Re: PATCH: Add AVX support to x86-64 _dl_runtime_profile

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Mon, Jul 6, 2009 at 10:16 PM, Ulrich Drepper<drepper@redhat.com> wrote:
> H.J. Lu wrote:
>> Since SSE registers are aliases of the lower 128bits of AVX register,
>> we only need to restore AVX register if it is available.
>
> No. ?This would only work if we would have had the AVX registers in the
> data structure from day one. ?The current ABI allows audit modules to
> modify the lr_xmm array and see the changes being reflected in the
> program run. ?You cannot after that fact require that the lr_ymm
> registers have to be modified.
>
> This is why this two-step process is needed.

Here is the updated patch with 2 testcases. They are passed under
AVX emulator.

>
>> Since AVX register is just SSE register extended to 256bit, we only
>> use the first 8 AVX/SSE registers for passing parameters.
>
> This makes no sense. ?There are 16 of them and if a function needs more
> than 8 it should be able to use it. ?Someone should fix gcc.
>

The current psABI only uses the first 8 vector registers to pass
function parameters. We only  eed to save/restore the first 8
vector registers

>> Here is the updated patch tested under AVX emulator.
>
> As said above, the restoring issue has to be addressed.

Done.  tst-audit5 changes xmm fields.  It worked under
AVX emulator.

>
>> It defines
>> La_x86_64_vector as 64byte since we will extend 256bit AVX register
>> to 512bit in the future.
>
> OK I guess.


Thanks.

-- 
H.J.
---
2009-07-07  H.J. Lu  <hongjiu.lu@intel.com>

	* config.h.in (HAVE_AVX_SUPPORT): New.

	* config.make.in (config-cflags-avx): New.

	* configure.in: Substitute libc_cv_cc_avx.
	* configure: Regenerated.

	* elf/Makefile (distribute): Add tst-audit4.c tst-auditmod4a.c
	tst-auditmod4b.c tst-audit5.c tst-auditmod5a.c
	tst-auditmod5b.c.
	(tests): Add tst-audit4 tst-audit5 for x86_64.
	(modules-names): Add tst-auditmod4a tst-auditmod4b
	tst-auditmod5a tst-auditmod5b.
	($(objpfx)tst-audit4): New.
	($(objpfx)tst-audit4.out): Likewise.
	(tst-audit4-ENV): Likewise.
	($(objpfx)tst-audit5): Likewise.
	($(objpfx)tst-audit5.out): Likewise.
	(tst-audit5-ENV): Likewise.
	(CFLAGS-tst-audit4.c): Likewise.
	(CFLAGS-tst-auditmod4a.c): Likewise.
	(CFLAGS-tst-auditmod4b.c): Likewise.

	* elf/tst-audit4.c: New.
	* elf/tst-audit5.c: Likewise.
	* elf/tst-auditmod4a.c: Likewise.
	* elf/tst-auditmod4b.c: Likewise.
	* elf/tst-auditmod5a.c: Likewise.
	* elf/tst-auditmod5b.c: Likewise.

	* sysdeps/x86_64/Makefile (gen-as-const-headers): Add
	link-defines.sym.

	* sysdeps/x86_64/bits/link.h (La_x86_64_ymm): New.
	(La_x86_64_vector): Likewise.
	(La_x86_64_regs): Append lr_vector.
	(La_x86_64_retval): Append lr_vector0/lrv_vector1.

	* sysdeps/x86_64/dl-trampoline.S (_dl_runtime_profile): Move
	saving and restoring SSE registers to ...
	* sysdeps/x86_64/dl-trampoline.h: This.  New.

	* sysdeps/x86_64/dl-trampoline.S: Include <config.h> and
	<link-defines.h>.
	(_dl_runtime_profile): Use LR_SIZE to allocate space for
	La_x86_64_regs.  Allocate extra space and jump to memory at
	save_and_restore_vector if HAVE_AVX_SUPPORT is defined.
	(save_and_restore_vector_sse): New.
	(save_and_restore_vector_avx): Likewise.
	(check_avx): Likewise.
	(save_and_restore_vector): Likewise.

	* sysdeps/x86_64/elf/configure.in: Set libc_cv_cc_avx and
	HAVE_AVX_SUPPORT.
	* sysdeps/x86_64/elf/configure: Regenerated.

	* sysdeps/x86_64/link-defines.sym: New.

Attachment: libc-avx-4.patch
Description: Text document