diff mbox series

Add optional _Float16 support

Message ID 20210701210537.51272-1-hjl.tools@gmail.com
State Not applicable
Headers
Series Add optional _Float16 support |

Checks

Context Check Description
dj/TryBot-apply_patch fail Patch failed to apply to master at the time it was sent
dj/TryBot-32bit fail Patch series failed to apply

Commit Message

H.J. Lu July 1, 2021, 9:05 p.m. UTC 
  1. Pass _Float16 and _Complex _Float16 values on stack.
2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
---
 low-level-sys-info.tex | 57 +++++++++++++++++++++++++++++-------------
 1 file changed, 40 insertions(+), 17 deletions(-)

Comments

Joseph Myers July 1, 2021, 10:10 p.m. UTC | #1 
On Thu, 1 Jul 2021, H.J. Lu via Gcc-patches wrote:

> 2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.

That restricts use of _Float16 to processors with SSE.  Is that what we 
want in the ABI, or should _Float16 be available with base 32-bit x86 
architecture features only, much like _Float128 and the decimal FP types 
are?  (If it is restricted to SSE, we can of course ensure relevant libgcc 
functions are built with SSE enabled, and likewise in glibc if that gains 
_Float16 functions, though maybe with some extra complications to get 
relevant testcases to run whenever possible.)
H.J. Lu July 1, 2021, 10:27 p.m. UTC | #2 
On Thu, Jul 1, 2021 at 3:10 PM Joseph Myers <joseph@codesourcery.com> wrote:
>
> On Thu, 1 Jul 2021, H.J. Lu via Gcc-patches wrote:
>
> > 2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
>
> That restricts use of _Float16 to processors with SSE.  Is that what we
> want in the ABI, or should _Float16 be available with base 32-bit x86
> architecture features only, much like _Float128 and the decimal FP types

Yes, _Float16 requires XMM registers.

> are?  (If it is restricted to SSE, we can of course ensure relevant libgcc
> functions are built with SSE enabled, and likewise in glibc if that gains
> _Float16 functions, though maybe with some extra complications to get
> relevant testcases to run whenever possible.)
>

_Float16 functions in libgcc should be compiled with SSE enabled.

BTW, _Float16 software emulation may require more than just SSE
since we need to do _Float16 load and store with XMM registers.
There is no 16bit load/store for XMM registers without AVX512FP16.
Joseph Myers July 1, 2021, 10:40 p.m. UTC | #3 
On Thu, 1 Jul 2021, H.J. Lu wrote:

> BTW, _Float16 software emulation may require more than just SSE
> since we need to do _Float16 load and store with XMM registers.
> There is no 16bit load/store for XMM registers without AVX512FP16.

You should be able to make the move go via general-purpose registers (for 
example) if you can't do a direct 16-bit load/store for XMM registers.
H.J. Lu July 1, 2021, 11:01 p.m. UTC | #4 
On Thu, Jul 1, 2021 at 3:40 PM Joseph Myers <joseph@codesourcery.com> wrote:
>
> On Thu, 1 Jul 2021, H.J. Lu wrote:
>
> > BTW, _Float16 software emulation may require more than just SSE
> > since we need to do _Float16 load and store with XMM registers.
> > There is no 16bit load/store for XMM registers without AVX512FP16.
>
> You should be able to make the move go via general-purpose registers (for
> example) if you can't do a direct 16-bit load/store for XMM registers.
>

There is no 16bit move between GPRs and XMM registers without
AVX512FP16.
Craig Topper July 1, 2021, 11:05 p.m. UTC | #5 
On Thu, Jul 1, 2021 at 4:02 PM H.J. Lu via llvm-dev <llvm-dev@lists.llvm.org>
wrote:

> On Thu, Jul 1, 2021 at 3:40 PM Joseph Myers <joseph@codesourcery.com>
> wrote:
> >
> > On Thu, 1 Jul 2021, H.J. Lu wrote:
> >
> > > BTW, _Float16 software emulation may require more than just SSE
> > > since we need to do _Float16 load and store with XMM registers.
> > > There is no 16bit load/store for XMM registers without AVX512FP16.
> >
> > You should be able to make the move go via general-purpose registers (for
> > example) if you can't do a direct 16-bit load/store for XMM registers.
> >
>
> There is no 16bit move between GPRs and XMM registers without
> AVX512FP16.
>
>
Isn't PINSRW supported since SSE1?


>
> --
> H.J.
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev@lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
Jacob Lifshay July 1, 2021, 11:33 p.m. UTC | #6 
On Thu, Jul 1, 2021, 15:28 H.J. Lu via llvm-dev <llvm-dev@lists.llvm.org>
wrote:

> On Thu, Jul 1, 2021 at 3:10 PM Joseph Myers <joseph@codesourcery.com>
> wrote:
> >
> > On Thu, 1 Jul 2021, H.J. Lu via Gcc-patches wrote:
> >
> > > 2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1
> registers.
> >
> > That restricts use of _Float16 to processors with SSE.  Is that what we
> > want in the ABI, or should _Float16 be available with base 32-bit x86
> > architecture features only, much like _Float128 and the decimal FP types
>
> Yes, _Float16 requires XMM registers.
>
> > are?  (If it is restricted to SSE, we can of course ensure relevant
> libgcc
> > functions are built with SSE enabled, and likewise in glibc if that gains
> > _Float16 functions, though maybe with some extra complications to get
> > relevant testcases to run whenever possible.)
> >
>
> _Float16 functions in libgcc should be compiled with SSE enabled.
>
> BTW, _Float16 software emulation may require more than just SSE
> since we need to do _Float16 load and store with XMM registers.
> There is no 16bit load/store for XMM registers without AVX512FP16.
>

Umm, if you just need to load/store 16-bit scalars in XMM registers you can
use pextrw and pinsrw which don't require AVX. f16x8 can use any of the
standard full-register load/stores.

https://gcc.godbolt.org/z/ncznr9TM1

Jacob
Richard Biener July 2, 2021, 7:45 a.m. UTC | #7 
On Fri, Jul 2, 2021 at 1:34 AM Jacob Lifshay via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Thu, Jul 1, 2021, 15:28 H.J. Lu via llvm-dev <llvm-dev@lists.llvm.org>
> wrote:
>
> > On Thu, Jul 1, 2021 at 3:10 PM Joseph Myers <joseph@codesourcery.com>
> > wrote:
> > >
> > > On Thu, 1 Jul 2021, H.J. Lu via Gcc-patches wrote:
> > >
> > > > 2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1
> > registers.
> > >
> > > That restricts use of _Float16 to processors with SSE.  Is that what we
> > > want in the ABI, or should _Float16 be available with base 32-bit x86
> > > architecture features only, much like _Float128 and the decimal FP types
> >
> > Yes, _Float16 requires XMM registers.
> >
> > > are?  (If it is restricted to SSE, we can of course ensure relevant
> > libgcc
> > > functions are built with SSE enabled, and likewise in glibc if that gains
> > > _Float16 functions, though maybe with some extra complications to get
> > > relevant testcases to run whenever possible.)
> > >
> >
> > _Float16 functions in libgcc should be compiled with SSE enabled.
> >
> > BTW, _Float16 software emulation may require more than just SSE
> > since we need to do _Float16 load and store with XMM registers.
> > There is no 16bit load/store for XMM registers without AVX512FP16.
> >
>
> Umm, if you just need to load/store 16-bit scalars in XMM registers you can
> use pextrw and pinsrw which don't require AVX. f16x8 can use any of the
> standard full-register load/stores.

It looks like that requires SSE2, with SSE only inserts/extracts
to/from MMX regs
are supported.  But of course GPR half-word loads and GPR->XMM moves of
full size would work.

> https://gcc.godbolt.org/z/ncznr9TM1
>
> Jacob
Hongtao Liu July 2, 2021, 8:03 a.m. UTC | #8 
On Fri, Jul 2, 2021 at 3:46 PM Richard Biener via llvm-dev
<llvm-dev@lists.llvm.org> wrote:
>
> On Fri, Jul 2, 2021 at 1:34 AM Jacob Lifshay via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > On Thu, Jul 1, 2021, 15:28 H.J. Lu via llvm-dev <llvm-dev@lists.llvm.org>
> > wrote:
> >
> > > On Thu, Jul 1, 2021 at 3:10 PM Joseph Myers <joseph@codesourcery.com>
> > > wrote:
> > > >
> > > > On Thu, 1 Jul 2021, H.J. Lu via Gcc-patches wrote:
> > > >
> > > > > 2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1
> > > registers.
> > > >
> > > > That restricts use of _Float16 to processors with SSE.  Is that what we
> > > > want in the ABI, or should _Float16 be available with base 32-bit x86
> > > > architecture features only, much like _Float128 and the decimal FP types
> > >
> > > Yes, _Float16 requires XMM registers.
> > >
> > > > are?  (If it is restricted to SSE, we can of course ensure relevant
> > > libgcc
> > > > functions are built with SSE enabled, and likewise in glibc if that gains
> > > > _Float16 functions, though maybe with some extra complications to get
> > > > relevant testcases to run whenever possible.)
> > > >
> > >
> > > _Float16 functions in libgcc should be compiled with SSE enabled.
> > >
> > > BTW, _Float16 software emulation may require more than just SSE
> > > since we need to do _Float16 load and store with XMM registers.
> > > There is no 16bit load/store for XMM registers without AVX512FP16.
> > >
> >
> > Umm, if you just need to load/store 16-bit scalars in XMM registers you can
> > use pextrw and pinsrw which don't require AVX. f16x8 can use any of the
> > standard full-register load/stores.
>
> It looks like that requires SSE2, with SSE only inserts/extracts
> to/from MMX regs
> are supported.  But of course GPR half-word loads and GPR->XMM moves of
> full size would work.
movd between sse registers and gpr also required sse2.
>
> > https://gcc.godbolt.org/z/ncznr9TM1
> >
> > Jacob
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev@lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
Jakub Jelinek July 2, 2021, 9:21 a.m. UTC | #9 
On Fri, Jul 02, 2021 at 09:45:46AM +0200, Richard Biener via Gcc-patches wrote:
> > > > are?  (If it is restricted to SSE, we can of course ensure relevant
> > > libgcc
> > > > functions are built with SSE enabled, and likewise in glibc if that gains
> > > > _Float16 functions, though maybe with some extra complications to get
> > > > relevant testcases to run whenever possible.)
> > > >
> > >
> > > _Float16 functions in libgcc should be compiled with SSE enabled.
> > >
> > > BTW, _Float16 software emulation may require more than just SSE
> > > since we need to do _Float16 load and store with XMM registers.
> > > There is no 16bit load/store for XMM registers without AVX512FP16.
> > >
> >
> > Umm, if you just need to load/store 16-bit scalars in XMM registers you can
> > use pextrw and pinsrw which don't require AVX. f16x8 can use any of the
> > standard full-register load/stores.
> 
> It looks like that requires SSE2, with SSE only inserts/extracts
> to/from MMX regs
> are supported.  But of course GPR half-word loads and GPR->XMM moves of
> full size would work.

Loads can be done in SSE2 directly with PINSRW, that supports 16-bit load
from memory to XMM reg.  But SSE2 PEXTRW only supports stores into GPR
and one needs SSE4.1 fo PEXTRW into memory.  So, for the stores and SSE2 one
needs secondary reload...

	Jakub
develop--- via Libc-alpha July 13, 2021, 3:59 a.m. UTC | #10 
> Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.

Can you please explain the behavior here? Is there difference between _Float16 and _Complex _Float16 when return? I.e.,
1, In which case will _Float16 values return in both %xmm0 and %xmm1?
2, For a single _Float16 value, are both real part and imaginary part returned in %xmm0? Or returned in %xmm0 and %xmm1 respectively?

Thanks
Pengfei

-----Original Message-----
From: llvm-dev <llvm-dev-bounces@lists.llvm.org> On Behalf Of H.J. Lu via llvm-dev
Sent: Friday, July 2, 2021 6:28 AM
To: Joseph Myers <joseph@codesourcery.com>
Cc: llvm-dev@lists.llvm.org; GCC Patches <gcc-patches@gcc.gnu.org>; GNU C Library <libc-alpha@sourceware.org>; IA32 System V Application Binary Interface <ia32-abi@googlegroups.com>
Subject: Re: [llvm-dev] [PATCH] Add optional _Float16 support

On Thu, Jul 1, 2021 at 3:10 PM Joseph Myers <joseph@codesourcery.com> wrote:
>
> On Thu, 1 Jul 2021, H.J. Lu via Gcc-patches wrote:
>
> > 2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
>
> That restricts use of _Float16 to processors with SSE.  Is that what 
> we want in the ABI, or should _Float16 be available with base 32-bit 
> x86 architecture features only, much like _Float128 and the decimal FP 
> types

Yes, _Float16 requires XMM registers.

> are?  (If it is restricted to SSE, we can of course ensure relevant 
> libgcc functions are built with SSE enabled, and likewise in glibc if 
> that gains
> _Float16 functions, though maybe with some extra complications to get 
> relevant testcases to run whenever possible.)
>

_Float16 functions in libgcc should be compiled with SSE enabled.

BTW, _Float16 software emulation may require more than just SSE since we need to do _Float16 load and store with XMM registers.
There is no 16bit load/store for XMM registers without AVX512FP16.

--
H.J.
H.J. Lu July 13, 2021, 2:26 p.m. UTC | #11 
On Mon, Jul 12, 2021 at 8:59 PM Wang, Pengfei <pengfei.wang@intel.com> wrote:
>
> > Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
>
> Can you please explain the behavior here? Is there difference between _Float16 and _Complex _Float16 when return? I.e.,
> 1, In which case will _Float16 values return in both %xmm0 and %xmm1?
> 2, For a single _Float16 value, are both real part and imaginary part returned in %xmm0? Or returned in %xmm0 and %xmm1 respectively?

Here is the v2 patch to add the missing _Float16 bits.   The PDF file is at

https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI

> Thanks
> Pengfei
>
> -----Original Message-----
> From: llvm-dev <llvm-dev-bounces@lists.llvm.org> On Behalf Of H.J. Lu via llvm-dev
> Sent: Friday, July 2, 2021 6:28 AM
> To: Joseph Myers <joseph@codesourcery.com>
> Cc: llvm-dev@lists.llvm.org; GCC Patches <gcc-patches@gcc.gnu.org>; GNU C Library <libc-alpha@sourceware.org>; IA32 System V Application Binary Interface <ia32-abi@googlegroups.com>
> Subject: Re: [llvm-dev] [PATCH] Add optional _Float16 support
>
> On Thu, Jul 1, 2021 at 3:10 PM Joseph Myers <joseph@codesourcery.com> wrote:
> >
> > On Thu, 1 Jul 2021, H.J. Lu via Gcc-patches wrote:
> >
> > > 2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
> >
> > That restricts use of _Float16 to processors with SSE.  Is that what
> > we want in the ABI, or should _Float16 be available with base 32-bit
> > x86 architecture features only, much like _Float128 and the decimal FP
> > types
>
> Yes, _Float16 requires XMM registers.
>
> > are?  (If it is restricted to SSE, we can of course ensure relevant
> > libgcc functions are built with SSE enabled, and likewise in glibc if
> > that gains
> > _Float16 functions, though maybe with some extra complications to get
> > relevant testcases to run whenever possible.)
> >
>
> _Float16 functions in libgcc should be compiled with SSE enabled.
>
> BTW, _Float16 software emulation may require more than just SSE since we need to do _Float16 load and store with XMM registers.
> There is no 16bit load/store for XMM registers without AVX512FP16.
>
> --
> H.J.
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev@lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
develop--- via Libc-alpha July 13, 2021, 2:48 p.m. UTC | #12 
Hi H.J.,

Our LLVM implementation currently use %xmm0 for both _Complex's real part and imaginary part. Do we have special reason to use two registers?
We are using one register on X64. Considering the performance, especially the register pressure, should it be better to use one register for _Complex _Float16 on 32 bits target?

Thanks
Pengfei

-----Original Message-----
From: H.J. Lu <hjl.tools@gmail.com> 
Sent: Tuesday, July 13, 2021 10:26 PM
To: Wang, Pengfei <pengfei.wang@intel.com>; llvm-dev@lists.llvm.org
Cc: Joseph Myers <joseph@codesourcery.com>; GCC Patches <gcc-patches@gcc.gnu.org>; GNU C Library <libc-alpha@sourceware.org>; IA32 System V Application Binary Interface <ia32-abi@googlegroups.com>
Subject: Re: [llvm-dev] [PATCH] Add optional _Float16 support

On Mon, Jul 12, 2021 at 8:59 PM Wang, Pengfei <pengfei.wang@intel.com> wrote:
>
> > Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
>
> Can you please explain the behavior here? Is there difference between 
> _Float16 and _Complex _Float16 when return? I.e., 1, In which case will _Float16 values return in both %xmm0 and %xmm1?
> 2, For a single _Float16 value, are both real part and imaginary part returned in %xmm0? Or returned in %xmm0 and %xmm1 respectively?

Here is the v2 patch to add the missing _Float16 bits.   The PDF file is at

https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI

> Thanks
> Pengfei
>
> -----Original Message-----
> From: llvm-dev <llvm-dev-bounces@lists.llvm.org> On Behalf Of H.J. Lu 
> via llvm-dev
> Sent: Friday, July 2, 2021 6:28 AM
> To: Joseph Myers <joseph@codesourcery.com>
> Cc: llvm-dev@lists.llvm.org; GCC Patches <gcc-patches@gcc.gnu.org>; 
> GNU C Library <libc-alpha@sourceware.org>; IA32 System V Application 
> Binary Interface <ia32-abi@googlegroups.com>
> Subject: Re: [llvm-dev] [PATCH] Add optional _Float16 support
>
> On Thu, Jul 1, 2021 at 3:10 PM Joseph Myers <joseph@codesourcery.com> wrote:
> >
> > On Thu, 1 Jul 2021, H.J. Lu via Gcc-patches wrote:
> >
> > > 2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
> >
> > That restricts use of _Float16 to processors with SSE.  Is that what 
> > we want in the ABI, or should _Float16 be available with base 32-bit
> > x86 architecture features only, much like _Float128 and the decimal 
> > FP types
>
> Yes, _Float16 requires XMM registers.
>
> > are?  (If it is restricted to SSE, we can of course ensure relevant 
> > libgcc functions are built with SSE enabled, and likewise in glibc 
> > if that gains
> > _Float16 functions, though maybe with some extra complications to 
> > get relevant testcases to run whenever possible.)
> >
>
> _Float16 functions in libgcc should be compiled with SSE enabled.
>
> BTW, _Float16 software emulation may require more than just SSE since we need to do _Float16 load and store with XMM registers.
> There is no 16bit load/store for XMM registers without AVX512FP16.
>
> --
> H.J.
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev@lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev



--
H.J.
H.J. Lu July 13, 2021, 3:04 p.m. UTC | #13 
On Tue, Jul 13, 2021 at 7:48 AM Wang, Pengfei <pengfei.wang@intel.com> wrote:
>
> Hi H.J.,
>
> Our LLVM implementation currently use %xmm0 for both _Complex's real part and imaginary part. Do we have special reason to use two registers?
> We are using one register on X64. Considering the performance, especially the register pressure, should it be better to use one register for _Complex _Float16 on 32 bits target?

x86-64 psABI is unrelated to i386 psABI.  Using a pair of registers is
more natural for
complex _Float16.  Since it is only used for function return value, I
don't think there is
a register pressure issue.

> Thanks
> Pengfei
>
> -----Original Message-----
> From: H.J. Lu <hjl.tools@gmail.com>
> Sent: Tuesday, July 13, 2021 10:26 PM
> To: Wang, Pengfei <pengfei.wang@intel.com>; llvm-dev@lists.llvm.org
> Cc: Joseph Myers <joseph@codesourcery.com>; GCC Patches <gcc-patches@gcc.gnu.org>; GNU C Library <libc-alpha@sourceware.org>; IA32 System V Application Binary Interface <ia32-abi@googlegroups.com>
> Subject: Re: [llvm-dev] [PATCH] Add optional _Float16 support
>
> On Mon, Jul 12, 2021 at 8:59 PM Wang, Pengfei <pengfei.wang@intel.com> wrote:
> >
> > > Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
> >
> > Can you please explain the behavior here? Is there difference between
> > _Float16 and _Complex _Float16 when return? I.e., 1, In which case will _Float16 values return in both %xmm0 and %xmm1?
> > 2, For a single _Float16 value, are both real part and imaginary part returned in %xmm0? Or returned in %xmm0 and %xmm1 respectively?
>
> Here is the v2 patch to add the missing _Float16 bits.   The PDF file is at
>
> https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI
>
> > Thanks
> > Pengfei
> >
> > -----Original Message-----
> > From: llvm-dev <llvm-dev-bounces@lists.llvm.org> On Behalf Of H.J. Lu
> > via llvm-dev
> > Sent: Friday, July 2, 2021 6:28 AM
> > To: Joseph Myers <joseph@codesourcery.com>
> > Cc: llvm-dev@lists.llvm.org; GCC Patches <gcc-patches@gcc.gnu.org>;
> > GNU C Library <libc-alpha@sourceware.org>; IA32 System V Application
> > Binary Interface <ia32-abi@googlegroups.com>
> > Subject: Re: [llvm-dev] [PATCH] Add optional _Float16 support
> >
> > On Thu, Jul 1, 2021 at 3:10 PM Joseph Myers <joseph@codesourcery.com> wrote:
> > >
> > > On Thu, 1 Jul 2021, H.J. Lu via Gcc-patches wrote:
> > >
> > > > 2. Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
> > >
> > > That restricts use of _Float16 to processors with SSE.  Is that what
> > > we want in the ABI, or should _Float16 be available with base 32-bit
> > > x86 architecture features only, much like _Float128 and the decimal
> > > FP types
> >
> > Yes, _Float16 requires XMM registers.
> >
> > > are?  (If it is restricted to SSE, we can of course ensure relevant
> > > libgcc functions are built with SSE enabled, and likewise in glibc
> > > if that gains
> > > _Float16 functions, though maybe with some extra complications to
> > > get relevant testcases to run whenever possible.)
> > >
> >
> > _Float16 functions in libgcc should be compiled with SSE enabled.
> >
> > BTW, _Float16 software emulation may require more than just SSE since we need to do _Float16 load and store with XMM registers.
> > There is no 16bit load/store for XMM registers without AVX512FP16.
> >
> > --
> > H.J.
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev@lists.llvm.org
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>
>
> --
> H.J.
Joseph Myers July 13, 2021, 3:41 p.m. UTC | #14 
On Tue, 13 Jul 2021, H.J. Lu wrote:

> On Mon, Jul 12, 2021 at 8:59 PM Wang, Pengfei <pengfei.wang@intel.com> wrote:
> >
> > > Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
> >
> > Can you please explain the behavior here? Is there difference between _Float16 and _Complex _Float16 when return? I.e.,
> > 1, In which case will _Float16 values return in both %xmm0 and %xmm1?
> > 2, For a single _Float16 value, are both real part and imaginary part returned in %xmm0? Or returned in %xmm0 and %xmm1 respectively?
> 
> Here is the v2 patch to add the missing _Float16 bits.   The PDF file is at
> 
> https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI

This PDF shows _Complex _Float16 as having a size of 2 bytes (should be 
4-byte size, 2-byte alignment).

It also seems to change double from 4-byte to 8-byte alignment, which is 
wrong.  And it's inconsistent about whether it covers the long double = 
double (Android) case - it shows that case for _Complex long double but 
not for long double itself.
H.J. Lu July 13, 2021, 4:24 p.m. UTC | #15 
On Tue, Jul 13, 2021 at 8:41 AM Joseph Myers <joseph@codesourcery.com> wrote:
>
> On Tue, 13 Jul 2021, H.J. Lu wrote:
>
> > On Mon, Jul 12, 2021 at 8:59 PM Wang, Pengfei <pengfei.wang@intel.com> wrote:
> > >
> > > > Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
> > >
> > > Can you please explain the behavior here? Is there difference between _Float16 and _Complex _Float16 when return? I.e.,
> > > 1, In which case will _Float16 values return in both %xmm0 and %xmm1?
> > > 2, For a single _Float16 value, are both real part and imaginary part returned in %xmm0? Or returned in %xmm0 and %xmm1 respectively?
> >
> > Here is the v2 patch to add the missing _Float16 bits.   The PDF file is at
> >
> > https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI
>
> This PDF shows _Complex _Float16 as having a size of 2 bytes (should be
> 4-byte size, 2-byte alignment).
>
> It also seems to change double from 4-byte to 8-byte alignment, which is
> wrong.  And it's inconsistent about whether it covers the long double =
> double (Android) case - it shows that case for _Complex long double but
> not for long double itself.

Here is the v3 patch with the fixes.  I also updated the PDF file.

> --
> Joseph S. Myers
> joseph@codesourcery.com
>
H.J. Lu July 29, 2021, 1:39 p.m. UTC | #16 
On Tue, Jul 13, 2021 at 9:24 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Tue, Jul 13, 2021 at 8:41 AM Joseph Myers <joseph@codesourcery.com> wrote:
> >
> > On Tue, 13 Jul 2021, H.J. Lu wrote:
> >
> > > On Mon, Jul 12, 2021 at 8:59 PM Wang, Pengfei <pengfei.wang@intel.com> wrote:
> > > >
> > > > > Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
> > > >
> > > > Can you please explain the behavior here? Is there difference between _Float16 and _Complex _Float16 when return? I.e.,
> > > > 1, In which case will _Float16 values return in both %xmm0 and %xmm1?
> > > > 2, For a single _Float16 value, are both real part and imaginary part returned in %xmm0? Or returned in %xmm0 and %xmm1 respectively?
> > >
> > > Here is the v2 patch to add the missing _Float16 bits.   The PDF file is at
> > >
> > > https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI
> >
> > This PDF shows _Complex _Float16 as having a size of 2 bytes (should be
> > 4-byte size, 2-byte alignment).
> >
> > It also seems to change double from 4-byte to 8-byte alignment, which is
> > wrong.  And it's inconsistent about whether it covers the long double =
> > double (Android) case - it shows that case for _Complex long double but
> > not for long double itself.
>
> Here is the v3 patch with the fixes.  I also updated the PDF file.

Here is the final patch I checked in.   _Complex _Float16 is changed to return
in XMM0 register.   The new PDF file is at

https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI
John McCall Aug. 24, 2021, 5:55 a.m. UTC | #17 
On Thu, Jul 29, 2021 at 9:40 AM H.J. Lu <hjl.tools@gmail.com> wrote:

> On Tue, Jul 13, 2021 at 9:24 AM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > On Tue, Jul 13, 2021 at 8:41 AM Joseph Myers <joseph@codesourcery.com>
> wrote:
> > >
> > > On Tue, 13 Jul 2021, H.J. Lu wrote:
> > >
> > > > On Mon, Jul 12, 2021 at 8:59 PM Wang, Pengfei <
> pengfei.wang@intel.com> wrote:
> > > > >
> > > > > > Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1
> registers.
> > > > >
> > > > > Can you please explain the behavior here? Is there difference
> between _Float16 and _Complex _Float16 when return? I.e.,
> > > > > 1, In which case will _Float16 values return in both %xmm0 and
> %xmm1?
> > > > > 2, For a single _Float16 value, are both real part and imaginary
> part returned in %xmm0? Or returned in %xmm0 and %xmm1 respectively?
> > > >
> > > > Here is the v2 patch to add the missing _Float16 bits.   The PDF
> file is at
> > > >
> > > > https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI
> > >
> > > This PDF shows _Complex _Float16 as having a size of 2 bytes (should be
> > > 4-byte size, 2-byte alignment).
> > >
> > > It also seems to change double from 4-byte to 8-byte alignment, which
> is
> > > wrong.  And it's inconsistent about whether it covers the long double =
> > > double (Android) case - it shows that case for _Complex long double but
> > > not for long double itself.
> >
> > Here is the v3 patch with the fixes.  I also updated the PDF file.
>
> Here is the final patch I checked in.   _Complex _Float16 is changed to
> return
> in XMM0 register.   The new PDF file is at
>
> https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI


This should be explicit that the real part is returned in bits 0..15 and
the imaginary part is returned in bits 16..31, or however we conventionally
designate subcomponents of a vector.

John.
H.J. Lu Aug. 25, 2021, 12:35 p.m. UTC | #18 
On Mon, Aug 23, 2021 at 10:55 PM John McCall <rjmccall@gmail.com> wrote:
>
> On Thu, Jul 29, 2021 at 9:40 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>>
>> On Tue, Jul 13, 2021 at 9:24 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>> >
>> > On Tue, Jul 13, 2021 at 8:41 AM Joseph Myers <joseph@codesourcery.com> wrote:
>> > >
>> > > On Tue, 13 Jul 2021, H.J. Lu wrote:
>> > >
>> > > > On Mon, Jul 12, 2021 at 8:59 PM Wang, Pengfei <pengfei.wang@intel.com> wrote:
>> > > > >
>> > > > > > Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1 registers.
>> > > > >
>> > > > > Can you please explain the behavior here? Is there difference between _Float16 and _Complex _Float16 when return? I.e.,
>> > > > > 1, In which case will _Float16 values return in both %xmm0 and %xmm1?
>> > > > > 2, For a single _Float16 value, are both real part and imaginary part returned in %xmm0? Or returned in %xmm0 and %xmm1 respectively?
>> > > >
>> > > > Here is the v2 patch to add the missing _Float16 bits.   The PDF file is at
>> > > >
>> > > > https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI
>> > >
>> > > This PDF shows _Complex _Float16 as having a size of 2 bytes (should be
>> > > 4-byte size, 2-byte alignment).
>> > >
>> > > It also seems to change double from 4-byte to 8-byte alignment, which is
>> > > wrong.  And it's inconsistent about whether it covers the long double =
>> > > double (Android) case - it shows that case for _Complex long double but
>> > > not for long double itself.
>> >
>> > Here is the v3 patch with the fixes.  I also updated the PDF file.
>>
>> Here is the final patch I checked in.   _Complex _Float16 is changed to return
>> in XMM0 register.   The new PDF file is at
>>
>> https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI
>
>
> This should be explicit that the real part is returned in bits 0..15 and the imaginary part is returned in bits 16..31, or however we conventionally designate subcomponents of a vector.
>
> John.

How about this?

diff --git a/low-level-sys-info.tex b/low-level-sys-info.tex
index 860ff66..8f527c1 100644
--- a/low-level-sys-info.tex
+++ b/low-level-sys-info.tex
@@ -457,6 +457,9 @@ and \texttt{unions}) are always returned in memory.
     & \texttt{__float128} & memory \\
     \hline
     & \texttt{_Complex _Float16} & \reg{xmm0} \\
+    & & The real part is returned in bits 0..15. The imaginary part is
+        returned \\
+    & & in bits 16..31.\\
     \cline{2-3}
     Complex & \texttt{_Complex float} & \EDX:\EAX \\
     floating- & & The real part is returned in \EAX. The imaginary part is

https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/uploads/89eb3e52c7e5eadd58f7597508e13f34/intel386-psABI-2021-08-25.pdf
John McCall Aug. 25, 2021, 8:32 p.m. UTC | #19 
On Wed, Aug 25, 2021 at 8:36 AM H.J. Lu <hjl.tools@gmail.com> wrote:

> On Mon, Aug 23, 2021 at 10:55 PM John McCall <rjmccall@gmail.com> wrote:
> > On Thu, Jul 29, 2021 at 9:40 AM H.J. Lu <hjl.tools@gmail.com> wrote:
> >> On Tue, Jul 13, 2021 at 9:24 AM H.J. Lu <hjl.tools@gmail.com> wrote:
> >> > On Tue, Jul 13, 2021 at 8:41 AM Joseph Myers <joseph@codesourcery.com>
> wrote:
> >> > > On Tue, 13 Jul 2021, H.J. Lu wrote:
> >> > > > On Mon, Jul 12, 2021 at 8:59 PM Wang, Pengfei <
> pengfei.wang@intel.com> wrote:
> >> > > > >
> >> > > > > > Return _Float16 and _Complex _Float16 values in %xmm0/%xmm1
> registers.
> >> > > > >
> >> > > > > Can you please explain the behavior here? Is there difference
> between _Float16 and _Complex _Float16 when return? I.e.,
> >> > > > > 1, In which case will _Float16 values return in both %xmm0 and
> %xmm1?
> >> > > > > 2, For a single _Float16 value, are both real part and
> imaginary part returned in %xmm0? Or returned in %xmm0 and %xmm1
> respectively?
> >> > > >
> >> > > > Here is the v2 patch to add the missing _Float16 bits.   The PDF
> file is at
> >> > > >
> >> > > > https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI
> >> > >
> >> > > This PDF shows _Complex _Float16 as having a size of 2 bytes
> (should be
> >> > > 4-byte size, 2-byte alignment).
> >> > >
> >> > > It also seems to change double from 4-byte to 8-byte alignment,
> which is
> >> > > wrong.  And it's inconsistent about whether it covers the long
> double =
> >> > > double (Android) case - it shows that case for _Complex long double
> but
> >> > > not for long double itself.
> >> >
> >> > Here is the v3 patch with the fixes.  I also updated the PDF file.
> >>
> >> Here is the final patch I checked in.   _Complex _Float16 is changed to
> return
> >> in XMM0 register.   The new PDF file is at
> >>
> >> https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/Intel386-psABI
> >
> >
> > This should be explicit that the real part is returned in bits 0..15 and
> the imaginary part is returned in bits 16..31, or however we conventionally
> designate subcomponents of a vector.
>
> How about this?
>
> diff --git a/low-level-sys-info.tex b/low-level-sys-info.tex
> index 860ff66..8f527c1 100644
> --- a/low-level-sys-info.tex
> +++ b/low-level-sys-info.tex
> @@ -457,6 +457,9 @@ and \texttt{unions}) are always returned in memory.
>      & \texttt{__float128} & memory \\
>      \hline
>      & \texttt{_Complex _Float16} & \reg{xmm0} \\
> +    & & The real part is returned in bits 0..15. The imaginary part is
> +        returned \\
> +    & & in bits 16..31.\\
>      \cline{2-3}
>      Complex & \texttt{_Complex float} & \EDX:\EAX \\
>      floating- & & The real part is returned in \EAX. The imaginary part is
>
>
> https://gitlab.com/x86-psABIs/i386-ABI/-/wikis/uploads/89eb3e52c7e5eadd58f7597508e13f34/intel386-psABI-2021-08-25.pdf


Looks good to me, thanks.

John.
diff mbox series

Patch

diff --git a/low-level-sys-info.tex b/low-level-sys-info.tex
index acaf30e..82956e3 100644
--- a/low-level-sys-info.tex
+++ b/low-level-sys-info.tex
@@ -30,7 +30,8 @@  object, and the term \emph{\textindex{\sixteenbyte{}}} refers to a
 \subsubsection{Fundamental Types}
 
 Table~\ref{basic-types} shows the correspondence between ISO C
-scalar types and the processor scalar types.  \code{__float80},
+scalar types and the processor scalar types.  \code{_Float16},
+\code{__float80},
 \code{__float128}, \code{__m64}, \code{__m128}, \code{__m256} and
 \code{__m512} types are optional.
 
@@ -79,22 +80,25 @@  scalar types and the processor scalar types.  \code{__float80},
     & \texttt{\textit{any-type} *} & 4 & 4 & unsigned \fourbyte \\
     & \texttt{\textit{any-type} (*)()} & & \\
     \hline
-    Floating-& \texttt{float} & 4 & 4 & single (IEEE-754) \\
     \cline{2-5}
-    point & \texttt{double} & 8 & 4 & double (IEEE-754) \\
-    & \texttt{long double}$^{\dagger\dagger\dagger\dagger}$  & & & \\
+    & \texttt{_Float16}$^{\dagger\dagger\dagger\dagger\dagger\dagger}$ & 2 & 2 & 16-bit (IEEE-754) \\
     \cline{2-5}
-    & \texttt{__float80}$^{\dagger\dagger}$  & 12 & 4 & 80-bit extended (IEEE-754) \\
-    & \texttt{long double}$^{\dagger\dagger\dagger\dagger}$  & & & \\
+    & \texttt{float} & 4 & 4 & single (IEEE-754) \\
+    \cline{2-5}
+    Floating- & \texttt{double} & 8
+	& 8$^{\dagger\dagger\dagger\dagger}$ & double (IEEE-754) \\
+    \cline{2-5}
+    point & \texttt{__float80}$^{\dagger\dagger}$  & 16 & 16 & 80-bit extended (IEEE-754) \\
+    & \texttt{long double}$^{\dagger\dagger\dagger\dagger\dagger}$  & 16 & 16 & 80-bit extended (IEEE-754) \\
     \cline{2-5}
     & \texttt{__float128}$^{\dagger\dagger}$ & 16 & 16 & 128-bit extended (IEEE-754) \\
     \hline
-    Complex& \texttt{_Complex float} & 8 & 4 & complex single (IEEE-754) \\
+    & \texttt{_Complex float} & 8 & 4 & complex single (IEEE-754) \\
     \cline{2-5}
-    Floating-& \texttt{_Complex double} & 16 & 4 & complex double (IEEE-754) \\
-    point & \texttt{_Complex long double}$^{\dagger\dagger\dagger\dagger}$ & & & \\
+    Complex& \texttt{_Complex double} & 16 & 4 & complex double (IEEE-754) \\
+    Floating-& \texttt{_Complex long double}$^{\dagger\dagger\dagger\dagger}$ & & & \\
     \cline{2-5}
-    & \texttt{_Complex __float80}$^{\dagger\dagger}$  & 24 & 4 & complex 80-bit extended (IEEE-754) \\
+    point & \texttt{_Complex __float80}$^{\dagger\dagger}$  & 24 & 4 & complex 80-bit extended (IEEE-754) \\
     & \texttt{_Complex long double}$^{\dagger\dagger\dagger\dagger}$  & & & \\
     \cline{2-5}
     & \texttt{_Complex __float128}$^{\dagger\dagger}$ & 32 & 16 & complex 128-bit extended (IEEE-754) \\
@@ -125,6 +129,8 @@  The \texttt{long double} type is 64-bit, the same as the \texttt{double}
 type, on the Android{\texttrademark} platform.  More information on the
 Android{\texttrademark} platform is available from
 \url{http://www.android.com/}.}\\
+\multicolumn{5}{p{13cm}}{\myfontsize $^{\dagger\dagger\dagger\dagger\dagger\dagger}$
+The \texttt{_Float16} type, from ISO/IEC TS 18661-3:2015, is optional.}\\
   \end{tabular}
 }
 \end{table}
@@ -323,6 +329,7 @@  at the time of the call.
 \begin{table}
 \Hrule
   \caption{Register Usage}
+  \myfontsize
   \label{fig-reg-usage}
   \begin{center}
     \begin{tabular}{l|p{8.35cm}|l}
@@ -346,13 +353,29 @@  of some 64bit return types & No \\
 \EBP & callee-saved register; optionally used as frame pointer & Yes \\
 \ESI & callee-saved register & yes \\
 \EDI & callee-saved register & yes \\
-\reg{xmm0}, \reg{ymm0} & scratch registers; also used to pass and return
-\code{__m128}, \code{__m256} parameters & No\\
-\reg{xmm1}--\reg{xmm2},& scratch registers; also used to pass
-\code{__m128}, & No \\
-\reg{ymm1}--\reg{ymm2} & \code{__m256} parameters & \\
-\reg{xmm3}--\reg{xmm7},& scratch registers & No \\
-\reg{ymm3}--\reg{ymm7} & & \\
+\reg{xmm0} & scratch register; also used to pass the first \code{__m128}
+             parameter and return \code{__m128}, \code{_Float16},
+	     the real part of \code{_Complex _Float16} & No \\
+\reg{ymm0} & scratch register; also used to pass the first \code{__m256}
+             parameter and return \code{__m256} & No \\
+\reg{zmm0} & scratch register; also used to pass the first \code{__m512}
+             parameter and return \code{__m512} & No \\
+\reg{xmm1} & scratch register; also used to pass the second \code{__m128}
+             parameter and return the imaginary part of
+	     \code{_Complex _Float16} & No \\
+\reg{ymm1} & scratch register; also used to pass the second \code{__m256}
+             parameters & No \\
+\reg{zmm1} & scratch register; also used to pass the second \code{__m512}
+             parameters & No \\
+\reg{xmm2} & scratch register; also used to pass the third \code{__m128}
+             parameters & No \\
+\reg{ymm2} & scratch register; also used to pass the third \code{__m256}
+             parameters & No \\
+\reg{zmm2} & scratch register; also used to pass the third \code{__m512}
+             parameters & No \\
+\reg{xmm3}--\reg{xmm7} & scratch registers & No \\
+\reg{ymm3}--\reg{ymm7} & scratch registers & No \\
+\reg{zmm3}--\reg{zmm7} & scratch registers & No \\
 \reg{mm0} & scratch register; also used to pass and return
 \code{__m64} parameter & No\\
 \reg{mm1}--\reg{mm2} & used to pass \code{__m64} parameters & No\\