion:] + 1480 frame #10: 0x0000000186580b84 UIKit`-[UIApplication workspaceDidEndTransaction:] + 184 frame #11: 0x0000000189d846ac FrontBoardServices`__31-[FBSSerialQueue performAsync:]_block_invoke + 28 frame #12: 0x0000000181c7a360 CoreFoundation`__CFRUNLOOP_IS_CALLING_OUT_TO_A_BLOCK__ + 20 frame #13: 0x0000000181c79468 CoreFoundation`__CFRunLoopDoBlocks + 312 frame #14: 0x0000000181c77a8c CoreFoundation`__CFRunLoopRun + 1756 frame #15: 0x0000000181ba5664 CoreFoundation`CFRunLoopRunSpecific + 396 frame #16: 0x0000000186363140 UIKit`-[UIApplication _run] + 552 frame #17: 0x000000018635e164 UIKit`UIApplicationMain + 1488 frame #18: 0x0000000100100268 HelloHaskell`main(argc=1, argv=0x000000016fd07a58) + 204 at main.m:24 frame #19: 0x00000001921eea08 libdyld.dylib`start + 4

On Tue, Aug 12, 2014 at 11:24 AM, Karel Gardas wrote:

...
On 08/12/14 11:03 AM, Luke Iannini wrote:

...
It looks like it's jumping somewhere strange; lldb tells me it's to 0x100e05110: .long 0x00000000 ; unknown opcode 0x100e05114: .long 0x00000000 ; unknown opcode 0x100e05118: .long 0x00000000 ; unknown opcode 0x100e0511c: .long 0x00000000 ; unknown opcode 0x100e05120: .long 0x00000000 ; unknown opcode 0x100e05124: .long 0x00000000 ; unknown opcode 0x100e05128: .long 0x00000000 ; unknown opcode 0x100e0512c: .long 0x00000000 ; unknown opcode

If I put a breakpoint on StgRun and step by instruction, I seem to make it to about: https://github.com/lukexi/ghc/blob/e99b7a41e64f3ddb9bb420c0d5583f 0e302e321e/rts/StgCRun.c#L770 (give or take a line)

strange that it's in the middle of the stp isns block. Anyway, this looks like a cpu exception doesn't it? You will need to find out the reg which holds the "exception reason" value and then look on it in your debugger to find out what's going wrong there.

Karel

Reply

Sign in to reply online Use email software

Ben Gamari

11:20 p.m.

Luke Iannini writes:

Indeed, the float register stuff was a red herring -- restoring it caused no problems and all my tests are working great. So yahoo!! We've got ARM64 support.

Yay! Awesome work! Cheers, - Ben

Reply

Sign in to reply online Use email software

Manuel M T Chakravarty

11:22 p.m.

That’s awesome — great work! Manuel Luke Iannini :

Indeed, the float register stuff was a red herring — restoring it caused no problems and all my tests are working great. So yahoo!! We've got ARM64 support.

I'll tidy up the patches and create a ticket for review and merge.

Luke

On Tue, Aug 12, 2014 at 4:47 PM, Luke Iannini wrote: Hi all, Yahoo, happy news — I think I've got it. Studying enough of the non-handwritten ASM that I was stepping through led me to make this change: https://github.com/lukexi/ghc/commit/1140e11db07817fcfc12446c74cd5a2bcdf9278... (I think disabling the floating point registers was just a red herring; I'll confirm that next)

And I can now call this fib code happily via the FFI: fibs :: [Int] fibs = 1:1:zipWith (+) fibs (tail fibs)

foreign export ccall fib :: Int -> Int fib :: Int -> Int fib = (fibs !!)

For posterity, more detail on the crashing case earlier (this is fixed now with proper storage and updating of the frame pointer): Calling fib(1) or fib(2) worked, but calling fib(3) triggered the crash. This was the backtrace, where you can see the errant 0x0000000100f05110 frame values. (lldb) bt * thread #1: tid = 0xac6ed, 0x0000000100f05110, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=2, address=0x100f05110) frame #0: 0x0000000100f05110 frame #1: 0x0000000100f05110 * frame #2: 0x00000001000ffc9c HelloHaskell`-[SPJViewController viewDidLoad](self=0x0000000144e0cf10, _cmd=0x0000000186ae429a) + 76 at SPJViewController.m:22 frame #3: 0x00000001862f8b70 UIKit`-[UIViewController loadViewIfRequired] + 692 frame #4: 0x00000001862f8880 UIKit`-[UIViewController view] + 32 frame #5: 0x00000001862feeb0 UIKit`-[UIWindow addRootViewControllerViewIfPossible] + 72 frame #6: 0x00000001862fc6d4 UIKit`-[UIWindow _setHidden:forced:] + 296 frame #7: 0x000000018636d2bc UIKit`-[UIWindow makeKeyAndVisible] + 56 frame #8: 0x000000018657ff74 UIKit`-[UIApplication _callInitializationDelegatesForMainScene:transitionContext:] + 2804 frame #9: 0x00000001865824ec UIKit`-[UIApplication _runWithMainScene:transitionContext:completion:] + 1480 frame #10: 0x0000000186580b84 UIKit`-[UIApplication workspaceDidEndTransaction:] + 184 frame #11: 0x0000000189d846ac FrontBoardServices`__31-[FBSSerialQueue performAsync:]_block_invoke + 28 frame #12: 0x0000000181c7a360 CoreFoundation`__CFRUNLOOP_IS_CALLING_OUT_TO_A_BLOCK__ + 20 frame #13: 0x0000000181c79468 CoreFoundation`__CFRunLoopDoBlocks + 312 frame #14: 0x0000000181c77a8c CoreFoundation`__CFRunLoopRun + 1756 frame #15: 0x0000000181ba5664 CoreFoundation`CFRunLoopRunSpecific + 396 frame #16: 0x0000000186363140 UIKit`-[UIApplication _run] + 552 frame #17: 0x000000018635e164 UIKit`UIApplicationMain + 1488 frame #18: 0x0000000100100268 HelloHaskell`main(argc=1, argv=0x000000016fd07a58) + 204 at main.m:24 frame #19: 0x00000001921eea08 libdyld.dylib`start + 4

On Tue, Aug 12, 2014 at 11:24 AM, Karel Gardas wrote: On 08/12/14 11:03 AM, Luke Iannini wrote: It looks like it's jumping somewhere strange; lldb tells me it's to 0x100e05110: .long 0x00000000 ; unknown opcode 0x100e05114: .long 0x00000000 ; unknown opcode 0x100e05118: .long 0x00000000 ; unknown opcode 0x100e0511c: .long 0x00000000 ; unknown opcode 0x100e05120: .long 0x00000000 ; unknown opcode 0x100e05124: .long 0x00000000 ; unknown opcode 0x100e05128: .long 0x00000000 ; unknown opcode 0x100e0512c: .long 0x00000000 ; unknown opcode

If I put a breakpoint on StgRun and step by instruction, I seem to make it to about: https://github.com/lukexi/ghc/blob/e99b7a41e64f3ddb9bb420c0d5583f0e302e321e/... (give or take a line)

strange that it's in the middle of the stp isns block. Anyway, this looks like a cpu exception doesn't it? You will need to find out the reg which holds the "exception reason" value and then look on it in your debugger to find out what's going wrong there.

Karel

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs

Reply

Sign in to reply online Use email software

3932

Age (days ago)

3969

Last active (days ago)

Download

11 comments

4 participants

tags

participants (4)

Ben Gamari
Karel Gardas
Luke Iannini
Manuel M T Chakravarty