/* {wcs|wcp|str|stp}cpy with 256/512-bit EVEX instructions.
   Copyright (C) 2021-2024 Free Software Foundation, Inc.
   This file is part of the GNU C Library.

   The GNU C Library is free software; you can redistribute it and/or
   modify it under the terms of the GNU Lesser General Public
   License as published by the Free Software Foundation; either
   version 2.1 of the License, or (at your option) any later version.

   The GNU C Library is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
   Lesser General Public License for more details.

   You should have received a copy of the GNU Lesser General Public
   License along with the GNU C Library; if not, see
   <https://www.gnu.org/licenses/>.  */

#include <isa-level.h>
#if ISA_SHOULD_BUILD (4)


	/* Use evex-masked stores for small sizes. Turned off at the
	   moment.  */
# define USE_EVEX_MASKED_STORE	0
	/* Use movsb in page cross case to save code size.  */
# define USE_MOVSB_IN_PAGE_CROSS	1

# include <sysdep.h>

# ifndef VEC_SIZE
#  include "x86-evex256-vecs.h"
# endif

# ifndef STRCPY
#  define STRCPY	__strcpy_evex
# endif


# ifdef USE_AS_WCSCPY
#  define VMOVU_MASK	vmovdqu32
#  define VPMIN	vpminud
#  define VPTESTN	vptestnmd
#  define VPTEST	vptestmd
#  define VPCMPEQ	vpcmpeqd
#  define CHAR_SIZE	4

#  define REP_MOVS	rep movsd

#  define USE_WIDE_CHAR
# else
#  define VMOVU_MASK	vmovdqu8
#  define VPMIN	vpminub
#  define VPTESTN	vptestnmb
#  define VPTEST	vptestmb
#  define VPCMPEQ	vpcmpeqb
#  define CHAR_SIZE	1

#  define REP_MOVS	rep movsb
# endif

# include "reg-macros.h"


# ifdef USE_AS_STPCPY
#  define END_REG	rax
# else
#  define END_REG	rdi, %rdx, CHAR_SIZE
# endif

# ifdef USE_AS_STRCAT
#  define PAGE_ALIGN_REG	edx
#  define PAGE_ALIGN_REG_64	rdx
# else
#  define PAGE_ALIGN_REG	eax
#  define PAGE_ALIGN_REG_64	rax
# endif

# define VZERO	VMM(7)
# define VZERO_128	VMM_128(7)


# define PAGE_SIZE	4096
# define CHAR_PER_VEC	(VEC_SIZE / CHAR_SIZE)


	.section SECTION(.text), "ax", @progbits
ENTRY(STRCPY)
# ifdef USE_AS_STRCAT
	movq	%rdi, %rax
#  include "strcat-strlen-evex.h.S"
# endif

	movl	%esi, %PAGE_ALIGN_REG
	andl	$(PAGE_SIZE - 1), %PAGE_ALIGN_REG
	cmpl	$(PAGE_SIZE - VEC_SIZE), %PAGE_ALIGN_REG
	ja	L(page_cross)
L(page_cross_continue):
	VMOVU	(%rsi), %VMM(0)
# if !defined USE_AS_STPCPY && !defined USE_AS_STRCAT
	movq	%rdi, %rax
# endif


	/* Two short string implementations. One with traditional
	   branching approach and one with masked instructions (which
	   have potential for dramatically bad perf if dst splits a
	   page and is not in the TLB).  */
# if USE_EVEX_MASKED_STORE
	VPTEST	%VMM(0), %VMM(0), %k0
	KMOV	%k0, %VRCX
#  ifdef USE_AS_WCSCPY
	subl	$((1 << CHAR_PER_VEC)- 1), %VRCX
#  else
	inc	%VRCX
#  endif
	jz	L(more_1x_vec)
	KMOV	%VRCX, %k1
	KXOR	%k0, %k1, %k1

	VMOVU_MASK %VMM(0), (%rdi){%k1}

#  ifdef USE_AS_STPCPY
	bsf	%VRCX, %VRCX
	leaq	(%rdi, %rcx, CHAR_SIZE), %rax
#  endif
	ret

# else
	VPTESTN	%VMM(0), %VMM(0), %k0
	KMOV	%k0, %VRCX
	test	%VRCX, %VRCX
	jz	L(more_1x_vec)

	xorl	%edx, %edx
	bsf	%VRCX, %VRDX
#  ifdef USE_AS_STPCPY
	leaq	(%rdi, %rdx, CHAR_SIZE), %rax
#  endif

	/* Use mask bits in rcx to detect which copy we need. If the low
	   mask is zero then there must be a bit set in the upper half.
	   I.e if rcx != 0 and ecx == 0, then match must be upper 32
	   bits so we use L(copy_32_63).  */
#  if VEC_SIZE == 64
#   ifdef USE_AS_WCSCPY
	testb	%cl, %cl
#   else
	testl	%ecx, %ecx
#   endif
	jz	L(copy_32_63)
#  endif

#  ifdef USE_AS_WCSCPY
	testb	$0xf, %cl
#  else
	testw	%cx, %cx
#  endif
	jz	L(copy_16_31)


#  ifdef USE_AS_WCSCPY
	testb	$0x3, %cl
#  else
	testb	%cl, %cl
#  endif
	jz	L(copy_8_15)


#  ifdef USE_AS_WCSCPY
	vmovd	%VMM_128(0), (%rdi)
	/* No need to copy, we know its zero.  */
	movl	$0, (%END_REG)

	ret
#  else

	testb	$0x7, %cl
	jz	L(copy_4_7)


	test	%edx, %edx
	jz	L(set_null_term)

	/* NB: make this `vmovw` if support for AVX512-FP16 is added.
	 */
	vmovd	%VMM_128(0), %esi
	movw	%si, (%rdi)

	.p2align 4,, 1
L(set_null_term):
	/* No need to copy, we know its zero.  */
	movb	$0, (%END_REG)
	ret
#  endif

#  if VEC_SIZE == 64
	.p2align 4,, 6
L(copy_32_63):
	VMOVU	-(32 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %VMM_256(1)
	VMOVU	%VMM_256(0), (%rdi)
	VMOVU	%VMM_256(1), -(32 - CHAR_SIZE)(%END_REG)
	ret
#  endif


	.p2align 4,, 6
L(copy_16_31):
	/* Use xmm1 explicitly here as it won't require a `vzeroupper`
	   and will save code size.  */
	vmovdqu	-(16 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %xmm1
	VMOVU	%VMM_128(0), (%rdi)
	vmovdqu	%xmm1, -(16 - CHAR_SIZE)(%END_REG)
	ret

	.p2align 4,, 8
L(copy_8_15):
#  ifdef USE_AS_WCSCPY
	movl	-(8 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %ecx
#  else
	movq	-(8 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %rcx
#  endif
	vmovq	%VMM_128(0), (%rdi)
	movq	%rcx, -(8 - CHAR_SIZE)(%END_REG)
	ret
# endif


# ifndef USE_AS_WCSCPY
	.p2align 4,, 12
L(copy_4_7):
	movl	-(4 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %ecx
	vmovd	%VMM_128(0), (%rdi)
	movl	%ecx, -(4 - CHAR_SIZE)(%END_REG)
	ret
# endif


	.p2align 4,, 8
L(more_1x_vec):
# if defined USE_AS_STPCPY || defined USE_AS_STRCAT
	VMOVU	%VMM(0), (%rdi)
# endif
	subq	%rsi, %rdi
	andq	$-(VEC_SIZE), %rsi
	addq	%rsi, %rdi
	VMOVA	(VEC_SIZE * 1)(%rsi), %VMM(1)

	/* Ideally we store after moves to minimize impact of potential
	   false-dependencies.  */
# if !defined USE_AS_STPCPY && !defined USE_AS_STRCAT
	VMOVU	%VMM(0), (%rax)
# endif

	VPTESTN	%VMM(1), %VMM(1), %k0
	KMOV	%k0, %VRCX
	test	%VRCX, %VRCX
	jnz	L(ret_vec_x1)

	VMOVA	(VEC_SIZE * 2)(%rsi), %VMM(2)
	VMOVU	%VMM(1), VEC_SIZE(%rdi)

	VPTESTN	%VMM(2), %VMM(2), %k0
	KMOV	%k0, %VRCX
	test	%VRCX, %VRCX
	jnz	L(ret_vec_x2)

	VMOVA	(VEC_SIZE * 3)(%rsi), %VMM(3)
	VMOVU	%VMM(2), (VEC_SIZE * 2)(%rdi)

	VPTESTN	%VMM(3), %VMM(3), %k0
	KMOV	%k0, %VRDX
	test	%VRDX, %VRDX
	jnz	L(ret_vec_x3)

	VMOVA	(VEC_SIZE * 4)(%rsi), %VMM(4)
	VMOVU	%VMM(3), (VEC_SIZE * 3)(%rdi)
	VPTESTN	%VMM(4), %VMM(4), %k0
	KMOV	%k0, %VRCX
	test	%VRCX, %VRCX
	jnz	L(ret_vec_x4)

	VMOVU	%VMM(4), (VEC_SIZE * 4)(%rdi)


	/* Align for 4x loop.  */
	subq	%rsi, %rdi

	/* + VEC_SIZE * 5 because we never added the original VEC_SIZE
	   we covered before aligning.  */
	subq	$-(VEC_SIZE * 5), %rsi
	andq	$-(VEC_SIZE * 4), %rsi


	/* Load first half of the loop before entry.  */
	VMOVA	(VEC_SIZE * 0 + 0)(%rsi), %VMM(0)
	VMOVA	(VEC_SIZE * 1 + 0)(%rsi), %VMM(1)
	VMOVA	(VEC_SIZE * 2 + 0)(%rsi), %VMM(2)
	VMOVA	(VEC_SIZE * 3 + 0)(%rsi), %VMM(3)

	VPMIN	%VMM(0), %VMM(1), %VMM(4)
	VPMIN	%VMM(2), %VMM(3), %VMM(6)
	VPTESTN	%VMM(4), %VMM(4), %k2
	VPTESTN	%VMM(6), %VMM(6), %k4
	KORTEST	%k2, %k4
	jnz	L(loop_4x_done)

	.p2align 4,, 11
L(loop_4x_vec):

	VMOVU	%VMM(0), (VEC_SIZE * 0 + 0)(%rdi, %rsi)
	VMOVU	%VMM(1), (VEC_SIZE * 1 + 0)(%rdi, %rsi)
	VMOVU	%VMM(2), (VEC_SIZE * 2 + 0)(%rdi, %rsi)
	VMOVU	%VMM(3), (VEC_SIZE * 3 + 0)(%rdi, %rsi)

	subq	$(VEC_SIZE * -4), %rsi

	VMOVA	(VEC_SIZE * 0 + 0)(%rsi), %VMM(0)
	VMOVA	(VEC_SIZE * 1 + 0)(%rsi), %VMM(1)
	VMOVA	(VEC_SIZE * 2 + 0)(%rsi), %VMM(2)
	VMOVA	(VEC_SIZE * 3 + 0)(%rsi), %VMM(3)


	VPMIN	%VMM(0), %VMM(1), %VMM(4)
	VPMIN	%VMM(2), %VMM(3), %VMM(6)
	VPTESTN	%VMM(4), %VMM(4), %k2
	VPTESTN	%VMM(6), %VMM(6), %k4
	KORTEST	%k2, %k4
	jz	L(loop_4x_vec)

L(loop_4x_done):
	VPTESTN	%VMM(0), %VMM(0), %k0
	KMOV	%k0, %VRCX
	/* Restore rdi (%rdi).  */
	addq	%rsi, %rdi
	test	%VRCX, %VRCX
	jnz	L(ret_vec_x0_end)
	VMOVU	%VMM(0), (VEC_SIZE * 0 + 0)(%rdi)

	KMOV	%k2, %VRCX
	test	%VRCX, %VRCX
	jnz	L(ret_vec_x1)
	VMOVU	%VMM(1), (VEC_SIZE * 1 + 0)(%rdi)

	VPTESTN	%VMM(2), %VMM(2), %k0
	KMOV	%k0, %VRCX
	test	%VRCX, %VRCX
	jnz	L(ret_vec_x2)
	VMOVU	%VMM(2), (VEC_SIZE * 2 + 0)(%rdi)
	/* Place L(ret_vec_x4) here to save code size.  We get a
	   meaningfuly benefit doing this for stpcpy.  */
	KMOV	%k4, %VRDX
L(ret_vec_x3):
	bsf	%VRDX, %VRDX
	VMOVU	((VEC_SIZE * 3)-(VEC_SIZE - CHAR_SIZE))(%rsi, %rdx, CHAR_SIZE), %VMM(0)
	VMOVU	%VMM(0), ((VEC_SIZE * 3 + 0)-(VEC_SIZE - CHAR_SIZE))(%rdi, %rdx, CHAR_SIZE)
# ifdef USE_AS_STPCPY
	leaq	(VEC_SIZE * 3 + 0)(%rdi, %rdx, CHAR_SIZE), %rax
# endif
L(return_end):
	ret

	.p2align 4,, 6
L(ret_vec_x0_end):
	bsf	%VRCX, %VRCX
# ifdef USE_AS_STPCPY
	leaq	(%rdi, %rcx, CHAR_SIZE), %rax
# endif
	inc	%VRCX
	VMOVU	(-(VEC_SIZE))(%rsi, %rcx, CHAR_SIZE), %VMM(0)
	VMOVU	%VMM(0), (-(VEC_SIZE))(%rdi, %rcx, CHAR_SIZE)
	ret

	.p2align 4,, 8
L(ret_vec_x1):
	bsf	%VRCX, %VRCX
	VMOVU	(VEC_SIZE -(VEC_SIZE - CHAR_SIZE))(%rsi, %rcx, CHAR_SIZE), %VMM(0)
	VMOVU	%VMM(0), (VEC_SIZE -(VEC_SIZE - CHAR_SIZE))(%rdi, %rcx, CHAR_SIZE)
# ifdef USE_AS_STPCPY
	leaq	VEC_SIZE(%rdi, %rcx, CHAR_SIZE), %rax
# endif
	ret

	.p2align 4,, 4
L(ret_vec_x2):
	bsf	%VRCX, %VRCX
	VMOVU	((VEC_SIZE * 2)-(VEC_SIZE - CHAR_SIZE))(%rsi, %rcx, CHAR_SIZE), %VMM(0)
	VMOVU	%VMM(0), ((VEC_SIZE * 2)-(VEC_SIZE - CHAR_SIZE))(%rdi, %rcx, CHAR_SIZE)
# ifdef USE_AS_STPCPY
	leaq	(VEC_SIZE * 2)(%rdi, %rcx, CHAR_SIZE), %rax
# endif
	ret

	/* ret_vec_x3 reuses return code after the loop.  */
	.p2align 4,, 6
L(ret_vec_x4):
	bsf	%VRCX, %VRCX
	VMOVU	((VEC_SIZE * 4)-(VEC_SIZE - CHAR_SIZE))(%rsi, %rcx, CHAR_SIZE), %VMM(0)
	VMOVU	%VMM(0), ((VEC_SIZE * 4)-(VEC_SIZE - CHAR_SIZE))(%rdi, %rcx, CHAR_SIZE)
# ifdef USE_AS_STPCPY
	leaq	(VEC_SIZE * 4)(%rdi, %rcx, CHAR_SIZE), %rax
# endif
	ret


	.p2align 4,, 4
L(page_cross):
# ifndef USE_AS_STRCAT
	vpxorq	%VZERO_128, %VZERO_128, %VZERO_128
# endif
	movq	%rsi, %rcx
	andq	$(VEC_SIZE * -1), %rcx

	VPCMPEQ	(%rcx), %VZERO, %k0
	KMOV	%k0, %VRCX
# ifdef USE_AS_WCSCPY
	andl	$(VEC_SIZE - 1), %PAGE_ALIGN_REG
	shrl	$2, %PAGE_ALIGN_REG
# endif
	shrx	%VGPR(PAGE_ALIGN_REG_64), %VRCX, %VRCX

# if USE_MOVSB_IN_PAGE_CROSS
	/* Optimizing more aggressively for space as this is very cold
	   code. This saves 2x cache lines.  */

	/* This adds once to the later result which will get correct
	   copy bounds. NB: this can never zero-out a non-zero RCX as
	   to be in the page cross case rsi cannot be aligned and we
	   already right-shift rcx by the misalignment.  */
	shl	%VRCX
	jz	L(page_cross_continue)
#  if !defined USE_AS_STPCPY && !defined USE_AS_STRCAT
	movq	%rdi, %rax
#  endif
	bsf	%VRCX, %VRCX
	REP_MOVS

#  ifdef USE_AS_STPCPY
	leaq	-CHAR_SIZE(%rdi), %rax
#  endif
	ret


# else
	/* Check if we found zero-char before end of page.  */
	test	%VRCX, %VRCX
	jz	L(page_cross_continue)

	/* Traditional copy case, essentially same as used in non-page-
	   cross case but since we can't reuse VMM(0) we need twice as
	   many loads from rsi.  */

#  ifndef USE_AS_STRCAT
	xorl	%edx, %edx
#  endif
	/* Dependency on rdi must already have been satisfied.  */
	bsf	%VRCX, %VRDX
#  ifdef USE_AS_STPCPY
	leaq	(%rdi, %rdx, CHAR_SIZE), %rax
#  elif !defined USE_AS_STRCAT
	movq	%rdi, %rax
#  endif

#  if VEC_SIZE == 64
#   ifdef USE_AS_WCSCPY
	testb	%cl, %cl
#   else
	test	%ecx, %ecx
#   endif
	jz	L(page_cross_copy_32_63)
#  endif

#  ifdef USE_AS_WCSCPY
	testb	$0xf, %cl
#  else
	testw	%cx, %cx
#  endif
	jz	L(page_cross_copy_16_31)

#  ifdef USE_AS_WCSCPY
	testb	$0x3, %cl
#  else
	testb	%cl, %cl
#  endif
	jz	L(page_cross_copy_8_15)

#  ifdef USE_AS_WCSCPY
	movl	(%rsi), %esi
	movl	%esi, (%rdi)
	movl	$0, (%END_REG)
	ret
#  else

	testb	$0x7, %cl
	jz	L(page_cross_copy_4_7)

	test	%edx, %edx
	jz	L(page_cross_set_null_term)
	movzwl	(%rsi), %ecx
	movw	%cx, (%rdi)
L(page_cross_set_null_term):
	movb	$0, (%END_REG)
	ret


	.p2align 4,, 4
L(page_cross_copy_4_7):
	movl	(%rsi), %ecx
	movl	-(4 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %esi
	movl	%ecx, (%rdi)
	movl	%esi, -(4 - CHAR_SIZE)(%END_REG)
	ret
#  endif

#  if VEC_SIZE == 64
	.p2align 4,, 4
L(page_cross_copy_32_63):
	VMOVU	(%rsi), %VMM_256(0)
	VMOVU	-(32 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %VMM_256(1)
	VMOVU	%VMM_256(0), (%rdi)
	VMOVU	%VMM_256(1), -(32 - CHAR_SIZE)(%END_REG)
	ret
#  endif

	.p2align 4,, 4
L(page_cross_copy_16_31):
	vmovdqu	(%rsi), %xmm0
	vmovdqu	-(16 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %xmm1
	vmovdqu	%xmm0, (%rdi)
	vmovdqu	%xmm1, -(16 - CHAR_SIZE)(%END_REG)
	ret

	.p2align 4,, 4
L(page_cross_copy_8_15):
	movq	(%rsi), %rcx
	movq	-(8 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %rsi
	movq	%rcx, (%rdi)
	movq	%rsi, -(8 - CHAR_SIZE)(%END_REG)
	ret
# endif
END(STRCPY)
#endif