404 lines
10 KiB
ArmAsm
404 lines
10 KiB
ArmAsm
/* Copyright (C) 2003-2007 Analog Devices Inc.
|
|
*
|
|
* This file is subject to the terms and conditions of the GNU General Public
|
|
* License.
|
|
*/
|
|
|
|
#define ASSEMBLY
|
|
|
|
#include <asm/linkage.h>
|
|
#include <asm/cplb.h>
|
|
#include <config.h>
|
|
#include <asm/blackfin.h>
|
|
|
|
.text
|
|
|
|
/* This is an external function being called by the user
|
|
* application through __flush_cache_all. Currently this function
|
|
* serves the purpose of flushing all the pending writes in
|
|
* in the instruction cache.
|
|
*/
|
|
|
|
ENTRY(_flush_instruction_cache)
|
|
[--SP] = ( R7:6, P5:4 );
|
|
LINK 12;
|
|
SP += -12;
|
|
P5.H = (ICPLB_ADDR0 >> 16);
|
|
P5.L = (ICPLB_ADDR0 & 0xFFFF);
|
|
P4.H = (ICPLB_DATA0 >> 16);
|
|
P4.L = (ICPLB_DATA0 & 0xFFFF);
|
|
R7 = CPLB_VALID | CPLB_L1_CHBL;
|
|
R6 = 16;
|
|
inext: R0 = [P5++];
|
|
R1 = [P4++];
|
|
[--SP] = RETS;
|
|
CALL _icplb_flush; /* R0 = page, R1 = data*/
|
|
RETS = [SP++];
|
|
iskip: R6 += -1;
|
|
CC = R6;
|
|
IF CC JUMP inext;
|
|
SSYNC;
|
|
SP += 12;
|
|
UNLINK;
|
|
( R7:6, P5:4 ) = [SP++];
|
|
RTS;
|
|
|
|
/* This is an internal function to flush all pending
|
|
* writes in the cache associated with a particular ICPLB.
|
|
*
|
|
* R0 - page's start address
|
|
* R1 - CPLB's data field.
|
|
*/
|
|
|
|
.align 2
|
|
ENTRY(_icplb_flush)
|
|
[--SP] = ( R7:0, P5:0 );
|
|
[--SP] = LC0;
|
|
[--SP] = LT0;
|
|
[--SP] = LB0;
|
|
[--SP] = LC1;
|
|
[--SP] = LT1;
|
|
[--SP] = LB1;
|
|
|
|
/* If it's a 1K or 4K page, then it's quickest to
|
|
* just systematically flush all the addresses in
|
|
* the page, regardless of whether they're in the
|
|
* cache, or dirty. If it's a 1M or 4M page, there
|
|
* are too many addresses, and we have to search the
|
|
* cache for lines corresponding to the page.
|
|
*/
|
|
|
|
CC = BITTST(R1, 17); /* 1MB or 4MB */
|
|
IF !CC JUMP iflush_whole_page;
|
|
|
|
/* We're only interested in the page's size, so extract
|
|
* this from the CPLB (bits 17:16), and scale to give an
|
|
* offset into the page_size and page_prefix tables.
|
|
*/
|
|
|
|
R1 <<= 14;
|
|
R1 >>= 30;
|
|
R1 <<= 2;
|
|
|
|
/* We can also determine the sub-bank used, because this is
|
|
* taken from bits 13:12 of the address.
|
|
*/
|
|
|
|
R3 = ((12<<8)|2); /* Extraction pattern */
|
|
nop; /* Anamoly 05000209 */
|
|
R4 = EXTRACT(R0, R3.L) (Z); /* Extract bits */
|
|
|
|
/* Save in extraction pattern for later deposit. */
|
|
R3.H = R4.L << 0;
|
|
|
|
/* So:
|
|
* R0 = Page start
|
|
* R1 = Page length (actually, offset into size/prefix tables)
|
|
* R3 = sub-bank deposit values
|
|
*
|
|
* The cache has 2 Ways, and 64 sets, so we iterate through
|
|
* the sets, accessing the tag for each Way, for our Bank and
|
|
* sub-bank, looking for dirty, valid tags that match our
|
|
* address prefix.
|
|
*/
|
|
|
|
P5.L = (ITEST_COMMAND & 0xFFFF);
|
|
P5.H = (ITEST_COMMAND >> 16);
|
|
P4.L = (ITEST_DATA0 & 0xFFFF);
|
|
P4.H = (ITEST_DATA0 >> 16);
|
|
|
|
P0.L = page_prefix_table;
|
|
P0.H = page_prefix_table;
|
|
P1 = R1;
|
|
R5 = 0; /* Set counter*/
|
|
P0 = P1 + P0;
|
|
R4 = [P0]; /* This is the address prefix*/
|
|
|
|
/* We're reading (bit 1==0) the tag (bit 2==0), and we
|
|
* don't care about which double-word, since we're only
|
|
* fetching tags, so we only have to set Set, Bank,
|
|
* Sub-bank and Way.
|
|
*/
|
|
|
|
P2 = 4;
|
|
LSETUP (ifs1, ife1) LC1 = P2;
|
|
ifs1: P0 = 32; /* iterate over all sets*/
|
|
LSETUP (ifs0, ife0) LC0 = P0;
|
|
ifs0: R6 = R5 << 5; /* Combine set*/
|
|
R6.H = R3.H << 0 ; /* and sub-bank*/
|
|
[P5] = R6; /* Issue Command*/
|
|
SSYNC; /* CSYNC will not work here :(*/
|
|
R7 = [P4]; /* and read Tag.*/
|
|
CC = BITTST(R7, 0); /* Check if valid*/
|
|
IF !CC JUMP ifskip; /* and skip if not.*/
|
|
|
|
/* Compare against the page address. First, plant bits 13:12
|
|
* into the tag, since those aren't part of the returned data.
|
|
*/
|
|
|
|
R7 = DEPOSIT(R7, R3); /* set 13:12*/
|
|
R1 = R7 & R4; /* Mask off lower bits*/
|
|
CC = R1 == R0; /* Compare against page start.*/
|
|
IF !CC JUMP ifskip; /* Skip it if it doesn't match.*/
|
|
|
|
/* Tag address matches against page, so this is an entry
|
|
* we must flush.
|
|
*/
|
|
|
|
R7 >>= 10; /* Mask off the non-address bits*/
|
|
R7 <<= 10;
|
|
P3 = R7;
|
|
IFLUSH [P3]; /* And flush the entry*/
|
|
ifskip:
|
|
ife0: R5 += 1; /* Advance to next Set*/
|
|
ife1: NOP;
|
|
|
|
ifinished:
|
|
SSYNC; /* Ensure the data gets out to mem.*/
|
|
|
|
/*Finished. Restore context.*/
|
|
LB1 = [SP++];
|
|
LT1 = [SP++];
|
|
LC1 = [SP++];
|
|
LB0 = [SP++];
|
|
LT0 = [SP++];
|
|
LC0 = [SP++];
|
|
( R7:0, P5:0 ) = [SP++];
|
|
RTS;
|
|
|
|
iflush_whole_page:
|
|
/* It's a 1K or 4K page, so quicker to just flush the
|
|
* entire page.
|
|
*/
|
|
|
|
P1 = 32; /* For 1K pages*/
|
|
P2 = P1 << 2; /* For 4K pages*/
|
|
P0 = R0; /* Start of page*/
|
|
CC = BITTST(R1, 16); /* Whether 1K or 4K*/
|
|
IF CC P1 = P2;
|
|
P1 += -1; /* Unroll one iteration*/
|
|
SSYNC;
|
|
IFLUSH [P0++]; /* because CSYNC can't end loops.*/
|
|
LSETUP (isall, ieall) LC0 = P1;
|
|
isall:IFLUSH [P0++];
|
|
ieall: NOP;
|
|
SSYNC;
|
|
JUMP ifinished;
|
|
|
|
/* This is an external function being called by the user
|
|
* application through __flush_cache_all. Currently this function
|
|
* serves the purpose of flushing all the pending writes in
|
|
* in the data cache.
|
|
*/
|
|
|
|
ENTRY(_flush_data_cache)
|
|
[--SP] = ( R7:6, P5:4 );
|
|
LINK 12;
|
|
SP += -12;
|
|
P5.H = (DCPLB_ADDR0 >> 16);
|
|
P5.L = (DCPLB_ADDR0 & 0xFFFF);
|
|
P4.H = (DCPLB_DATA0 >> 16);
|
|
P4.L = (DCPLB_DATA0 & 0xFFFF);
|
|
R7 = CPLB_VALID | CPLB_L1_CHBL | CPLB_DIRTY (Z);
|
|
R6 = 16;
|
|
next: R0 = [P5++];
|
|
R1 = [P4++];
|
|
CC = BITTST(R1, 14); /* Is it write-through?*/
|
|
IF CC JUMP skip; /* If so, ignore it.*/
|
|
R2 = R1 & R7; /* Is it a dirty, cached page?*/
|
|
CC = R2;
|
|
IF !CC JUMP skip; /* If not, ignore it.*/
|
|
[--SP] = RETS;
|
|
CALL _dcplb_flush; /* R0 = page, R1 = data*/
|
|
RETS = [SP++];
|
|
skip: R6 += -1;
|
|
CC = R6;
|
|
IF CC JUMP next;
|
|
SSYNC;
|
|
SP += 12;
|
|
UNLINK;
|
|
( R7:6, P5:4 ) = [SP++];
|
|
RTS;
|
|
|
|
/* This is an internal function to flush all pending
|
|
* writes in the cache associated with a particular DCPLB.
|
|
*
|
|
* R0 - page's start address
|
|
* R1 - CPLB's data field.
|
|
*/
|
|
|
|
.align 2
|
|
ENTRY(_dcplb_flush)
|
|
[--SP] = ( R7:0, P5:0 );
|
|
[--SP] = LC0;
|
|
[--SP] = LT0;
|
|
[--SP] = LB0;
|
|
[--SP] = LC1;
|
|
[--SP] = LT1;
|
|
[--SP] = LB1;
|
|
|
|
/* If it's a 1K or 4K page, then it's quickest to
|
|
* just systematically flush all the addresses in
|
|
* the page, regardless of whether they're in the
|
|
* cache, or dirty. If it's a 1M or 4M page, there
|
|
* are too many addresses, and we have to search the
|
|
* cache for lines corresponding to the page.
|
|
*/
|
|
|
|
CC = BITTST(R1, 17); /* 1MB or 4MB */
|
|
IF !CC JUMP dflush_whole_page;
|
|
|
|
/* We're only interested in the page's size, so extract
|
|
* this from the CPLB (bits 17:16), and scale to give an
|
|
* offset into the page_size and page_prefix tables.
|
|
*/
|
|
|
|
R1 <<= 14;
|
|
R1 >>= 30;
|
|
R1 <<= 2;
|
|
|
|
/* The page could be mapped into Bank A or Bank B, depending
|
|
* on (a) whether both banks are configured as cache, and
|
|
* (b) on whether address bit A[x] is set. x is determined
|
|
* by DCBS in DMEM_CONTROL
|
|
*/
|
|
|
|
R2 = 0; /* Default to Bank A (Bank B would be 1)*/
|
|
|
|
P0.L = (DMEM_CONTROL & 0xFFFF);
|
|
P0.H = (DMEM_CONTROL >> 16);
|
|
|
|
R3 = [P0]; /* If Bank B is not enabled as cache*/
|
|
CC = BITTST(R3, 2); /* then Bank A is our only option.*/
|
|
IF CC JUMP bank_chosen;
|
|
|
|
R4 = 1<<14; /* If DCBS==0, use A[14].*/
|
|
R5 = R4 << 7; /* If DCBS==1, use A[23];*/
|
|
CC = BITTST(R3, 4);
|
|
IF CC R4 = R5; /* R4 now has either bit 14 or bit 23 set.*/
|
|
R5 = R0 & R4; /* Use it to test the Page address*/
|
|
CC = R5; /* and if that bit is set, we use Bank B,*/
|
|
R2 = CC; /* else we use Bank A.*/
|
|
R2 <<= 23; /* The Bank selection's at posn 23.*/
|
|
|
|
bank_chosen:
|
|
|
|
/* We can also determine the sub-bank used, because this is
|
|
* taken from bits 13:12 of the address.
|
|
*/
|
|
|
|
R3 = ((12<<8)|2); /* Extraction pattern */
|
|
nop; /*Anamoly 05000209*/
|
|
R4 = EXTRACT(R0, R3.L) (Z); /* Extract bits*/
|
|
/* Save in extraction pattern for later deposit.*/
|
|
R3.H = R4.L << 0;
|
|
|
|
/* So:
|
|
* R0 = Page start
|
|
* R1 = Page length (actually, offset into size/prefix tables)
|
|
* R2 = Bank select mask
|
|
* R3 = sub-bank deposit values
|
|
*
|
|
* The cache has 2 Ways, and 64 sets, so we iterate through
|
|
* the sets, accessing the tag for each Way, for our Bank and
|
|
* sub-bank, looking for dirty, valid tags that match our
|
|
* address prefix.
|
|
*/
|
|
|
|
P5.L = (DTEST_COMMAND & 0xFFFF);
|
|
P5.H = (DTEST_COMMAND >> 16);
|
|
P4.L = (DTEST_DATA0 & 0xFFFF);
|
|
P4.H = (DTEST_DATA0 >> 16);
|
|
|
|
P0.L = page_prefix_table;
|
|
P0.H = page_prefix_table;
|
|
P1 = R1;
|
|
R5 = 0; /* Set counter*/
|
|
P0 = P1 + P0;
|
|
R4 = [P0]; /* This is the address prefix*/
|
|
|
|
|
|
/* We're reading (bit 1==0) the tag (bit 2==0), and we
|
|
* don't care about which double-word, since we're only
|
|
* fetching tags, so we only have to set Set, Bank,
|
|
* Sub-bank and Way.
|
|
*/
|
|
|
|
P2 = 2;
|
|
LSETUP (fs1, fe1) LC1 = P2;
|
|
fs1: P0 = 64; /* iterate over all sets*/
|
|
LSETUP (fs0, fe0) LC0 = P0;
|
|
fs0: R6 = R5 << 5; /* Combine set*/
|
|
R6.H = R3.H << 0 ; /* and sub-bank*/
|
|
R6 = R6 | R2; /* and Bank. Leave Way==0 at first.*/
|
|
BITSET(R6,14);
|
|
[P5] = R6; /* Issue Command*/
|
|
SSYNC;
|
|
R7 = [P4]; /* and read Tag.*/
|
|
CC = BITTST(R7, 0); /* Check if valid*/
|
|
IF !CC JUMP fskip; /* and skip if not.*/
|
|
CC = BITTST(R7, 1); /* Check if dirty*/
|
|
IF !CC JUMP fskip; /* and skip if not.*/
|
|
|
|
/* Compare against the page address. First, plant bits 13:12
|
|
* into the tag, since those aren't part of the returned data.
|
|
*/
|
|
|
|
R7 = DEPOSIT(R7, R3); /* set 13:12*/
|
|
R1 = R7 & R4; /* Mask off lower bits*/
|
|
CC = R1 == R0; /* Compare against page start.*/
|
|
IF !CC JUMP fskip; /* Skip it if it doesn't match.*/
|
|
|
|
/* Tag address matches against page, so this is an entry
|
|
* we must flush.
|
|
*/
|
|
|
|
R7 >>= 10; /* Mask off the non-address bits*/
|
|
R7 <<= 10;
|
|
P3 = R7;
|
|
SSYNC;
|
|
FLUSHINV [P3]; /* And flush the entry*/
|
|
fskip:
|
|
fe0: R5 += 1; /* Advance to next Set*/
|
|
fe1: BITSET(R2, 26); /* Go to next Way.*/
|
|
|
|
dfinished:
|
|
SSYNC; /* Ensure the data gets out to mem.*/
|
|
|
|
/*Finished. Restore context.*/
|
|
LB1 = [SP++];
|
|
LT1 = [SP++];
|
|
LC1 = [SP++];
|
|
LB0 = [SP++];
|
|
LT0 = [SP++];
|
|
LC0 = [SP++];
|
|
( R7:0, P5:0 ) = [SP++];
|
|
RTS;
|
|
|
|
dflush_whole_page:
|
|
|
|
/* It's a 1K or 4K page, so quicker to just flush the
|
|
* entire page.
|
|
*/
|
|
|
|
P1 = 32; /* For 1K pages*/
|
|
P2 = P1 << 2; /* For 4K pages*/
|
|
P0 = R0; /* Start of page*/
|
|
CC = BITTST(R1, 16); /* Whether 1K or 4K*/
|
|
IF CC P1 = P2;
|
|
P1 += -1; /* Unroll one iteration*/
|
|
SSYNC;
|
|
FLUSHINV [P0++]; /* because CSYNC can't end loops.*/
|
|
LSETUP (eall, eall) LC0 = P1;
|
|
eall: FLUSHINV [P0++];
|
|
SSYNC;
|
|
JUMP dfinished;
|
|
|
|
.align 4;
|
|
page_prefix_table:
|
|
.byte4 0xFFFFFC00; /* 1K */
|
|
.byte4 0xFFFFF000; /* 4K */
|
|
.byte4 0xFFF00000; /* 1M */
|
|
.byte4 0xFFC00000; /* 4M */
|
|
.page_prefix_table.end:
|