Commit 411ccd30 authored by David S. Miller's avatar David S. Miller

[SPARC64]: Simplify and optimize ultra3 memory copies.

- Use one unrolled loop instead of two
- Improve performance for well aligned
  copies smaller than 256 bytes
- Use better named local label names so
  that diffing between the various instances
  is much easier to read
- Reverify exception handling
Signed-off-by: default avatarDavid S. Miller <davem@redhat.com>
parent 47228097
/* $Id: U3copy_from_user.S,v 1.4 2002/01/15 07:16:26 davem Exp $
* U3memcpy.S: UltraSparc-III optimized copy from userspace.
/* U3copy_from_user.S: UltraSparc-III optimized copy from userspace.
*
* Copyright (C) 1999, 2000 David S. Miller (davem@redhat.com)
* Copyright (C) 1999, 2000, 2004 David S. Miller (davem@redhat.com)
*/
#ifdef __KERNEL__
#include <asm/visasm.h>
#include <asm/asi.h>
#include <asm/dcu.h>
#include <asm/spitfire.h>
#undef SMALL_COPY_USES_FPU
#define XCC xcc
#define EXNV_RAW(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
99: ba U3cfu_fixup; \
a, b, %o1; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#define EXNV(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
99: VISExitHalf; \
99: add %o1, %o3, %o0; \
ba U3cfu_fixup; \
a, b, %o1; \
.section __ex_table; \
......@@ -22,6 +33,32 @@
.word 98b, 99b; \
.text; \
.align 4;
#define EXNV4(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
99: add %o1, %o3, %o0; \
a, b, %o1; \
ba U3cfu_fixup; \
add %o1, 4, %o1; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#define EXNV8(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
99: add %o1, %o3, %o0; \
a, b, %o1; \
ba U3cfu_fixup; \
add %o1, 8, %o1; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#define EX(x,y,a,b) \
98: x,y; \
.section .fixup; \
......@@ -77,18 +114,9 @@
.word 98b, 99b; \
.text; \
.align 4;
#else
#define ASI_BLK_P 0xf0
#define FPRS_FEF 0x04
#define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs
#define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
#define SMALL_COPY_USES_FPU
#define EXNV(x,y,a,b) x,y;
#define EX(x,y,a,b) x,y;
#define EX2(x,y) x,y;
#define EX3(x,y) x,y;
#define EX4(x,y) x,y;
#endif
.register %g2,#scratch
.register %g3,#scratch
/* Special/non-trivial issues of this code:
*
......@@ -109,81 +137,55 @@
* of up to 2.4GB per second.
*/
.globl U3copy_from_user
U3copy_from_user: /* %o0=dst, %o1=src, %o2=len */
#ifndef __KERNEL__
/* Save away original 'dst' for memcpy return value. */
mov %o0, %g3 ! A0 Group
#endif
/* Anything to copy at all? */
cmp %o2, 0 ! A1
ble,pn %icc, U3copy_from_user_short_ret! BR
/* Extremely small copy? */
cmp %o2, 31 ! A0 Group
ble,pn %icc, U3copy_from_user_short ! BR
/* Large enough to use unrolled prefetch loops? */
cmp %o2, 0x100 ! A1
bge,a,pt %icc, U3copy_from_user_enter ! BR Group
andcc %o0, 0x3f, %g2 ! A0
ba,pt %xcc, U3copy_from_user_toosmall ! BR Group
andcc %o0, 0x7, %g2 ! A0
.align 32
U3copy_from_user_short:
/* Copy %o2 bytes from src to dst, one byte at a time. */
EXNV(lduba [%o1 + 0x00] %asi, %o3, add %o2, %g0)! MS Group
add %o1, 0x1, %o1 ! A0
add %o0, 0x1, %o0 ! A1
subcc %o2, 1, %o2 ! A0 Group
bg,pt %icc, U3copy_from_user_short ! BR
stb %o3, [%o0 + -1] ! MS Group (1-cycle stall)
U3copy_from_user_short_ret:
#ifdef __KERNEL__
retl ! BR Group (0-4 cycle stall)
clr %o0 ! A0
#else
retl ! BR Group (0-4 cycle stall)
mov %g3, %o0 ! A0
#endif
/* Here len >= (6 * 64) and condition codes reflect execution
.globl U3copy_from_user
U3copy_from_user: /* %o0=dst, %o1=src, %o2=len */
cmp %o2, 0
be,pn %XCC, out
or %o0, %o1, %o3
cmp %o2, 16
bleu,a,pn %XCC, small_copy
or %o3, %o2, %o3
cmp %o2, 256
blu,pt %XCC, medium_copy
andcc %o3, 0x7, %g0
ba,pt %xcc, enter
andcc %o0, 0x3f, %g2
/* Here len >= 256 and condition codes reflect execution
* of "andcc %o0, 0x7, %g2", done by caller.
*/
.align 64
U3copy_from_user_enter:
enter:
/* Is 'dst' already aligned on an 64-byte boundary? */
be,pt %xcc, 2f ! BR
be,pt %XCC, 2f
/* Compute abs((dst & 0x3f) - 0x40) into %g2. This is the number
* of bytes to copy to make 'dst' 64-byte aligned. We pre-
* subtract this from 'len'.
*/
sub %g2, 0x40, %g2 ! A0 Group
sub %g0, %g2, %g2 ! A0 Group
sub %o2, %g2, %o2 ! A0 Group
sub %g2, 0x40, %g2
sub %g0, %g2, %g2
sub %o2, %g2, %o2
/* Copy %g2 bytes from src to dst, one byte at a time. */
1: EXNV(lduba [%o1 + 0x00] %asi, %o3, add %o2, %g2)! MS (Group)
add %o1, 0x1, %o1 ! A1
add %o0, 0x1, %o0 ! A0 Group
subcc %g2, 0x1, %g2 ! A1
1: EXNV_RAW(lduba [%o1 + 0x00] %asi, %o3, add %o2, %g2)
add %o1, 0x1, %o1
add %o0, 0x1, %o0
subcc %g2, 0x1, %g2
bg,pt %icc, 1b ! BR Group
stb %o3, [%o0 + -1] ! MS Group
bg,pt %XCC, 1b
stb %o3, [%o0 + -1]
2: VISEntryHalf ! MS+MS
and %o1, 0x7, %g1 ! A1
ba,pt %xcc, U3copy_from_user_begin ! BR
alignaddr %o1, %g0, %o1 ! MS (Break-after)
2: VISEntryHalf
and %o1, 0x7, %g1
ba,pt %xcc, begin
alignaddr %o1, %g0, %o1
.align 64
U3copy_from_user_begin:
#ifdef __KERNEL__
begin:
.globl U3copy_from_user_nop_1_6
U3copy_from_user_nop_1_6:
ldxa [%g0] ASI_DCU_CONTROL_REG, %g3
......@@ -192,315 +194,225 @@ U3copy_from_user_nop_1_6:
or %g3, %o3, %o3
stxa %o3, [%g0] ASI_DCU_CONTROL_REG ! Enable P-cache
membar #Sync
#endif
prefetcha [%o1 + 0x000] %asi, #one_read ! MS Group1
prefetcha [%o1 + 0x040] %asi, #one_read ! MS Group2
andn %o2, (0x40 - 1), %o4 ! A0
prefetcha [%o1 + 0x080] %asi, #one_read ! MS Group3
cmp %o4, 0x140 ! A0
prefetcha [%o1 + 0x0c0] %asi, #one_read ! MS Group4
EX(ldda [%o1 + 0x000] %asi, %f0, add %o2, %g0) ! MS Group5 (%f0 results at G8)
bge,a,pt %icc, 1f ! BR
prefetcha [%o1 + 0x100] %asi, #one_read ! MS Group6
1: EX(ldda [%o1 + 0x008] %asi, %f2, add %o2, %g0) ! AX (%f2 results at G9)
cmp %o4, 0x180 ! A1
bge,a,pt %icc, 1f ! BR
prefetcha [%o1 + 0x140] %asi, #one_read ! MS Group7
1: EX(ldda [%o1 + 0x010] %asi, %f4, add %o2, %g0) ! AX (%f4 results at G10)
cmp %o4, 0x1c0 ! A1
bge,a,pt %icc, 1f ! BR
prefetcha [%o1 + 0x180] %asi, #one_read ! MS Group8
1: faligndata %f0, %f2, %f16 ! FGA Group9 (%f16 at G12)
EX(ldda [%o1 + 0x018] %asi, %f6, add %o2, %g0) ! AX (%f6 results at G12)
faligndata %f2, %f4, %f18 ! FGA Group10 (%f18 results at G13)
EX(ldda [%o1 + 0x020] %asi, %f8, add %o2, %g0) ! MS (%f8 results at G13)
faligndata %f4, %f6, %f20 ! FGA Group12 (1-cycle stall,%f20 at G15)
EX(ldda [%o1 + 0x028] %asi, %f10, add %o2, %g0) ! MS (%f10 results at G15)
faligndata %f6, %f8, %f22 ! FGA Group13 (%f22 results at G16)
EX(ldda [%o1 + 0x030] %asi, %f12, add %o2, %g0) ! MS (%f12 results at G16)
faligndata %f8, %f10, %f24 ! FGA Group15 (1-cycle stall,%f24 at G18)
EX(ldda [%o1 + 0x038] %asi, %f14, add %o2, %g0) ! MS (%f14 results at G18)
faligndata %f10, %f12, %f26 ! FGA Group16 (%f26 results at G19)
EX(ldda [%o1 + 0x040] %asi, %f0, add %o2, %g0) ! MS (%f0 results at G19)
/* We only use the first loop if len > (7 * 64). */
subcc %o4, 0x1c0, %o4 ! A0 Group17
bg,pt %icc, U3copy_from_user_loop1 ! BR
add %o1, 0x40, %o1 ! A1
add %o4, 0x140, %o4 ! A0 Group18
ba,pt %xcc, U3copy_from_user_loop2 ! BR
srl %o4, 6, %o3 ! A0 Group19
nop
nop
nop
nop
nop
nop
nop
/* This loop performs the copy and queues new prefetches.
* We drop into the second loop when len <= (5 * 64). Note
* that this (5 * 64) factor has been subtracted from len
* already.
*/
U3copy_from_user_loop1:
EX2(ldda [%o1 + 0x008] %asi, %f2) ! MS Group2 (%f2 results at G5)
faligndata %f12, %f14, %f28 ! FGA (%f28 results at G5)
EX2(ldda [%o1 + 0x010] %asi, %f4) ! MS Group3 (%f4 results at G6)
faligndata %f14, %f0, %f30 ! FGA Group4 (1-cycle stall, %f30 at G7)
stda %f16, [%o0] ASI_BLK_P ! MS
EX2(ldda [%o1 + 0x018] %asi, %f6) ! AX (%f6 results at G7)
faligndata %f0, %f2, %f16 ! FGA Group12 (7-cycle stall)
EX2(ldda [%o1 + 0x020] %asi, %f8) ! MS (%f8 results at G15)
faligndata %f2, %f4, %f18 ! FGA Group13 (%f18 results at G16)
EX2(ldda [%o1 + 0x028] %asi, %f10) ! MS (%f10 results at G16)
faligndata %f4, %f6, %f20 ! FGA Group14 (%f20 results at G17)
EX2(ldda [%o1 + 0x030] %asi, %f12) ! MS (%f12 results at G17)
faligndata %f6, %f8, %f22 ! FGA Group15 (%f22 results at G18)
EX2(ldda [%o1 + 0x038] %asi, %f14) ! MS (%f14 results at G18)
faligndata %f8, %f10, %f24 ! FGA Group16 (%f24 results at G19)
EX2(ldda [%o1 + 0x040] %asi, %f0) ! AX (%f0 results at G19)
prefetcha [%o1 + 0x180] %asi, #one_read ! MS
faligndata %f10, %f12, %f26 ! FGA Group17 (%f26 results at G20)
subcc %o4, 0x40, %o4 ! A0
add %o1, 0x40, %o1 ! A1
bg,pt %xcc, U3copy_from_user_loop1 ! BR
add %o0, 0x40, %o0 ! A0 Group18
U3copy_from_user_loop2_enter:
mov 5, %o3 ! A1
/* This loop performs on the copy, no new prefetches are
* queued. We do things this way so that we do not perform
* any spurious prefetches past the end of the src buffer.
*/
U3copy_from_user_loop2:
EX3(ldda [%o1 + 0x008] %asi, %f2) ! MS
faligndata %f12, %f14, %f28 ! FGA Group2
EX3(ldda [%o1 + 0x010] %asi, %f4) ! MS
faligndata %f14, %f0, %f30 ! FGA Group4 (1-cycle stall)
stda %f16, [%o0] ASI_BLK_P ! MS
EX3(ldda [%o1 + 0x018] %asi, %f6) ! AX
faligndata %f0, %f2, %f16 ! FGA Group12 (7-cycle stall)
EX3(ldda [%o1 + 0x020] %asi, %f8) ! MS
faligndata %f2, %f4, %f18 ! FGA Group13
EX3(ldda [%o1 + 0x028] %asi, %f10) ! MS
faligndata %f4, %f6, %f20 ! FGA Group14
EX3(ldda [%o1 + 0x030] %asi, %f12) ! MS
faligndata %f6, %f8, %f22 ! FGA Group15
EX3(ldda [%o1 + 0x038] %asi, %f14) ! MS
faligndata %f8, %f10, %f24 ! FGA Group16
EX3(ldda [%o1 + 0x040] %asi, %f0) ! AX
faligndata %f10, %f12, %f26 ! FGA Group17
subcc %o3, 0x01, %o3 ! A0
add %o1, 0x40, %o1 ! A1
bg,pt %xcc, U3copy_from_user_loop2 ! BR
add %o0, 0x40, %o0 ! A0 Group18
prefetcha [%o1 + 0x000] %asi, #one_read
prefetcha [%o1 + 0x040] %asi, #one_read
andn %o2, (0x40 - 1), %o4
prefetcha [%o1 + 0x080] %asi, #one_read
prefetcha [%o1 + 0x0c0] %asi, #one_read
EX(ldda [%o1 + 0x000] %asi, %f0, add %o2, %g0)
prefetcha [%o1 + 0x100] %asi, #one_read
EX(ldda [%o1 + 0x008] %asi, %f2, add %o2, %g0)
prefetcha [%o1 + 0x140] %asi, #one_read
EX(ldda [%o1 + 0x010] %asi, %f4, add %o2, %g0)
prefetcha [%o1 + 0x180] %asi, #one_read
faligndata %f0, %f2, %f16
EX(ldda [%o1 + 0x018] %asi, %f6, add %o2, %g0)
faligndata %f2, %f4, %f18
EX(ldda [%o1 + 0x020] %asi, %f8, add %o2, %g0)
faligndata %f4, %f6, %f20
EX(ldda [%o1 + 0x028] %asi, %f10, add %o2, %g0)
faligndata %f6, %f8, %f22
EX(ldda [%o1 + 0x030] %asi, %f12, add %o2, %g0)
faligndata %f8, %f10, %f24
EX(ldda [%o1 + 0x038] %asi, %f14, add %o2, %g0)
faligndata %f10, %f12, %f26
EX(ldda [%o1 + 0x040] %asi, %f0, add %o2, %g0)
sub %o4, 0x80, %o4
add %o1, 0x40, %o1
ba,pt %xcc, loop
srl %o4, 6, %o3
.align 64
loop:
EX3(ldda [%o1 + 0x008] %asi, %f2)
faligndata %f12, %f14, %f28
EX3(ldda [%o1 + 0x010] %asi, %f4)
faligndata %f14, %f0, %f30
stda %f16, [%o0] ASI_BLK_P
EX3(ldda [%o1 + 0x018] %asi, %f6)
faligndata %f0, %f2, %f16
EX3(ldda [%o1 + 0x020] %asi, %f8)
faligndata %f2, %f4, %f18
EX3(ldda [%o1 + 0x028] %asi, %f10)
faligndata %f4, %f6, %f20
EX3(ldda [%o1 + 0x030] %asi, %f12)
faligndata %f6, %f8, %f22
EX3(ldda [%o1 + 0x038] %asi, %f14)
faligndata %f8, %f10, %f24
EX3(ldda [%o1 + 0x040] %asi, %f0)
prefetcha [%o1 + 0x180] %asi, #one_read
faligndata %f10, %f12, %f26
subcc %o3, 0x01, %o3
add %o1, 0x40, %o1
bg,pt %XCC, loop
add %o0, 0x40, %o0
/* Finally we copy the last full 64-byte block. */
U3copy_from_user_loopfini:
EX3(ldda [%o1 + 0x008] %asi, %f2) ! MS
faligndata %f12, %f14, %f28 ! FGA
EX3(ldda [%o1 + 0x010] %asi, %f4) ! MS Group19
faligndata %f14, %f0, %f30 ! FGA
stda %f16, [%o0] ASI_BLK_P ! MS Group20
EX3(ldda [%o1 + 0x018] %asi, %f6) ! AX
faligndata %f0, %f2, %f16 ! FGA Group11 (7-cycle stall)
EX3(ldda [%o1 + 0x020] %asi, %f8) ! MS
faligndata %f2, %f4, %f18 ! FGA Group12
EX3(ldda [%o1 + 0x028] %asi, %f10) ! MS
faligndata %f4, %f6, %f20 ! FGA Group13
EX3(ldda [%o1 + 0x030] %asi, %f12) ! MS
faligndata %f6, %f8, %f22 ! FGA Group14
EX3(ldda [%o1 + 0x038] %asi, %f14) ! MS
faligndata %f8, %f10, %f24 ! FGA Group15
cmp %g1, 0 ! A0
be,pt %icc, 1f ! BR
add %o0, 0x40, %o0 ! A1
EX4(ldda [%o1 + 0x040] %asi, %f0) ! MS
1: faligndata %f10, %f12, %f26 ! FGA Group16
faligndata %f12, %f14, %f28 ! FGA Group17
faligndata %f14, %f0, %f30 ! FGA Group18
stda %f16, [%o0] ASI_BLK_P ! MS
add %o0, 0x40, %o0 ! A0
add %o1, 0x40, %o1 ! A1
#ifdef __KERNEL__
loopfini:
EX3(ldda [%o1 + 0x008] %asi, %f2)
faligndata %f12, %f14, %f28
EX3(ldda [%o1 + 0x010] %asi, %f4)
faligndata %f14, %f0, %f30
stda %f16, [%o0] ASI_BLK_P
EX3(ldda [%o1 + 0x018] %asi, %f6)
faligndata %f0, %f2, %f16
EX3(ldda [%o1 + 0x020] %asi, %f8)
faligndata %f2, %f4, %f18
EX3(ldda [%o1 + 0x028] %asi, %f10)
faligndata %f4, %f6, %f20
EX3(ldda [%o1 + 0x030] %asi, %f12)
faligndata %f6, %f8, %f22
EX3(ldda [%o1 + 0x038] %asi, %f14)
faligndata %f8, %f10, %f24
cmp %g1, 0
be,pt %XCC, 1f
add %o0, 0x40, %o0
EX4(ldda [%o1 + 0x040] %asi, %f0)
1: faligndata %f10, %f12, %f26
faligndata %f12, %f14, %f28
faligndata %f14, %f0, %f30
stda %f16, [%o0] ASI_BLK_P
add %o0, 0x40, %o0
add %o1, 0x40, %o1
.globl U3copy_from_user_nop_2_3
U3copy_from_user_nop_2_3:
mov PRIMARY_CONTEXT, %o3
stxa %g0, [%o3] ASI_DMMU ! Flush P-cache
stxa %g3, [%g0] ASI_DCU_CONTROL_REG ! Disable P-cache
#endif
membar #Sync ! MS Group26 (7-cycle stall)
membar #Sync
/* Now we copy the (len modulo 64) bytes at the end.
* Note how we borrow the %f0 loaded above.
*
* Also notice how this code is careful not to perform a
* load past the end of the src buffer just like similar
* code found in U3copy_from_user_toosmall processing.
* load past the end of the src buffer.
*/
U3copy_from_user_loopend:
and %o2, 0x3f, %o2 ! A0 Group
andcc %o2, 0x38, %g2 ! A0 Group
be,pn %icc, U3copy_from_user_endcruft ! BR
subcc %g2, 0x8, %g2 ! A1
be,pn %icc, U3copy_from_user_endcruft ! BR Group
cmp %g1, 0 ! A0
be,a,pt %icc, 1f ! BR Group
EX(ldda [%o1 + 0x00] %asi, %f0, add %o2, %g0) ! MS
1: EX(ldda [%o1 + 0x08] %asi, %f2, add %o2, %g0) ! MS Group
add %o1, 0x8, %o1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f0, %f2, %f8 ! FGA Group
std %f8, [%o0 + 0x00] ! MS (XXX does it stall here? XXX)
be,pn %icc, U3copy_from_user_endcruft ! BR
add %o0, 0x8, %o0 ! A0
EX(ldda [%o1 + 0x08] %asi, %f0, add %o2, %g0) ! MS Group
add %o1, 0x8, %o1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f2, %f0, %f8 ! FGA
std %f8, [%o0 + 0x00] ! MS (XXX does it stall here? XXX)
bne,pn %icc, 1b ! BR
add %o0, 0x8, %o0 ! A0 Group
loopend:
and %o2, 0x3f, %o2
andcc %o2, 0x38, %g2
be,pn %XCC, endcruft
subcc %g2, 0x8, %g2
be,pn %XCC, endcruft
cmp %g1, 0
be,a,pt %XCC, 1f
EX(ldda [%o1 + 0x00] %asi, %f0, add %o2, %g0)
1: EX(ldda [%o1 + 0x08] %asi, %f2, add %o2, %g0)
add %o1, 0x8, %o1
sub %o2, 0x8, %o2
subcc %g2, 0x8, %g2
faligndata %f0, %f2, %f8
std %f8, [%o0 + 0x00]
be,pn %XCC, endcruft
add %o0, 0x8, %o0
EX(ldda [%o1 + 0x08] %asi, %f0, add %o2, %g0)
add %o1, 0x8, %o1
sub %o2, 0x8, %o2
subcc %g2, 0x8, %g2
faligndata %f2, %f0, %f8
std %f8, [%o0 + 0x00]
bne,pn %XCC, 1b
add %o0, 0x8, %o0
/* If anything is left, we copy it one byte at a time.
* Note that %g1 is (src & 0x3) saved above before the
* alignaddr was performed.
*/
U3copy_from_user_endcruft:
endcruft:
cmp %o2, 0
add %o1, %g1, %o1
VISExitHalf
be,pn %icc, U3copy_from_user_short_ret
nop
ba,a,pt %xcc, U3copy_from_user_short
/* If we get here, then 32 <= len < (6 * 64) */
U3copy_from_user_toosmall:
#ifdef SMALL_COPY_USES_FPU
/* Is 'dst' already aligned on an 8-byte boundary? */
be,pt %xcc, 2f ! BR Group
/* Compute abs((dst & 7) - 8) into %g2. This is the number
* of bytes to copy to make 'dst' 8-byte aligned. We pre-
* subtract this from 'len'.
*/
sub %g2, 0x8, %g2 ! A0
sub %g0, %g2, %g2 ! A0 Group (reg-dep)
sub %o2, %g2, %o2 ! A0 Group (reg-dep)
be,pn %XCC, out
sub %o0, %o1, %o3
/* Copy %g2 bytes from src to dst, one byte at a time. */
1: EXNV(lduba [%o1 + 0x00] %asi, %o3, add %o2, %g2)! MS (Group) (%o3 in 3 cycles)
add %o1, 0x1, %o1 ! A1
add %o0, 0x1, %o0 ! A0 Group
subcc %g2, 0x1, %g2 ! A1
bg,pt %icc, 1b ! BR Group
stb %o3, [%o0 + -1] ! MS Group
andcc %g1, 0x7, %g0
bne,pn %icc, small_copy_unaligned
andcc %o2, 0x8, %g0
be,pt %icc, 1f
nop
EXNV(ldxa [%o1] %asi, %o5, add %o2, %g0)
stx %o5, [%o1 + %o3]
add %o1, 0x8, %o1
2: VISEntryHalf ! MS+MS
1: andcc %o2, 0x4, %g0
be,pt %icc, 1f
nop
EXNV(lduwa [%o1] %asi, %o5, and %o2, 0x7)
stw %o5, [%o1 + %o3]
add %o1, 0x4, %o1
/* Compute (len - (len % 8)) into %g2. This is guaranteed
* to be nonzero.
*/
andn %o2, 0x7, %g2 ! A0 Group
/* You may read this and believe that it allows reading
* one 8-byte longword past the end of src. It actually
* does not, as %g2 is subtracted as loads are done from
* src, so we always stop before running off the end.
* Also, we are guaranteed to have at least 0x10 bytes
* to move here.
*/
sub %g2, 0x8, %g2 ! A0 Group (reg-dep)
alignaddr %o1, %g0, %g1 ! MS (Break-after)
EX(ldda [%g1 + 0x00] %asi, %f0, add %o2, %g0) ! MS Group (1-cycle stall)
add %g1, 0x8, %g1 ! A0
1: EX(ldda [%g1 + 0x00] %asi, %f2, add %o2, %g0) ! MS Group
add %g1, 0x8, %g1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f0, %f2, %f8 ! FGA Group (1-cycle stall)
std %f8, [%o0 + 0x00] ! MS Group (2-cycle stall)
add %o1, 0x8, %o1 ! A0
be,pn %icc, 2f ! BR
add %o0, 0x8, %o0 ! A1
EX(ldda [%g1 + 0x00] %asi, %f0, add %o2, %g0) ! MS Group
add %g1, 0x8, %g1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f2, %f0, %f8 ! FGA Group (1-cycle stall)
std %f8, [%o0 + 0x00] ! MS Group (2-cycle stall)
add %o1, 0x8, %o1 ! A0
bne,pn %icc, 1b ! BR
add %o0, 0x8, %o0 ! A1
/* Nothing left to copy? */
2: cmp %o2, 0 ! A0 Group
VISExitHalf ! A0+MS
be,pn %icc, U3copy_from_user_short_ret! BR Group
nop ! A0
ba,a,pt %xcc, U3copy_from_user_short ! BR Group
#else /* !(SMALL_COPY_USES_FPU) */
xor %o1, %o0, %g2
andcc %g2, 0x7, %g0
bne,pn %icc, U3copy_from_user_short
andcc %o1, 0x7, %g2
be,pt %xcc, 2f
sub %g2, 0x8, %g2
sub %g0, %g2, %g2
sub %o2, %g2, %o2
1: andcc %o2, 0x2, %g0
be,pt %icc, 1f
nop
EXNV(lduha [%o1] %asi, %o5, and %o2, 0x3)
sth %o5, [%o1 + %o3]
add %o1, 0x2, %o1
1: EXNV(lduba [%o1 + 0x00] %asi, %o3, add %o2, %g2)
add %o1, 0x1, %o1
add %o0, 0x1, %o0
subcc %g2, 0x1, %g2
bg,pt %icc, 1b
stb %o3, [%o0 + -1]
1: andcc %o2, 0x1, %g0
be,pt %icc, out
nop
EXNV(lduba [%o1] %asi, %o5, and %o2, 0x1)
ba,pt %xcc, out
stb %o5, [%o1 + %o3]
medium_copy: /* 16 < len <= 64 */
bne,pn %XCC, small_copy_unaligned
sub %o0, %o1, %o3
medium_copy_aligned:
andn %o2, 0x7, %o4
and %o2, 0x7, %o2
1: subcc %o4, 0x8, %o4
EXNV8(ldxa [%o1] %asi, %o5, add %o2, %o4)
stx %o5, [%o1 + %o3]
bgu,pt %XCC, 1b
add %o1, 0x8, %o1
andcc %o2, 0x4, %g0
be,pt %XCC, 1f
nop
sub %o2, 0x4, %o2
EXNV4(lduwa [%o1] %asi, %o5, add %o2, %g0)
stw %o5, [%o1 + %o3]
add %o1, 0x4, %o1
1: cmp %o2, 0
be,pt %XCC, out
nop
ba,pt %xcc, small_copy_unaligned
nop
2: andn %o2, 0x7, %g2
sub %o2, %g2, %o2
small_copy: /* 0 < len <= 16 */
andcc %o3, 0x3, %g0
bne,pn %XCC, small_copy_unaligned
sub %o0, %o1, %o3
3: EXNV(ldxa [%o1 + 0x00] %asi, %o3, add %o2, %g2)
add %o1, 0x8, %o1
add %o0, 0x8, %o0
subcc %g2, 0x8, %g2
bg,pt %icc, 3b
stx %o3, [%o0 + -8]
small_copy_aligned:
subcc %o2, 4, %o2
EXNV(lduwa [%o1] %asi, %g1, add %o2, %g0)
stw %g1, [%o1 + %o3]
bgu,pt %XCC, small_copy_aligned
add %o1, 4, %o1
cmp %o2, 0
bne,pn %icc, U3copy_from_user_short
nop
ba,a,pt %xcc, U3copy_from_user_short_ret
out: retl
clr %o0
#endif /* !(SMALL_COPY_USES_FPU) */
.align 32
small_copy_unaligned:
subcc %o2, 1, %o2
EXNV(lduba [%o1] %asi, %g1, add %o2, %g0)
stb %g1, [%o1 + %o3]
bgu,pt %XCC, small_copy_unaligned
add %o1, 1, %o1
retl
clr %o0
#ifdef __KERNEL__
.globl U3cfu_fixup
U3cfu_fixup:
/* Since this is copy_from_user(), zero out the rest of the
* kernel buffer.
......@@ -516,4 +428,3 @@ U3cfu_fixup:
2: retl
mov %o1, %o0
#endif
/* $Id: U3copy_in_user.S,v 1.4 2001/03/21 05:58:47 davem Exp $
* U3memcpy.S: UltraSparc-III optimized copy within userspace.
/* U3copy_in_user.S: UltraSparc-III optimized memcpy.
*
* Copyright (C) 1999, 2000 David S. Miller (davem@redhat.com)
* Copyright (C) 1999, 2000, 2004 David S. Miller (davem@redhat.com)
*/
#ifdef __KERNEL__
#include <asm/visasm.h>
#include <asm/asi.h>
#undef SMALL_COPY_USES_FPU
#include <asm/dcu.h>
#include <asm/spitfire.h>
#define XCC xcc
#define EXNV(x,y,a,b) \
98: x,y; \
.section .fixup; \
......@@ -19,7 +21,7 @@
.word 98b, 99b; \
.text; \
.align 4;
#define EXNV2(x,y,a,b) \
#define EXNV1(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
......@@ -31,501 +33,108 @@
.word 98b, 99b; \
.text; \
.align 4;
#define EXNV3(x,y,a,b) \
#define EXNV4(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
99: a, b, %o0; \
retl; \
add %o0, 8, %o0; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#define EX(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
99: VISExitHalf; \
retl; \
a, b, %o0; \
add %o0, 4, %o0; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#define EXBLK1(x,y) \
#define EXNV8(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
99: VISExitHalf; \
add %o4, 0x1c0, %o1; \
and %o2, (0x40 - 1), %o2; \
retl; \
add %o1, %o2, %o0; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#define EXBLK2(x,y) \
98: x,y; \
.section .fixup; \
.align 4; \
99: VISExitHalf; \
sll %o3, 6, %o3; \
and %o2, (0x40 - 1), %o2; \
add %o3, 0x80, %o1; \
retl; \
add %o1, %o2, %o0; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#define EXBLK3(x,y) \
98: x,y; \
.section .fixup; \
.align 4; \
99: VISExitHalf; \
and %o2, (0x40 - 1), %o2; \
retl; \
add %o2, 0x80, %o0; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#define EXBLK4(x,y) \
98: x,y; \
.section .fixup; \
.align 4; \
99: VISExitHalf; \
and %o2, (0x40 - 1), %o2; \
99: a, b, %o0; \
retl; \
add %o2, 0x40, %o0; \
add %o0, 8, %o0; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#else
#define ASI_AIUS 0x80
#define ASI_BLK_AIUS 0xf0
#define FPRS_FEF 0x04
#define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs
#define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
#define SMALL_COPY_USES_FPU
#define EXNV(x,y,a,b) x,y;
#define EXNV2(x,y,a,b) x,y;
#define EXNV3(x,y,a,b) x,y;
#define EX(x,y,a,b) x,y;
#define EXBLK1(x,y) x,y;
#define EXBLK2(x,y) x,y;
#define EXBLK3(x,y) x,y;
#define EXBLK4(x,y) x,y;
#endif
/* Special/non-trivial issues of this code:
*
* 1) %o5 is preserved from VISEntryHalf to VISExitHalf
* 2) Only low 32 FPU registers are used so that only the
* lower half of the FPU register set is dirtied by this
* code. This is especially important in the kernel.
* 3) This code never prefetches cachelines past the end
* of the source buffer.
*
* XXX Actually, Cheetah can buffer up to 8 concurrent
* XXX prefetches, revisit this...
*/
.register %g2,#scratch
.register %g3,#scratch
.text
.align 32
/* The cheetah's flexible spine, oversized liver, enlarged heart,
* slender muscular body, and claws make it the swiftest hunter
* in Africa and the fastest animal on land. Can reach speeds
* of up to 2.4GB per second.
/* Don't try to get too fancy here, just nice and
* simple. This is predominantly used for well aligned
* small copies in the compat layer. It is also used
* to copy register windows around during thread cloning.
*/
.globl U3copy_in_user
U3copy_in_user: /* %o0=dst, %o1=src, %o2=len */
.globl U3copy_in_user
U3copy_in_user: /* %o0=dst, %o1=src, %o2=len */
/* Writing to %asi is _expensive_ so we hardcode it.
* Reading %asi to check for KERNEL_DS is comparatively
* cheap.
*/
rd %asi, %g1 ! MS Group (4 cycles)
cmp %g1, ASI_AIUS ! A0 Group
bne U3memcpy ! BR
nop ! A1
#ifndef __KERNEL__
/* Save away original 'dst' for memcpy return value. */
mov %o0, %g3 ! A0 Group
#endif
/* Anything to copy at all? */
cmp %o2, 0 ! A1
ble,pn %icc, U3copy_in_user_short_ret ! BR
/* Extremely small copy? */
cmp %o2, 31 ! A0 Group
ble,pn %icc, U3copy_in_user_short ! BR
/* Large enough to use unrolled prefetch loops? */
cmp %o2, 0x100 ! A1
bge,a,pt %icc, U3copy_in_user_enter ! BR Group
andcc %o0, 0x3f, %g2 ! A0
ba,pt %xcc, U3copy_in_user_toosmall ! BR Group
andcc %o0, 0x7, %g2 ! A0
.align 32
U3copy_in_user_short:
/* Copy %o2 bytes from src to dst, one byte at a time. */
EXNV(lduba [%o1 + 0x00] %asi, %o3, add %o2, %g0)! MS Group
add %o1, 0x1, %o1 ! A0
add %o0, 0x1, %o0 ! A1
subcc %o2, 1, %o2 ! A0 Group
bg,pt %icc, U3copy_in_user_short ! BR
EXNV(stba %o3, [%o0 + -1] %asi, add %o2, 1) ! MS Group (1-cycle stall)
U3copy_in_user_short_ret:
#ifdef __KERNEL__
retl ! BR Group (0-4 cycle stall)
clr %o0 ! A0
#else
retl ! BR Group (0-4 cycle stall)
mov %g3, %o0 ! A0
#endif
/* Here len >= (6 * 64) and condition codes reflect execution
* of "andcc %o0, 0x7, %g2", done by caller.
*/
.align 64
U3copy_in_user_enter:
/* Is 'dst' already aligned on an 64-byte boundary? */
be,pt %xcc, 2f ! BR
/* Compute abs((dst & 0x3f) - 0x40) into %g2. This is the number
* of bytes to copy to make 'dst' 64-byte aligned. We pre-
* subtract this from 'len'.
*/
sub %g2, 0x40, %g2 ! A0 Group
sub %g0, %g2, %g2 ! A0 Group
sub %o2, %g2, %o2 ! A0 Group
/* Copy %g2 bytes from src to dst, one byte at a time. */
1: EXNV(lduba [%o1 + 0x00] %asi, %o3, add %o2, %g2)! MS (Group)
add %o1, 0x1, %o1 ! A1
add %o0, 0x1, %o0 ! A0 Group
subcc %g2, 0x1, %g2 ! A1
bg,pt %icc, 1b ! BR Group
EXNV2(stba %o3, [%o0 + -1] %asi, add %o2, %g2) ! MS Group
2: VISEntryHalf ! MS+MS
and %o1, 0x7, %g1 ! A1
ba,pt %xcc, U3copy_in_user_begin ! BR
alignaddr %o1, %g0, %o1 ! MS (Break-after)
.align 64
U3copy_in_user_begin:
prefetcha [%o1 + 0x000] %asi, #one_read ! MS Group1
prefetcha [%o1 + 0x040] %asi, #one_read ! MS Group2
andn %o2, (0x40 - 1), %o4 ! A0
prefetcha [%o1 + 0x080] %asi, #one_read ! MS Group3
cmp %o4, 0x140 ! A0
prefetcha [%o1 + 0x0c0] %asi, #one_read ! MS Group4
EX(ldda [%o1 + 0x000] %asi, %f0, add %o2, %g0) ! MS Group5 (%f0 results at G8)
bge,a,pt %icc, 1f ! BR
prefetcha [%o1 + 0x100] %asi, #one_read ! MS Group6
1: EX(ldda [%o1 + 0x008] %asi, %f2, add %o2, %g0) ! AX (%f2 results at G9)
cmp %o4, 0x180 ! A1
bge,a,pt %icc, 1f ! BR
prefetcha [%o1 + 0x140] %asi, #one_read ! MS Group7
1: EX(ldda [%o1 + 0x010] %asi, %f4, add %o2, %g0) ! AX (%f4 results at G10)
cmp %o4, 0x1c0 ! A1
bge,a,pt %icc, 1f ! BR
prefetcha [%o1 + 0x180] %asi, #one_read ! MS Group8
1: faligndata %f0, %f2, %f16 ! FGA Group9 (%f16 at G12)
EX(ldda [%o1 + 0x018] %asi, %f6, add %o2, %g0) ! AX (%f6 results at G12)
faligndata %f2, %f4, %f18 ! FGA Group10 (%f18 results at G13)
EX(ldda [%o1 + 0x020] %asi, %f8, add %o2, %g0) ! MS (%f8 results at G13)
faligndata %f4, %f6, %f20 ! FGA Group12 (1-cycle stall,%f20 at G15)
EX(ldda [%o1 + 0x028] %asi, %f10, add %o2, %g0) ! MS (%f10 results at G15)
faligndata %f6, %f8, %f22 ! FGA Group13 (%f22 results at G16)
EX(ldda [%o1 + 0x030] %asi, %f12, add %o2, %g0) ! MS (%f12 results at G16)
faligndata %f8, %f10, %f24 ! FGA Group15 (1-cycle stall,%f24 at G18)
EX(ldda [%o1 + 0x038] %asi, %f14, add %o2, %g0) ! MS (%f14 results at G18)
faligndata %f10, %f12, %f26 ! FGA Group16 (%f26 results at G19)
EX(ldda [%o1 + 0x040] %asi, %f0, add %o2, %g0) ! MS (%f0 results at G19)
/* We only use the first loop if len > (7 * 64). */
subcc %o4, 0x1c0, %o4 ! A0 Group17
bg,pt %icc, U3copy_in_user_loop1 ! BR
add %o1, 0x40, %o1 ! A1
add %o4, 0x140, %o4 ! A0 Group18
ba,pt %xcc, U3copy_in_user_loop2 ! BR
srl %o4, 6, %o3 ! A0 Group19
nop
nop
nop
nop
nop
nop
nop
/* This loop performs the copy and queues new prefetches.
* We drop into the second loop when len <= (5 * 64). Note
* that this (5 * 64) factor has been subtracted from len
* already.
*/
U3copy_in_user_loop1:
EXBLK1(ldda [%o1 + 0x008] %asi, %f2) ! MS Group2 (%f2 results at G5)
faligndata %f12, %f14, %f28 ! FGA (%f28 results at G5)
EXBLK1(ldda [%o1 + 0x010] %asi, %f4) ! MS Group3 (%f4 results at G6)
faligndata %f14, %f0, %f30 ! FGA Group4 (1-cycle stall, %f30 at G7)
EXBLK1(stda %f16, [%o0] ASI_BLK_AIUS) ! MS
EXBLK1(ldda [%o1 + 0x018] %asi, %f6) ! AX (%f6 results at G7)
faligndata %f0, %f2, %f16 ! FGA Group12 (7-cycle stall)
EXBLK1(ldda [%o1 + 0x020] %asi, %f8) ! MS (%f8 results at G15)
faligndata %f2, %f4, %f18 ! FGA Group13 (%f18 results at G16)
EXBLK1(ldda [%o1 + 0x028] %asi, %f10) ! MS (%f10 results at G16)
faligndata %f4, %f6, %f20 ! FGA Group14 (%f20 results at G17)
EXBLK1(ldda [%o1 + 0x030] %asi, %f12) ! MS (%f12 results at G17)
faligndata %f6, %f8, %f22 ! FGA Group15 (%f22 results at G18)
EXBLK1(ldda [%o1 + 0x038] %asi, %f14) ! MS (%f14 results at G18)
faligndata %f8, %f10, %f24 ! FGA Group16 (%f24 results at G19)
EXBLK1(ldda [%o1 + 0x040] %asi, %f0) ! AX (%f0 results at G19)
prefetcha [%o1 + 0x180] %asi, #one_read ! MS
faligndata %f10, %f12, %f26 ! FGA Group17 (%f26 results at G20)
subcc %o4, 0x40, %o4 ! A0
add %o1, 0x40, %o1 ! A1
bg,pt %xcc, U3copy_in_user_loop1 ! BR
add %o0, 0x40, %o0 ! A0 Group18
U3copy_in_user_loop2_enter:
mov 5, %o3 ! A1
/* This loop performs on the copy, no new prefetches are
* queued. We do things this way so that we do not perform
* any spurious prefetches past the end of the src buffer.
*/
U3copy_in_user_loop2:
EXBLK2(ldda [%o1 + 0x008] %asi, %f2) ! MS
faligndata %f12, %f14, %f28 ! FGA Group2
EXBLK2(ldda [%o1 + 0x010] %asi, %f4) ! MS
faligndata %f14, %f0, %f30 ! FGA Group4 (1-cycle stall)
EXBLK2(stda %f16, [%o0] ASI_BLK_AIUS) ! MS
EXBLK2(ldda [%o1 + 0x018] %asi, %f6) ! AX
faligndata %f0, %f2, %f16 ! FGA Group12 (7-cycle stall)
EXBLK2(ldda [%o1 + 0x020] %asi, %f8) ! MS
faligndata %f2, %f4, %f18 ! FGA Group13
EXBLK2(ldda [%o1 + 0x028] %asi, %f10) ! MS
faligndata %f4, %f6, %f20 ! FGA Group14
EXBLK2(ldda [%o1 + 0x030] %asi, %f12) ! MS
faligndata %f6, %f8, %f22 ! FGA Group15
EXBLK2(ldda [%o1 + 0x038] %asi, %f14) ! MS
faligndata %f8, %f10, %f24 ! FGA Group16
EXBLK2(ldda [%o1 + 0x040] %asi, %f0) ! AX
faligndata %f10, %f12, %f26 ! FGA Group17
subcc %o3, 0x01, %o3 ! A0
add %o1, 0x40, %o1 ! A1
bg,pt %xcc, U3copy_in_user_loop2 ! BR
add %o0, 0x40, %o0 ! A0 Group18
/* Finally we copy the last full 64-byte block. */
U3copy_in_user_loopfini:
EXBLK3(ldda [%o1 + 0x008] %asi, %f2) ! MS
faligndata %f12, %f14, %f28 ! FGA
EXBLK3(ldda [%o1 + 0x010] %asi, %f4) ! MS Group19
faligndata %f14, %f0, %f30 ! FGA
EXBLK3(stda %f16, [%o0] ASI_BLK_AIUS) ! MS Group20
EXBLK4(ldda [%o1 + 0x018] %asi, %f6) ! AX
faligndata %f0, %f2, %f16 ! FGA Group11 (7-cycle stall)
EXBLK4(ldda [%o1 + 0x020] %asi, %f8) ! MS
faligndata %f2, %f4, %f18 ! FGA Group12
EXBLK4(ldda [%o1 + 0x028] %asi, %f10) ! MS
faligndata %f4, %f6, %f20 ! FGA Group13
EXBLK4(ldda [%o1 + 0x030] %asi, %f12) ! MS
faligndata %f6, %f8, %f22 ! FGA Group14
EXBLK4(ldda [%o1 + 0x038] %asi, %f14) ! MS
faligndata %f8, %f10, %f24 ! FGA Group15
cmp %g1, 0 ! A0
be,pt %icc, 1f ! BR
add %o0, 0x40, %o0 ! A1
EXBLK4(ldda [%o1 + 0x040] %asi, %f0) ! MS
1: faligndata %f10, %f12, %f26 ! FGA Group16
faligndata %f12, %f14, %f28 ! FGA Group17
faligndata %f14, %f0, %f30 ! FGA Group18
EXBLK4(stda %f16, [%o0] ASI_BLK_AIUS) ! MS
add %o0, 0x40, %o0 ! A0
add %o1, 0x40, %o1 ! A1
membar #Sync ! MS Group26 (7-cycle stall)
/* Now we copy the (len modulo 64) bytes at the end.
* Note how we borrow the %f0 loaded above.
*
* Also notice how this code is careful not to perform a
* load past the end of the src buffer just like similar
* code found in U3copy_in_user_toosmall processing.
*/
U3copy_in_user_loopend:
and %o2, 0x3f, %o2 ! A0 Group
andcc %o2, 0x38, %g2 ! A0 Group
be,pn %icc, U3copy_in_user_endcruft ! BR
subcc %g2, 0x8, %g2 ! A1
be,pn %icc, U3copy_in_user_endcruft ! BR Group
cmp %g1, 0 ! A0
be,a,pt %icc, 1f ! BR Group
EX(ldda [%o1 + 0x00] %asi, %f0, add %o2, %g0) ! MS
1: EX(ldda [%o1 + 0x08] %asi, %f2, add %o2, %g0) ! MS Group
add %o1, 0x8, %o1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f0, %f2, %f8 ! FGA Group
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8) ! MS (XXX does it stall here? XXX)
be,pn %icc, U3copy_in_user_endcruft ! BR
add %o0, 0x8, %o0 ! A0
EX(ldda [%o1 + 0x08] %asi, %f0, add %o2, %g0) ! MS Group
add %o1, 0x8, %o1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f2, %f0, %f8 ! FGA
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8) ! MS (XXX does it stall here? XXX)
bne,pn %icc, 1b ! BR
add %o0, 0x8, %o0 ! A0 Group
rd %asi, %g1
cmp %g1, ASI_AIUS
bne,pn %icc, U3memcpy_user_stub
nop
/* If anything is left, we copy it one byte at a time.
* Note that %g1 is (src & 0x3) saved above before the
* alignaddr was performed.
*/
U3copy_in_user_endcruft:
cmp %o2, 0
add %o1, %g1, %o1
VISExitHalf
be,pn %icc, U3copy_in_user_short_ret
be,pn %XCC, out
or %o0, %o1, %o3
cmp %o2, 16
bleu,a,pn %XCC, small_copy
or %o3, %o2, %o3
medium_copy: /* 16 < len <= 64 */
andcc %o3, 0x7, %g0
bne,pn %XCC, small_copy_unaligned
sub %o0, %o1, %o3
medium_copy_aligned:
andn %o2, 0x7, %o4
and %o2, 0x7, %o2
1: subcc %o4, 0x8, %o4
EXNV8(ldxa [%o1] %asi, %o5, add %o4, %o2)
EXNV8(stxa %o5, [%o1 + %o3] ASI_AIUS, add %o4, %o2)
bgu,pt %XCC, 1b
add %o1, 0x8, %o1
andcc %o2, 0x4, %g0
be,pt %XCC, 1f
nop
sub %o2, 0x4, %o2
EXNV4(lduwa [%o1] %asi, %o5, add %o4, %o2)
EXNV4(stwa %o5, [%o1 + %o3] ASI_AIUS, add %o4, %o2)
add %o1, 0x4, %o1
1: cmp %o2, 0
be,pt %XCC, out
nop
ba,pt %xcc, small_copy_unaligned
nop
ba,a,pt %xcc, U3copy_in_user_short
/* If we get here, then 32 <= len < (6 * 64) */
U3copy_in_user_toosmall:
#ifdef SMALL_COPY_USES_FPU
/* Is 'dst' already aligned on an 8-byte boundary? */
be,pt %xcc, 2f ! BR Group
/* Compute abs((dst & 7) - 8) into %g2. This is the number
* of bytes to copy to make 'dst' 8-byte aligned. We pre-
* subtract this from 'len'.
*/
sub %g2, 0x8, %g2 ! A0
sub %g0, %g2, %g2 ! A0 Group (reg-dep)
sub %o2, %g2, %o2 ! A0 Group (reg-dep)
/* Copy %g2 bytes from src to dst, one byte at a time. */
1: EXNV2(lduba [%o1 + 0x00] %asi, %o3, add %o2, %g2)! MS (Group) (%o3 in 3 cycles)
add %o1, 0x1, %o1 ! A1
add %o0, 0x1, %o0 ! A0 Group
subcc %g2, 0x1, %g2 ! A1
bg,pt %icc, 1b ! BR Group
EXNV2(stba %o3, [%o0 + -1] %asi, add %o2, %g2) ! MS Group
2: VISEntryHalf ! MS+MS
/* Compute (len - (len % 8)) into %g2. This is guaranteed
* to be nonzero.
*/
andn %o2, 0x7, %g2 ! A0 Group
/* You may read this and believe that it allows reading
* one 8-byte longword past the end of src. It actually
* does not, as %g2 is subtracted as loads are done from
* src, so we always stop before running off the end.
* Also, we are guaranteed to have at least 0x10 bytes
* to move here.
*/
sub %g2, 0x8, %g2 ! A0 Group (reg-dep)
alignaddr %o1, %g0, %g1 ! MS (Break-after)
EX(ldda [%g1 + 0x00] %asi, %f0, add %o2, %g0) ! MS Group (1-cycle stall)
add %g1, 0x8, %g1 ! A0
1: EX(ldda [%g1 + 0x00] %asi, %f2, add %o2, %g0) ! MS Group
add %g1, 0x8, %g1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f0, %f2, %f8 ! FGA Group (1-cycle stall)
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8) ! MS Group (2-cycle stall)
add %o1, 0x8, %o1 ! A0
be,pn %icc, 2f ! BR
add %o0, 0x8, %o0 ! A1
EX(ldda [%g1 + 0x00] %asi, %f0, add %o2, %g0) ! MS Group
add %g1, 0x8, %g1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f2, %f0, %f8 ! FGA Group (1-cycle stall)
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8) ! MS Group (2-cycle stall)
add %o1, 0x8, %o1 ! A0
bne,pn %icc, 1b ! BR
add %o0, 0x8, %o0 ! A1
/* Nothing left to copy? */
2: cmp %o2, 0 ! A0 Group
VISExitHalf ! A0+MS
be,pn %icc, U3copy_in_user_short_ret ! BR Group
nop ! A0
ba,a,pt %xcc, U3copy_in_user_short ! BR Group
#else /* !(SMALL_COPY_USES_FPU) */
xor %o1, %o0, %g2
andcc %g2, 0x7, %g0
bne,pn %icc, U3copy_in_user_short
andcc %o1, 0x7, %g2
be,pt %xcc, 2f
sub %g2, 0x8, %g2
sub %g0, %g2, %g2
sub %o2, %g2, %o2
1: EXNV2(lduba [%o1 + 0x00] %asi, %o3, add %o2, %g2)
add %o1, 0x1, %o1
add %o0, 0x1, %o0
subcc %g2, 0x1, %g2
bg,pt %icc, 1b
EXNV2(stba %o3, [%o0 + -1] %asi, add %o2, %g2)
2: andn %o2, 0x7, %g2
sub %o2, %g2, %o2
small_copy: /* 0 < len <= 16 */
andcc %o3, 0x3, %g0
bne,pn %XCC, small_copy_unaligned
sub %o0, %o1, %o3
3: EXNV3(ldxa [%o1 + 0x00] %asi, %o3, add %o2, %g2)
add %o1, 0x8, %o1
add %o0, 0x8, %o0
subcc %g2, 0x8, %g2
bg,pt %icc, 3b
EXNV3(stxa %o3, [%o0 + -8] %asi, add %o2, %g2)
small_copy_aligned:
subcc %o2, 4, %o2
EXNV4(lduwa [%o1] %asi, %g1, add %o2, %g0)
EXNV4(stwa %g1, [%o1 + %o3] ASI_AIUS, add %o2, %g0)
bgu,pt %XCC, small_copy_aligned
add %o1, 4, %o1
cmp %o2, 0
bne,pn %icc, U3copy_in_user_short
nop
ba,a,pt %xcc, U3copy_in_user_short_ret
out: retl
clr %o0
#endif /* !(SMALL_COPY_USES_FPU) */
.align 32
small_copy_unaligned:
subcc %o2, 1, %o2
EXNV1(lduba [%o1] %asi, %g1, add %o2, %g0)
EXNV1(stba %g1, [%o1 + %o3] ASI_AIUS, add %o2, %g0)
bgu,pt %XCC, small_copy_unaligned
add %o1, 1, %o1
retl
clr %o0
/* $Id: U3copy_to_user.S,v 1.3 2000/11/01 09:29:19 davem Exp $
* U3memcpy.S: UltraSparc-III optimized copy to userspace.
/* U3copy_to_user.S: UltraSparc-III optimized memcpy.
*
* Copyright (C) 1999, 2000 David S. Miller (davem@redhat.com)
* Copyright (C) 1999, 2000, 2004 David S. Miller (davem@redhat.com)
*/
#ifdef __KERNEL__
#include <asm/visasm.h>
#include <asm/asi.h>
#include <asm/dcu.h>
#include <asm/spitfire.h>
#undef SMALL_COPY_USES_FPU
#define XCC xcc
#define EXNV(x,y,a,b) \
98: x,y; \
.section .fixup; \
......@@ -34,6 +34,18 @@
.text; \
.align 4;
#define EXNV3(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
99: a, b, %o0; \
retl; \
add %o0, 4, %o0; \
.section __ex_table; \
.align 4; \
.word 98b, 99b; \
.text; \
.align 4;
#define EXNV4(x,y,a,b) \
98: x,y; \
.section .fixup; \
.align 4; \
......@@ -112,22 +124,9 @@
.word 98b, 99b; \
.text; \
.align 4;
#else
#define ASI_AIUS 0x80
#define ASI_BLK_AIUS 0xf0
#define FPRS_FEF 0x04
#define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs
#define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
#define SMALL_COPY_USES_FPU
#define EXNV(x,y,a,b) x,y;
#define EXNV2(x,y,a,b) x,y;
#define EXNV3(x,y,a,b) x,y;
#define EX(x,y,a,b) x,y;
#define EXBLK1(x,y) x,y;
#define EXBLK2(x,y) x,y;
#define EXBLK3(x,y) x,y;
#define EXBLK4(x,y) x,y;
#endif
.register %g2,#scratch
.register %g3,#scratch
/* Special/non-trivial issues of this code:
*
......@@ -148,89 +147,64 @@
* of up to 2.4GB per second.
*/
.globl U3copy_to_user
U3copy_to_user: /* %o0=dst, %o1=src, %o2=len */
.globl U3copy_to_user
U3copy_to_user: /* %o0=dst, %o1=src, %o2=len */
/* Writing to %asi is _expensive_ so we hardcode it.
* Reading %asi to check for KERNEL_DS is comparatively
* cheap.
*/
rd %asi, %g1 ! MS Group (4 cycles)
cmp %g1, ASI_AIUS ! A0 Group
bne U3memcpy ! BR
nop ! A1
#ifndef __KERNEL__
/* Save away original 'dst' for memcpy return value. */
mov %o0, %g3 ! A0 Group
#endif
/* Anything to copy at all? */
cmp %o2, 0 ! A1
ble,pn %icc, U3copy_to_user_short_ret ! BR
/* Extremely small copy? */
cmp %o2, 31 ! A0 Group
ble,pn %icc, U3copy_to_user_short ! BR
/* Large enough to use unrolled prefetch loops? */
cmp %o2, 0x100 ! A1
bge,a,pt %icc, U3copy_to_user_enter ! BR Group
andcc %o0, 0x3f, %g2 ! A0
ba,pt %xcc, U3copy_to_user_toosmall ! BR Group
andcc %o0, 0x7, %g2 ! A0
.align 32
U3copy_to_user_short:
/* Copy %o2 bytes from src to dst, one byte at a time. */
ldub [%o1 + 0x00], %o3 ! MS Group
add %o1, 0x1, %o1 ! A0
add %o0, 0x1, %o0 ! A1
subcc %o2, 1, %o2 ! A0 Group
bg,pt %icc, U3copy_to_user_short ! BR
EXNV(stba %o3, [%o0 + -1] %asi, add %o2, 1) ! MS Group (1-cycle stall)
U3copy_to_user_short_ret:
#ifdef __KERNEL__
retl ! BR Group (0-4 cycle stall)
clr %o0 ! A0
#else
retl ! BR Group (0-4 cycle stall)
mov %g3, %o0 ! A0
#endif
/* Here len >= (6 * 64) and condition codes reflect execution
rd %asi, %g1
cmp %g1, ASI_AIUS
bne,pn %icc, U3memcpy_user_stub
nop
cmp %o2, 0
be,pn %XCC, out
or %o0, %o1, %o3
cmp %o2, 16
bleu,a,pn %XCC, small_copy
or %o3, %o2, %o3
cmp %o2, 256
blu,pt %XCC, medium_copy
andcc %o3, 0x7, %g0
ba,pt %xcc, enter
andcc %o0, 0x3f, %g2
/* Here len >= 256 and condition codes reflect execution
* of "andcc %o0, 0x7, %g2", done by caller.
*/
.align 64
U3copy_to_user_enter:
enter:
/* Is 'dst' already aligned on an 64-byte boundary? */
be,pt %xcc, 2f ! BR
be,pt %XCC, 2f
/* Compute abs((dst & 0x3f) - 0x40) into %g2. This is the number
* of bytes to copy to make 'dst' 64-byte aligned. We pre-
* subtract this from 'len'.
*/
sub %g2, 0x40, %g2 ! A0 Group
sub %g0, %g2, %g2 ! A0 Group
sub %o2, %g2, %o2 ! A0 Group
sub %g2, 0x40, %g2
sub %g0, %g2, %g2
sub %o2, %g2, %o2
/* Copy %g2 bytes from src to dst, one byte at a time. */
1: ldub [%o1 + 0x00], %o3 ! MS (Group)
add %o1, 0x1, %o1 ! A1
add %o0, 0x1, %o0 ! A0 Group
subcc %g2, 0x1, %g2 ! A1
1: ldub [%o1 + 0x00], %o3
add %o1, 0x1, %o1
add %o0, 0x1, %o0
subcc %g2, 0x1, %g2
bg,pt %icc, 1b ! BR Group
EXNV2(stba %o3, [%o0 + -1] %asi, add %o2, %g2) ! MS Group
bg,pt %XCC, 1b
EXNV2(stba %o3, [%o0 + -1] %asi, add %o2, %g2)
2: VISEntryHalf ! MS+MS
and %o1, 0x7, %g1 ! A1
ba,pt %xcc, U3copy_to_user_begin ! BR
alignaddr %o1, %g0, %o1 ! MS (Break-after)
2: VISEntryHalf
and %o1, 0x7, %g1
ba,pt %xcc, begin
alignaddr %o1, %g0, %o1
.align 64
U3copy_to_user_begin:
#ifdef __KERNEL__
begin:
.globl U3copy_to_user_nop_1_6
U3copy_to_user_nop_1_6:
ldxa [%g0] ASI_DCU_CONTROL_REG, %g3
......@@ -239,309 +213,221 @@ U3copy_to_user_nop_1_6:
or %g3, %o3, %o3
stxa %o3, [%g0] ASI_DCU_CONTROL_REG ! Enable P-cache
membar #Sync
#endif
prefetch [%o1 + 0x000], #one_read ! MS Group1
prefetch [%o1 + 0x040], #one_read ! MS Group2
andn %o2, (0x40 - 1), %o4 ! A0
prefetch [%o1 + 0x080], #one_read ! MS Group3
cmp %o4, 0x140 ! A0
prefetch [%o1 + 0x0c0], #one_read ! MS Group4
ldd [%o1 + 0x000], %f0 ! MS Group5 (%f0 results at G8)
bge,a,pt %icc, 1f ! BR
prefetch [%o1 + 0x100], #one_read ! MS Group6
1: ldd [%o1 + 0x008], %f2 ! AX (%f2 results at G9)
cmp %o4, 0x180 ! A1
bge,a,pt %icc, 1f ! BR
prefetch [%o1 + 0x140], #one_read ! MS Group7
1: ldd [%o1 + 0x010], %f4 ! AX (%f4 results at G10)
cmp %o4, 0x1c0 ! A1
bge,a,pt %icc, 1f ! BR
prefetch [%o1 + 0x180], #one_read ! MS Group8
1: faligndata %f0, %f2, %f16 ! FGA Group9 (%f16 at G12)
ldd [%o1 + 0x018], %f6 ! AX (%f6 results at G12)
faligndata %f2, %f4, %f18 ! FGA Group10 (%f18 results at G13)
ldd [%o1 + 0x020], %f8 ! MS (%f8 results at G13)
faligndata %f4, %f6, %f20 ! FGA Group12 (1-cycle stall,%f20 at G15)
ldd [%o1 + 0x028], %f10 ! MS (%f10 results at G15)
faligndata %f6, %f8, %f22 ! FGA Group13 (%f22 results at G16)
ldd [%o1 + 0x030], %f12 ! MS (%f12 results at G16)
faligndata %f8, %f10, %f24 ! FGA Group15 (1-cycle stall,%f24 at G18)
ldd [%o1 + 0x038], %f14 ! MS (%f14 results at G18)
faligndata %f10, %f12, %f26 ! FGA Group16 (%f26 results at G19)
ldd [%o1 + 0x040], %f0 ! MS (%f0 results at G19)
/* We only use the first loop if len > (7 * 64). */
subcc %o4, 0x1c0, %o4 ! A0 Group17
bg,pt %icc, U3copy_to_user_loop1 ! BR
add %o1, 0x40, %o1 ! A1
add %o4, 0x140, %o4 ! A0 Group18
ba,pt %xcc, U3copy_to_user_loop2 ! BR
srl %o4, 6, %o3 ! A0 Group19
nop
nop
nop
nop
nop
nop
nop
/* This loop performs the copy and queues new prefetches.
* We drop into the second loop when len <= (5 * 64). Note
* that this (5 * 64) factor has been subtracted from len
* already.
*/
U3copy_to_user_loop1:
ldd [%o1 + 0x008], %f2 ! MS Group2 (%f2 results at G5)
faligndata %f12, %f14, %f28 ! FGA (%f28 results at G5)
ldd [%o1 + 0x010], %f4 ! MS Group3 (%f4 results at G6)
faligndata %f14, %f0, %f30 ! FGA Group4 (1-cycle stall, %f30 at G7)
EXBLK1(stda %f16, [%o0] ASI_BLK_AIUS) ! MS
ldd [%o1 + 0x018], %f6 ! AX (%f6 results at G7)
faligndata %f0, %f2, %f16 ! FGA Group12 (7-cycle stall)
ldd [%o1 + 0x020], %f8 ! MS (%f8 results at G15)
faligndata %f2, %f4, %f18 ! FGA Group13 (%f18 results at G16)
ldd [%o1 + 0x028], %f10 ! MS (%f10 results at G16)
faligndata %f4, %f6, %f20 ! FGA Group14 (%f20 results at G17)
ldd [%o1 + 0x030], %f12 ! MS (%f12 results at G17)
faligndata %f6, %f8, %f22 ! FGA Group15 (%f22 results at G18)
ldd [%o1 + 0x038], %f14 ! MS (%f14 results at G18)
faligndata %f8, %f10, %f24 ! FGA Group16 (%f24 results at G19)
ldd [%o1 + 0x040], %f0 ! AX (%f0 results at G19)
prefetch [%o1 + 0x180], #one_read ! MS
faligndata %f10, %f12, %f26 ! FGA Group17 (%f26 results at G20)
subcc %o4, 0x40, %o4 ! A0
add %o1, 0x40, %o1 ! A1
bg,pt %xcc, U3copy_to_user_loop1 ! BR
add %o0, 0x40, %o0 ! A0 Group18
U3copy_to_user_loop2_enter:
mov 5, %o3 ! A1
/* This loop performs on the copy, no new prefetches are
* queued. We do things this way so that we do not perform
* any spurious prefetches past the end of the src buffer.
*/
U3copy_to_user_loop2:
ldd [%o1 + 0x008], %f2 ! MS
faligndata %f12, %f14, %f28 ! FGA Group2
ldd [%o1 + 0x010], %f4 ! MS
faligndata %f14, %f0, %f30 ! FGA Group4 (1-cycle stall)
EXBLK2(stda %f16, [%o0] ASI_BLK_AIUS) ! MS
ldd [%o1 + 0x018], %f6 ! AX
faligndata %f0, %f2, %f16 ! FGA Group12 (7-cycle stall)
ldd [%o1 + 0x020], %f8 ! MS
faligndata %f2, %f4, %f18 ! FGA Group13
ldd [%o1 + 0x028], %f10 ! MS
faligndata %f4, %f6, %f20 ! FGA Group14
ldd [%o1 + 0x030], %f12 ! MS
faligndata %f6, %f8, %f22 ! FGA Group15
ldd [%o1 + 0x038], %f14 ! MS
faligndata %f8, %f10, %f24 ! FGA Group16
ldd [%o1 + 0x040], %f0 ! AX
faligndata %f10, %f12, %f26 ! FGA Group17
subcc %o3, 0x01, %o3 ! A0
add %o1, 0x40, %o1 ! A1
bg,pt %xcc, U3copy_to_user_loop2 ! BR
add %o0, 0x40, %o0 ! A0 Group18
prefetch [%o1 + 0x000], #one_read
prefetch [%o1 + 0x040], #one_read
andn %o2, (0x40 - 1), %o4
prefetch [%o1 + 0x080], #one_read
prefetch [%o1 + 0x0c0], #one_read
ldd [%o1 + 0x000], %f0
prefetch [%o1 + 0x100], #one_read
ldd [%o1 + 0x008], %f2
prefetch [%o1 + 0x140], #one_read
ldd [%o1 + 0x010], %f4
prefetch [%o1 + 0x180], #one_read
faligndata %f0, %f2, %f16
ldd [%o1 + 0x018], %f6
faligndata %f2, %f4, %f18
ldd [%o1 + 0x020], %f8
faligndata %f4, %f6, %f20
ldd [%o1 + 0x028], %f10
faligndata %f6, %f8, %f22
ldd [%o1 + 0x030], %f12
faligndata %f8, %f10, %f24
ldd [%o1 + 0x038], %f14
faligndata %f10, %f12, %f26
ldd [%o1 + 0x040], %f0
sub %o4, 0x80, %o4
add %o1, 0x40, %o1
ba,pt %xcc, loop
srl %o4, 6, %o3
.align 64
loop:
ldd [%o1 + 0x008], %f2
faligndata %f12, %f14, %f28
ldd [%o1 + 0x010], %f4
faligndata %f14, %f0, %f30
EXBLK2(stda %f16, [%o0] ASI_BLK_AIUS)
ldd [%o1 + 0x018], %f6
faligndata %f0, %f2, %f16
ldd [%o1 + 0x020], %f8
faligndata %f2, %f4, %f18
ldd [%o1 + 0x028], %f10
faligndata %f4, %f6, %f20
ldd [%o1 + 0x030], %f12
faligndata %f6, %f8, %f22
ldd [%o1 + 0x038], %f14
faligndata %f8, %f10, %f24
ldd [%o1 + 0x040], %f0
prefetch [%o1 + 0x180], #one_read
faligndata %f10, %f12, %f26
subcc %o3, 0x01, %o3
add %o1, 0x40, %o1
bg,pt %XCC, loop
add %o0, 0x40, %o0
/* Finally we copy the last full 64-byte block. */
U3copy_to_user_loopfini:
ldd [%o1 + 0x008], %f2 ! MS
faligndata %f12, %f14, %f28 ! FGA
ldd [%o1 + 0x010], %f4 ! MS Group19
faligndata %f14, %f0, %f30 ! FGA
EXBLK3(stda %f16, [%o0] ASI_BLK_AIUS) ! MS Group20
ldd [%o1 + 0x018], %f6 ! AX
faligndata %f0, %f2, %f16 ! FGA Group11 (7-cycle stall)
ldd [%o1 + 0x020], %f8 ! MS
faligndata %f2, %f4, %f18 ! FGA Group12
ldd [%o1 + 0x028], %f10 ! MS
faligndata %f4, %f6, %f20 ! FGA Group13
ldd [%o1 + 0x030], %f12 ! MS
faligndata %f6, %f8, %f22 ! FGA Group14
ldd [%o1 + 0x038], %f14 ! MS
faligndata %f8, %f10, %f24 ! FGA Group15
cmp %g1, 0 ! A0
be,pt %icc, 1f ! BR
add %o0, 0x40, %o0 ! A1
ldd [%o1 + 0x040], %f0 ! MS
1: faligndata %f10, %f12, %f26 ! FGA Group16
faligndata %f12, %f14, %f28 ! FGA Group17
faligndata %f14, %f0, %f30 ! FGA Group18
EXBLK4(stda %f16, [%o0] ASI_BLK_AIUS) ! MS
add %o0, 0x40, %o0 ! A0
add %o1, 0x40, %o1 ! A1
#ifdef __KERNEL__
loopfini:
ldd [%o1 + 0x008], %f2
faligndata %f12, %f14, %f28
ldd [%o1 + 0x010], %f4
faligndata %f14, %f0, %f30
EXBLK3(stda %f16, [%o0] ASI_BLK_AIUS)
ldd [%o1 + 0x018], %f6
faligndata %f0, %f2, %f16
ldd [%o1 + 0x020], %f8
faligndata %f2, %f4, %f18
ldd [%o1 + 0x028], %f10
faligndata %f4, %f6, %f20
ldd [%o1 + 0x030], %f12
faligndata %f6, %f8, %f22
ldd [%o1 + 0x038], %f14
faligndata %f8, %f10, %f24
cmp %g1, 0
be,pt %XCC, 1f
add %o0, 0x40, %o0
ldd [%o1 + 0x040], %f0
1: faligndata %f10, %f12, %f26
faligndata %f12, %f14, %f28
faligndata %f14, %f0, %f30
EXBLK4(stda %f16, [%o0] ASI_BLK_AIUS)
add %o0, 0x40, %o0
add %o1, 0x40, %o1
.globl U3copy_to_user_nop_2_3
U3copy_to_user_nop_2_3:
mov PRIMARY_CONTEXT, %o3
stxa %g0, [%o3] ASI_DMMU ! Flush P-cache
stxa %g3, [%g0] ASI_DCU_CONTROL_REG ! Disable P-cache
#endif
membar #Sync ! MS Group26 (7-cycle stall)
membar #Sync
/* Now we copy the (len modulo 64) bytes at the end.
* Note how we borrow the %f0 loaded above.
*
* Also notice how this code is careful not to perform a
* load past the end of the src buffer just like similar
* code found in U3copy_to_user_toosmall processing.
* load past the end of the src buffer.
*/
U3copy_to_user_loopend:
and %o2, 0x3f, %o2 ! A0 Group
andcc %o2, 0x38, %g2 ! A0 Group
be,pn %icc, U3copy_to_user_endcruft ! BR
subcc %g2, 0x8, %g2 ! A1
be,pn %icc, U3copy_to_user_endcruft ! BR Group
cmp %g1, 0 ! A0
be,a,pt %icc, 1f ! BR Group
ldd [%o1 + 0x00], %f0 ! MS
1: ldd [%o1 + 0x08], %f2 ! MS Group
add %o1, 0x8, %o1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f0, %f2, %f8 ! FGA Group
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8) ! MS (XXX does it stall here? XXX)
be,pn %icc, U3copy_to_user_endcruft ! BR
add %o0, 0x8, %o0 ! A0
ldd [%o1 + 0x08], %f0 ! MS Group
add %o1, 0x8, %o1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f2, %f0, %f8 ! FGA
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8) ! MS (XXX does it stall here? XXX)
bne,pn %icc, 1b ! BR
add %o0, 0x8, %o0 ! A0 Group
loopend:
and %o2, 0x3f, %o2
andcc %o2, 0x38, %g2
be,pn %XCC, endcruft
subcc %g2, 0x8, %g2
be,pn %XCC, endcruft
cmp %g1, 0
be,a,pt %XCC, 1f
ldd [%o1 + 0x00], %f0
1: ldd [%o1 + 0x08], %f2
add %o1, 0x8, %o1
sub %o2, 0x8, %o2
subcc %g2, 0x8, %g2
faligndata %f0, %f2, %f8
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8)
be,pn %XCC, endcruft
add %o0, 0x8, %o0
ldd [%o1 + 0x08], %f0
add %o1, 0x8, %o1
sub %o2, 0x8, %o2
subcc %g2, 0x8, %g2
faligndata %f2, %f0, %f8
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8)
bne,pn %XCC, 1b
add %o0, 0x8, %o0
/* If anything is left, we copy it one byte at a time.
* Note that %g1 is (src & 0x3) saved above before the
* alignaddr was performed.
*/
U3copy_to_user_endcruft:
endcruft:
cmp %o2, 0
add %o1, %g1, %o1
VISExitHalf
be,pn %icc, U3copy_to_user_short_ret
nop
ba,a,pt %xcc, U3copy_to_user_short
/* If we get here, then 32 <= len < (6 * 64) */
U3copy_to_user_toosmall:
#ifdef SMALL_COPY_USES_FPU
/* Is 'dst' already aligned on an 8-byte boundary? */
be,pt %xcc, 2f ! BR Group
/* Compute abs((dst & 7) - 8) into %g2. This is the number
* of bytes to copy to make 'dst' 8-byte aligned. We pre-
* subtract this from 'len'.
*/
sub %g2, 0x8, %g2 ! A0
sub %g0, %g2, %g2 ! A0 Group (reg-dep)
sub %o2, %g2, %o2 ! A0 Group (reg-dep)
/* Copy %g2 bytes from src to dst, one byte at a time. */
1: ldub [%o1 + 0x00], %o3 ! MS (Group) (%o3 in 3 cycles)
add %o1, 0x1, %o1 ! A1
add %o0, 0x1, %o0 ! A0 Group
subcc %g2, 0x1, %g2 ! A1
be,pn %XCC, out
sub %o0, %o1, %o3
bg,pt %icc, 1b ! BR Group
EXNV2(stba %o3, [%o0 + -1] %asi, add %o2, %g2) ! MS Group
andcc %g1, 0x7, %g0
bne,pn %icc, small_copy_unaligned
andcc %o2, 0x8, %g0
be,pt %icc, 1f
nop
ldx [%o1], %o5
EXNV(stxa %o5, [%o1 + %o3] ASI_AIUS, add %o2, %g0)
add %o1, 0x8, %o1
2: VISEntryHalf ! MS+MS
1: andcc %o2, 0x4, %g0
be,pt %icc, 1f
nop
lduw [%o1], %o5
EXNV(stwa %o5, [%o1 + %o3] ASI_AIUS, and %o2, 0x7)
add %o1, 0x4, %o1
/* Compute (len - (len % 8)) into %g2. This is guaranteed
* to be nonzero.
*/
andn %o2, 0x7, %g2 ! A0 Group
/* You may read this and believe that it allows reading
* one 8-byte longword past the end of src. It actually
* does not, as %g2 is subtracted as loads are done from
* src, so we always stop before running off the end.
* Also, we are guaranteed to have at least 0x10 bytes
* to move here.
*/
sub %g2, 0x8, %g2 ! A0 Group (reg-dep)
alignaddr %o1, %g0, %g1 ! MS (Break-after)
ldd [%g1 + 0x00], %f0 ! MS Group (1-cycle stall)
add %g1, 0x8, %g1 ! A0
1: ldd [%g1 + 0x00], %f2 ! MS Group
add %g1, 0x8, %g1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f0, %f2, %f8 ! FGA Group (1-cycle stall)
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8) ! MS Group (2-cycle stall)
add %o1, 0x8, %o1 ! A0
be,pn %icc, 2f ! BR
add %o0, 0x8, %o0 ! A1
ldd [%g1 + 0x00], %f0 ! MS Group
add %g1, 0x8, %g1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f2, %f0, %f8 ! FGA Group (1-cycle stall)
EX(stda %f8, [%o0 + 0x00] %asi, add %o2, 0x8) ! MS Group (2-cycle stall)
add %o1, 0x8, %o1 ! A0
bne,pn %icc, 1b ! BR
add %o0, 0x8, %o0 ! A1
/* Nothing left to copy? */
2: cmp %o2, 0 ! A0 Group
VISExitHalf ! A0+MS
be,pn %icc, U3copy_to_user_short_ret ! BR Group
nop ! A0
ba,a,pt %xcc, U3copy_to_user_short ! BR Group
#else /* !(SMALL_COPY_USES_FPU) */
xor %o1, %o0, %g2
andcc %g2, 0x7, %g0
bne,pn %icc, U3copy_to_user_short
andcc %o1, 0x7, %g2
be,pt %xcc, 2f
sub %g2, 0x8, %g2
sub %g0, %g2, %g2
sub %o2, %g2, %o2
1: andcc %o2, 0x2, %g0
be,pt %icc, 1f
nop
lduh [%o1], %o5
EXNV(stha %o5, [%o1 + %o3] ASI_AIUS, and %o2, 0x3)
add %o1, 0x2, %o1
1: ldub [%o1 + 0x00], %o3
add %o1, 0x1, %o1
add %o0, 0x1, %o0
subcc %g2, 0x1, %g2
bg,pt %icc, 1b
EXNV2(stba %o3, [%o0 + -1] %asi, add %o2, %g2)
1: andcc %o2, 0x1, %g0
be,pt %icc, out
nop
ldub [%o1], %o5
ba,pt %xcc, out
EXNV(stba %o5, [%o1 + %o3] ASI_AIUS, and %o2, 0x1)
medium_copy: /* 16 < len <= 64 */
bne,pn %XCC, small_copy_unaligned
sub %o0, %o1, %o3
medium_copy_aligned:
andn %o2, 0x7, %o4
and %o2, 0x7, %o2
1: subcc %o4, 0x8, %o4
ldx [%o1], %o5
EXNV4(stxa %o5, [%o1 + %o3] ASI_AIUS, add %o2, %o4)
bgu,pt %XCC, 1b
add %o1, 0x8, %o1
andcc %o2, 0x4, %g0
be,pt %XCC, 1f
nop
sub %o2, 0x4, %o2
lduw [%o1], %o5
EXNV3(stwa %o5, [%o1 + %o3] ASI_AIUS, add %o2, %g0)
add %o1, 0x4, %o1
1: cmp %o2, 0
be,pt %XCC, out
nop
ba,pt %xcc, small_copy_unaligned
nop
2: andn %o2, 0x7, %g2
sub %o2, %g2, %o2
small_copy: /* 0 < len <= 16 */
andcc %o3, 0x3, %g0
bne,pn %XCC, small_copy_unaligned
sub %o0, %o1, %o3
3: ldx [%o1 + 0x00], %o3
add %o1, 0x8, %o1
add %o0, 0x8, %o0
subcc %g2, 0x8, %g2
bg,pt %icc, 3b
EXNV3(stxa %o3, [%o0 + -8] %asi, add %o2, %g2)
small_copy_aligned:
subcc %o2, 4, %o2
lduw [%o1], %g1
EXNV3(stwa %g1, [%o1 + %o3] ASI_AIUS, add %o2, %g0)
bgu,pt %XCC, small_copy_aligned
add %o1, 4, %o1
cmp %o2, 0
bne,pn %icc, U3copy_to_user_short
nop
ba,a,pt %xcc, U3copy_to_user_short_ret
out: retl
clr %o0
#endif /* !(SMALL_COPY_USES_FPU) */
.align 32
small_copy_unaligned:
subcc %o2, 1, %o2
ldub [%o1], %g1
EXNV2(stba %g1, [%o1 + %o3] ASI_AIUS, add %o2, %g0)
bgu,pt %XCC, small_copy_unaligned
add %o1, 1, %o1
retl
clr %o0
/* $Id: U3memcpy.S,v 1.2 2000/11/01 09:29:19 davem Exp $
* U3memcpy.S: UltraSparc-III optimized memcpy.
/* U3memcpy.S: UltraSparc-III optimized memcpy.
*
* Copyright (C) 1999, 2000 David S. Miller (davem@redhat.com)
* Copyright (C) 1999, 2000, 2004 David S. Miller (davem@redhat.com)
*/
#ifdef __KERNEL__
......@@ -9,15 +8,20 @@
#include <asm/asi.h>
#include <asm/dcu.h>
#include <asm/spitfire.h>
#undef SMALL_COPY_USES_FPU
#else
#define ASI_BLK_P 0xf0
#define FPRS_FEF 0x04
#define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs
#define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
#define SMALL_COPY_USES_FPU
#endif
#ifndef XCC
#define XCC xcc
#endif
.register %g2,#scratch
.register %g3,#scratch
/* Special/non-trivial issues of this code:
*
* 1) %o5 is preserved from VISEntryHalf to VISExitHalf
......@@ -37,80 +41,55 @@
* of up to 2.4GB per second.
*/
.globl U3memcpy
U3memcpy: /* %o0=dst, %o1=src, %o2=len */
#ifndef __KERNEL__
/* Save away original 'dst' for memcpy return value. */
mov %o0, %g3 ! A0 Group
#endif
/* Anything to copy at all? */
cmp %o2, 0 ! A1
ble,pn %icc, U3memcpy_short_ret ! BR
/* Extremely small copy? */
cmp %o2, 31 ! A0 Group
ble,pn %icc, U3memcpy_short ! BR
/* Large enough to use unrolled prefetch loops? */
cmp %o2, 0x100 ! A1
bge,a,pt %icc, U3memcpy_enter ! BR Group
andcc %o0, 0x3f, %g2 ! A0
ba,pt %xcc, U3memcpy_toosmall ! BR Group
andcc %o0, 0x7, %g2 ! A0
.align 32
U3memcpy_short:
/* Copy %o2 bytes from src to dst, one byte at a time. */
ldub [%o1 + 0x00], %o3 ! MS Group
add %o1, 0x1, %o1 ! A0
add %o0, 0x1, %o0 ! A1
subcc %o2, 1, %o2 ! A0 Group
bg,pt %icc, U3memcpy_short ! BR
stb %o3, [%o0 + -1] ! MS Group (1-cycle stall)
U3memcpy_short_ret:
#ifdef __KERNEL__
retl ! BR Group (0-4 cycle stall)
clr %o0 ! A0
#else
retl ! BR Group (0-4 cycle stall)
mov %g3, %o0 ! A0
#endif
.globl U3memcpy
U3memcpy: /* %o0=dst, %o1=src, %o2=len */
mov %o0, %g5
cmp %o2, 0
be,pn %XCC, out
or %o0, %o1, %o3
cmp %o2, 16
bleu,a,pn %XCC, small_copy
or %o3, %o2, %o3
/* Here len >= (6 * 64) and condition codes reflect execution
cmp %o2, 256
blu,pt %XCC, medium_copy
andcc %o3, 0x7, %g0
ba,pt %xcc, enter
andcc %o0, 0x3f, %g2
/* Here len >= 256 and condition codes reflect execution
* of "andcc %o0, 0x7, %g2", done by caller.
*/
.align 64
U3memcpy_enter:
enter:
/* Is 'dst' already aligned on an 64-byte boundary? */
be,pt %xcc, 2f ! BR
be,pt %XCC, 2f
/* Compute abs((dst & 0x3f) - 0x40) into %g2. This is the number
* of bytes to copy to make 'dst' 64-byte aligned. We pre-
* subtract this from 'len'.
*/
sub %g2, 0x40, %g2 ! A0 Group
sub %g0, %g2, %g2 ! A0 Group
sub %o2, %g2, %o2 ! A0 Group
sub %g2, 0x40, %g2
sub %g0, %g2, %g2
sub %o2, %g2, %o2
/* Copy %g2 bytes from src to dst, one byte at a time. */
1: ldub [%o1 + 0x00], %o3 ! MS (Group)
add %o1, 0x1, %o1 ! A1
add %o0, 0x1, %o0 ! A0 Group
subcc %g2, 0x1, %g2 ! A1
1: ldub [%o1 + 0x00], %o3
add %o1, 0x1, %o1
add %o0, 0x1, %o0
subcc %g2, 0x1, %g2
bg,pt %icc, 1b ! BR Group
stb %o3, [%o0 + -1] ! MS Group
bg,pt %XCC, 1b
stb %o3, [%o0 + -1]
2: VISEntryHalf ! MS+MS
and %o1, 0x7, %g1 ! A1
ba,pt %xcc, U3memcpy_begin ! BR
alignaddr %o1, %g0, %o1 ! MS (Break-after)
2: VISEntryHalf
and %o1, 0x7, %g1
ba,pt %xcc, begin
alignaddr %o1, %g0, %o1
.align 64
U3memcpy_begin:
begin:
#ifdef __KERNEL__
.globl U3memcpy_nop_1_6
U3memcpy_nop_1_6:
......@@ -121,146 +100,90 @@ U3memcpy_nop_1_6:
stxa %o3, [%g0] ASI_DCU_CONTROL_REG ! Enable P-cache
membar #Sync
#endif
prefetch [%o1 + 0x000], #one_read ! MS Group1
prefetch [%o1 + 0x040], #one_read ! MS Group2
andn %o2, (0x40 - 1), %o4 ! A0
prefetch [%o1 + 0x080], #one_read ! MS Group3
cmp %o4, 0x140 ! A0
prefetch [%o1 + 0x0c0], #one_read ! MS Group4
ldd [%o1 + 0x000], %f0 ! MS Group5 (%f0 results at G8)
bge,a,pt %icc, 1f ! BR
prefetch [%o1 + 0x100], #one_read ! MS Group6
1: ldd [%o1 + 0x008], %f2 ! AX (%f2 results at G9)
cmp %o4, 0x180 ! A1
bge,a,pt %icc, 1f ! BR
prefetch [%o1 + 0x140], #one_read ! MS Group7
1: ldd [%o1 + 0x010], %f4 ! AX (%f4 results at G10)
cmp %o4, 0x1c0 ! A1
bge,a,pt %icc, 1f ! BR
prefetch [%o1 + 0x180], #one_read ! MS Group8
1: faligndata %f0, %f2, %f16 ! FGA Group9 (%f16 at G12)
ldd [%o1 + 0x018], %f6 ! AX (%f6 results at G12)
faligndata %f2, %f4, %f18 ! FGA Group10 (%f18 results at G13)
ldd [%o1 + 0x020], %f8 ! MS (%f8 results at G13)
faligndata %f4, %f6, %f20 ! FGA Group12 (1-cycle stall,%f20 at G15)
ldd [%o1 + 0x028], %f10 ! MS (%f10 results at G15)
faligndata %f6, %f8, %f22 ! FGA Group13 (%f22 results at G16)
ldd [%o1 + 0x030], %f12 ! MS (%f12 results at G16)
faligndata %f8, %f10, %f24 ! FGA Group15 (1-cycle stall,%f24 at G18)
ldd [%o1 + 0x038], %f14 ! MS (%f14 results at G18)
faligndata %f10, %f12, %f26 ! FGA Group16 (%f26 results at G19)
ldd [%o1 + 0x040], %f0 ! MS (%f0 results at G19)
/* We only use the first loop if len > (7 * 64). */
subcc %o4, 0x1c0, %o4 ! A0 Group17
bg,pt %icc, U3memcpy_loop1 ! BR
add %o1, 0x40, %o1 ! A1
add %o4, 0x140, %o4 ! A0 Group18
ba,pt %xcc, U3memcpy_loop2 ! BR
srl %o4, 6, %o3 ! A0 Group19
nop
nop
nop
nop
nop
nop
nop
/* This loop performs the copy and queues new prefetches.
* We drop into the second loop when len <= (5 * 64). Note
* that this (5 * 64) factor has been subtracted from len
* already.
*/
U3memcpy_loop1:
ldd [%o1 + 0x008], %f2 ! MS Group2 (%f2 results at G5)
faligndata %f12, %f14, %f28 ! FGA (%f28 results at G5)
ldd [%o1 + 0x010], %f4 ! MS Group3 (%f4 results at G6)
faligndata %f14, %f0, %f30 ! FGA Group4 (1-cycle stall, %f30 at G7)
stda %f16, [%o0] ASI_BLK_P ! MS
ldd [%o1 + 0x018], %f6 ! AX (%f6 results at G7)
faligndata %f0, %f2, %f16 ! FGA Group12 (7-cycle stall)
ldd [%o1 + 0x020], %f8 ! MS (%f8 results at G15)
faligndata %f2, %f4, %f18 ! FGA Group13 (%f18 results at G16)
ldd [%o1 + 0x028], %f10 ! MS (%f10 results at G16)
faligndata %f4, %f6, %f20 ! FGA Group14 (%f20 results at G17)
ldd [%o1 + 0x030], %f12 ! MS (%f12 results at G17)
faligndata %f6, %f8, %f22 ! FGA Group15 (%f22 results at G18)
ldd [%o1 + 0x038], %f14 ! MS (%f14 results at G18)
faligndata %f8, %f10, %f24 ! FGA Group16 (%f24 results at G19)
ldd [%o1 + 0x040], %f0 ! AX (%f0 results at G19)
prefetch [%o1 + 0x180], #one_read ! MS
faligndata %f10, %f12, %f26 ! FGA Group17 (%f26 results at G20)
subcc %o4, 0x40, %o4 ! A0
add %o1, 0x40, %o1 ! A1
bg,pt %xcc, U3memcpy_loop1 ! BR
add %o0, 0x40, %o0 ! A0 Group18
U3memcpy_loop2_enter:
mov 5, %o3 ! A1
/* This loop performs on the copy, no new prefetches are
* queued. We do things this way so that we do not perform
* any spurious prefetches past the end of the src buffer.
*/
U3memcpy_loop2:
ldd [%o1 + 0x008], %f2 ! MS
faligndata %f12, %f14, %f28 ! FGA Group2
ldd [%o1 + 0x010], %f4 ! MS
faligndata %f14, %f0, %f30 ! FGA Group4 (1-cycle stall)
stda %f16, [%o0] ASI_BLK_P ! MS
ldd [%o1 + 0x018], %f6 ! AX
faligndata %f0, %f2, %f16 ! FGA Group12 (7-cycle stall)
ldd [%o1 + 0x020], %f8 ! MS
faligndata %f2, %f4, %f18 ! FGA Group13
ldd [%o1 + 0x028], %f10 ! MS
faligndata %f4, %f6, %f20 ! FGA Group14
ldd [%o1 + 0x030], %f12 ! MS
faligndata %f6, %f8, %f22 ! FGA Group15
ldd [%o1 + 0x038], %f14 ! MS
faligndata %f8, %f10, %f24 ! FGA Group16
ldd [%o1 + 0x040], %f0 ! AX
faligndata %f10, %f12, %f26 ! FGA Group17
subcc %o3, 0x01, %o3 ! A0
add %o1, 0x40, %o1 ! A1
bg,pt %xcc, U3memcpy_loop2 ! BR
add %o0, 0x40, %o0 ! A0 Group18
prefetch [%o1 + 0x000], #one_read
prefetch [%o1 + 0x040], #one_read
andn %o2, (0x40 - 1), %o4
prefetch [%o1 + 0x080], #one_read
prefetch [%o1 + 0x0c0], #one_read
ldd [%o1 + 0x000], %f0
prefetch [%o1 + 0x100], #one_read
ldd [%o1 + 0x008], %f2
prefetch [%o1 + 0x140], #one_read
ldd [%o1 + 0x010], %f4
prefetch [%o1 + 0x180], #one_read
faligndata %f0, %f2, %f16
ldd [%o1 + 0x018], %f6
faligndata %f2, %f4, %f18
ldd [%o1 + 0x020], %f8
faligndata %f4, %f6, %f20
ldd [%o1 + 0x028], %f10
faligndata %f6, %f8, %f22
ldd [%o1 + 0x030], %f12
faligndata %f8, %f10, %f24
ldd [%o1 + 0x038], %f14
faligndata %f10, %f12, %f26
ldd [%o1 + 0x040], %f0
sub %o4, 0x80, %o4
add %o1, 0x40, %o1
ba,pt %xcc, loop
srl %o4, 6, %o3
.align 64
loop:
ldd [%o1 + 0x008], %f2
faligndata %f12, %f14, %f28
ldd [%o1 + 0x010], %f4
faligndata %f14, %f0, %f30
stda %f16, [%o0] ASI_BLK_P
ldd [%o1 + 0x018], %f6
faligndata %f0, %f2, %f16
ldd [%o1 + 0x020], %f8
faligndata %f2, %f4, %f18
ldd [%o1 + 0x028], %f10
faligndata %f4, %f6, %f20
ldd [%o1 + 0x030], %f12
faligndata %f6, %f8, %f22
ldd [%o1 + 0x038], %f14
faligndata %f8, %f10, %f24
ldd [%o1 + 0x040], %f0
prefetch [%o1 + 0x180], #one_read
faligndata %f10, %f12, %f26
subcc %o3, 0x01, %o3
add %o1, 0x40, %o1
bg,pt %XCC, loop
add %o0, 0x40, %o0
/* Finally we copy the last full 64-byte block. */
U3memcpy_loopfini:
ldd [%o1 + 0x008], %f2 ! MS
faligndata %f12, %f14, %f28 ! FGA
ldd [%o1 + 0x010], %f4 ! MS Group19
faligndata %f14, %f0, %f30 ! FGA
stda %f16, [%o0] ASI_BLK_P ! MS Group20
ldd [%o1 + 0x018], %f6 ! AX
faligndata %f0, %f2, %f16 ! FGA Group11 (7-cycle stall)
ldd [%o1 + 0x020], %f8 ! MS
faligndata %f2, %f4, %f18 ! FGA Group12
ldd [%o1 + 0x028], %f10 ! MS
faligndata %f4, %f6, %f20 ! FGA Group13
ldd [%o1 + 0x030], %f12 ! MS
faligndata %f6, %f8, %f22 ! FGA Group14
ldd [%o1 + 0x038], %f14 ! MS
faligndata %f8, %f10, %f24 ! FGA Group15
cmp %g1, 0 ! A0
be,pt %icc, 1f ! BR
add %o0, 0x40, %o0 ! A1
ldd [%o1 + 0x040], %f0 ! MS
1: faligndata %f10, %f12, %f26 ! FGA Group16
faligndata %f12, %f14, %f28 ! FGA Group17
faligndata %f14, %f0, %f30 ! FGA Group18
stda %f16, [%o0] ASI_BLK_P ! MS
add %o0, 0x40, %o0 ! A0
add %o1, 0x40, %o1 ! A1
loopfini:
ldd [%o1 + 0x008], %f2
faligndata %f12, %f14, %f28
ldd [%o1 + 0x010], %f4
faligndata %f14, %f0, %f30
stda %f16, [%o0] ASI_BLK_P
ldd [%o1 + 0x018], %f6
faligndata %f0, %f2, %f16
ldd [%o1 + 0x020], %f8
faligndata %f2, %f4, %f18
ldd [%o1 + 0x028], %f10
faligndata %f4, %f6, %f20
ldd [%o1 + 0x030], %f12
faligndata %f6, %f8, %f22
ldd [%o1 + 0x038], %f14
faligndata %f8, %f10, %f24
cmp %g1, 0
be,pt %XCC, 1f
add %o0, 0x40, %o0
ldd [%o1 + 0x040], %f0
1: faligndata %f10, %f12, %f26
faligndata %f12, %f14, %f28
faligndata %f14, %f0, %f30
stda %f16, [%o0] ASI_BLK_P
add %o0, 0x40, %o0
add %o1, 0x40, %o1
#ifdef __KERNEL__
.globl U3memcpy_nop_2_3
U3memcpy_nop_2_3:
......@@ -268,161 +191,143 @@ U3memcpy_nop_2_3:
stxa %g0, [%o3] ASI_DMMU ! Flush P-cache
stxa %g3, [%g0] ASI_DCU_CONTROL_REG ! Disable P-cache
#endif
membar #Sync ! MS Group26 (7-cycle stall)
membar #Sync
/* Now we copy the (len modulo 64) bytes at the end.
* Note how we borrow the %f0 loaded above.
*
* Also notice how this code is careful not to perform a
* load past the end of the src buffer just like similar
* code found in U3memcpy_toosmall processing.
* load past the end of the src buffer.
*/
U3memcpy_loopend:
and %o2, 0x3f, %o2 ! A0 Group
andcc %o2, 0x38, %g2 ! A0 Group
be,pn %icc, U3memcpy_endcruft ! BR
subcc %g2, 0x8, %g2 ! A1
be,pn %icc, U3memcpy_endcruft ! BR Group
cmp %g1, 0 ! A0
be,a,pt %icc, 1f ! BR Group
ldd [%o1 + 0x00], %f0 ! MS
1: ldd [%o1 + 0x08], %f2 ! MS Group
add %o1, 0x8, %o1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f0, %f2, %f8 ! FGA Group
std %f8, [%o0 + 0x00] ! MS (XXX does it stall here? XXX)
be,pn %icc, U3memcpy_endcruft ! BR
add %o0, 0x8, %o0 ! A0
ldd [%o1 + 0x08], %f0 ! MS Group
add %o1, 0x8, %o1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f2, %f0, %f8 ! FGA
std %f8, [%o0 + 0x00] ! MS (XXX does it stall here? XXX)
bne,pn %icc, 1b ! BR
add %o0, 0x8, %o0 ! A0 Group
loopend:
and %o2, 0x3f, %o2
andcc %o2, 0x38, %g2
be,pn %XCC, endcruft
subcc %g2, 0x8, %g2
be,pn %XCC, endcruft
cmp %g1, 0
be,a,pt %XCC, 1f
ldd [%o1 + 0x00], %f0
1: ldd [%o1 + 0x08], %f2
add %o1, 0x8, %o1
sub %o2, 0x8, %o2
subcc %g2, 0x8, %g2
faligndata %f0, %f2, %f8
std %f8, [%o0 + 0x00]
be,pn %XCC, endcruft
add %o0, 0x8, %o0
ldd [%o1 + 0x08], %f0
add %o1, 0x8, %o1
sub %o2, 0x8, %o2
subcc %g2, 0x8, %g2
faligndata %f2, %f0, %f8
std %f8, [%o0 + 0x00]
bne,pn %XCC, 1b
add %o0, 0x8, %o0
/* If anything is left, we copy it one byte at a time.
* Note that %g1 is (src & 0x3) saved above before the
* alignaddr was performed.
*/
U3memcpy_endcruft:
endcruft:
cmp %o2, 0
add %o1, %g1, %o1
VISExitHalf
be,pn %icc, U3memcpy_short_ret
nop
ba,a,pt %xcc, U3memcpy_short
be,pn %XCC, out
sub %o0, %o1, %o3
/* If we get here, then 32 <= len < (6 * 64) */
U3memcpy_toosmall:
andcc %g1, 0x7, %g0
bne,pn %icc, small_copy_unaligned
andcc %o2, 0x8, %g0
be,pt %icc, 1f
nop
ldx [%o1], %o5
stx %o5, [%o1 + %o3]
add %o1, 0x8, %o1
#ifdef SMALL_COPY_USES_FPU
1: andcc %o2, 0x4, %g0
be,pt %icc, 1f
nop
lduw [%o1], %o5
stw %o5, [%o1 + %o3]
add %o1, 0x4, %o1
/* Is 'dst' already aligned on an 8-byte boundary? */
be,pt %xcc, 2f ! BR Group
1: andcc %o2, 0x2, %g0
be,pt %icc, 1f
nop
lduh [%o1], %o5
sth %o5, [%o1 + %o3]
add %o1, 0x2, %o1
/* Compute abs((dst & 7) - 8) into %g2. This is the number
* of bytes to copy to make 'dst' 8-byte aligned. We pre-
* subtract this from 'len'.
*/
sub %g2, 0x8, %g2 ! A0
sub %g0, %g2, %g2 ! A0 Group (reg-dep)
sub %o2, %g2, %o2 ! A0 Group (reg-dep)
1: andcc %o2, 0x1, %g0
be,pt %icc, out
nop
ldub [%o1], %o5
ba,pt %xcc, out
stb %o5, [%o1 + %o3]
medium_copy: /* 16 < len <= 64 */
bne,pn %XCC, small_copy_unaligned
sub %o0, %o1, %o3
medium_copy_aligned:
andn %o2, 0x7, %o4
and %o2, 0x7, %o2
1: subcc %o4, 0x8, %o4
ldx [%o1], %o5
stx %o5, [%o1 + %o3]
bgu,pt %XCC, 1b
add %o1, 0x8, %o1
andcc %o2, 0x4, %g0
be,pt %XCC, 1f
nop
sub %o2, 0x4, %o2
lduw [%o1], %o5
stw %o5, [%o1 + %o3]
add %o1, 0x4, %o1
1: cmp %o2, 0
be,pt %XCC, out
nop
ba,pt %xcc, small_copy_unaligned
nop
/* Copy %g2 bytes from src to dst, one byte at a time. */
1: ldub [%o1 + 0x00], %o3 ! MS (Group) (%o3 in 3 cycles)
add %o1, 0x1, %o1 ! A1
add %o0, 0x1, %o0 ! A0 Group
subcc %g2, 0x1, %g2 ! A1
small_copy: /* 0 < len <= 16 */
andcc %o3, 0x3, %g0
bne,pn %XCC, small_copy_unaligned
sub %o0, %o1, %o3
bg,pt %icc, 1b ! BR Group
stb %o3, [%o0 + -1] ! MS Group
small_copy_aligned:
subcc %o2, 4, %o2
lduw [%o1], %g1
stw %g1, [%o1 + %o3]
bgu,pt %XCC, small_copy_aligned
add %o1, 4, %o1
2: VISEntryHalf ! MS+MS
out: retl
mov %g5, %o0
/* Compute (len - (len % 8)) into %g2. This is guaranteed
* to be nonzero.
*/
andn %o2, 0x7, %g2 ! A0 Group
/* You may read this and believe that it allows reading
* one 8-byte longword past the end of src. It actually
* does not, as %g2 is subtracted as loads are done from
* src, so we always stop before running off the end.
* Also, we are guaranteed to have at least 0x10 bytes
* to move here.
.align 32
small_copy_unaligned:
subcc %o2, 1, %o2
ldub [%o1], %g1
stb %g1, [%o1 + %o3]
bgu,pt %XCC, small_copy_unaligned
add %o1, 1, %o1
retl
mov %g5, %o0
/* Act like copy_{to,in}_user(), ie. return zero instead
* of original destination pointer. This is invoked when
* copy_{to,in}_user() finds that %asi is kernel space.
*/
sub %g2, 0x8, %g2 ! A0 Group (reg-dep)
alignaddr %o1, %g0, %g1 ! MS (Break-after)
ldd [%g1 + 0x00], %f0 ! MS Group (1-cycle stall)
add %g1, 0x8, %g1 ! A0
1: ldd [%g1 + 0x00], %f2 ! MS Group
add %g1, 0x8, %g1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f0, %f2, %f8 ! FGA Group (1-cycle stall)
std %f8, [%o0 + 0x00] ! MS Group (2-cycle stall)
add %o1, 0x8, %o1 ! A0
be,pn %icc, 2f ! BR
add %o0, 0x8, %o0 ! A1
ldd [%g1 + 0x00], %f0 ! MS Group
add %g1, 0x8, %g1 ! A0
sub %o2, 0x8, %o2 ! A1
subcc %g2, 0x8, %g2 ! A0 Group
faligndata %f2, %f0, %f8 ! FGA Group (1-cycle stall)
std %f8, [%o0 + 0x00] ! MS Group (2-cycle stall)
add %o1, 0x8, %o1 ! A0
bne,pn %icc, 1b ! BR
add %o0, 0x8, %o0 ! A1
/* Nothing left to copy? */
2: cmp %o2, 0 ! A0 Group
VISExitHalf ! A0+MS
be,pn %icc, U3memcpy_short_ret ! BR Group
nop ! A0
ba,a,pt %xcc, U3memcpy_short ! BR Group
#else /* !(SMALL_COPY_USES_FPU) */
xor %o1, %o0, %g2
andcc %g2, 0x7, %g0
bne,pn %icc, U3memcpy_short
andcc %o1, 0x7, %g2
be,pt %xcc, 2f
sub %g2, 0x8, %g2
sub %g0, %g2, %g2
sub %o2, %g2, %o2
1: ldub [%o1 + 0x00], %o3
add %o1, 0x1, %o1
add %o0, 0x1, %o0
subcc %g2, 0x1, %g2
bg,pt %icc, 1b
stb %o3, [%o0 + -1]
2: andn %o2, 0x7, %g2
sub %o2, %g2, %o2
3: ldx [%o1 + 0x00], %o3
add %o1, 0x8, %o1
add %o0, 0x8, %o0
subcc %g2, 0x8, %g2
bg,pt %icc, 3b
stx %o3, [%o0 + -8]
cmp %o2, 0
bne,pn %icc, U3memcpy_short
nop
ba,a,pt %xcc, U3memcpy_short_ret
#endif /* !(SMALL_COPY_USES_FPU) */
.globl U3memcpy_user_stub
U3memcpy_user_stub:
save %sp, -192, %sp
mov %i0, %o0
mov %i1, %o1
call U3memcpy
mov %i2, %o2
ret
restore %g0, %g0, %o0
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment