| /* Overlay manager for SPU. |
| |
| Copyright (C) 2006-2016 Free Software Foundation, Inc. |
| |
| This file is part of the GNU Binutils. |
| |
| This program is free software; you can redistribute it and/or modify |
| it under the terms of the GNU General Public License as published by |
| the Free Software Foundation; either version 3 of the License, or |
| (at your option) any later version. |
| |
| This program 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 General Public License for more details. |
| |
| You should have received a copy of the GNU General Public License |
| along with this program; if not, write to the Free Software |
| Foundation, Inc., 51 Franklin Street - Fifth Floor, Boston, |
| MA 02110-1301, USA. */ |
| |
| /* MFC DMA defn's. */ |
| #define MFC_GET_CMD 0x40 |
| #define MFC_MAX_DMA_SIZE 0x4000 |
| #define MFC_TAG_UPDATE_ALL 2 |
| #define MFC_TAG_ID 0 |
| |
| /* Register usage. */ |
| #define reserved1 $75 |
| #define parm $75 |
| #define tab1 reserved1 |
| #define tab2 reserved1 |
| #define vma reserved1 |
| #define oldvma reserved1 |
| #define newmask reserved1 |
| #define map reserved1 |
| |
| #define reserved2 $76 |
| #define off1 reserved2 |
| #define off2 reserved2 |
| #define present1 reserved2 |
| #define present2 reserved2 |
| #define sz reserved2 |
| #define cmp reserved2 |
| #define add64 reserved2 |
| #define cgbits reserved2 |
| #define off3 reserved2 |
| #define off4 reserved2 |
| #define addr4 reserved2 |
| #define off5 reserved2 |
| #define tagstat reserved2 |
| |
| #define reserved3 $77 |
| #define size1 reserved3 |
| #define size2 reserved3 |
| #define rv3 reserved3 |
| #define ealo reserved3 |
| #define cmd reserved3 |
| #define off64 reserved3 |
| #define tab3 reserved3 |
| #define tab4 reserved3 |
| #define tab5 reserved3 |
| |
| #define reserved4 $78 |
| #define ovl reserved4 |
| #define rv2 reserved4 |
| #define rv5 reserved4 |
| #define cgshuf reserved4 |
| #define newovl reserved4 |
| #define irqtmp1 reserved4 |
| #define irqtmp2 reserved4 |
| |
| #define reserved5 $79 |
| #define target reserved5 |
| |
| #define save1 $74 |
| #define rv4 save1 |
| #define rv7 save1 |
| #define tagid save1 |
| #define maxsize save1 |
| #define pbyte save1 |
| #define pbit save1 |
| |
| #define save2 $73 |
| #define cur save2 |
| #define rv6 save2 |
| #define osize save2 |
| #define zovl save2 |
| #define oldovl save2 |
| #define newvma save2 |
| |
| #define save3 $72 |
| #define rv1 save3 |
| #define ea64 save3 |
| #define buf3 save3 |
| #define genwi save3 |
| #define newmap save3 |
| #define oldmask save3 |
| |
| #define save4 $71 |
| #define irq_stat save4 |
| |
| .text |
| .align 4 |
| .type __rv_pattern, @object |
| .size __rv_pattern, 16 |
| __rv_pattern: |
| .word 0x00010203, 0x10111213, 0x80808080, 0x80808080 |
| |
| .type __cg_pattern, @object |
| .size __cg_pattern, 16 |
| __cg_pattern: |
| .word 0x04050607, 0x80808080, 0x80808080, 0x80808080 |
| |
| .type __ovly_current, @object |
| .size __ovly_current, 16 |
| __ovly_current: |
| .space 16 |
| |
| /* |
| * __ovly_return - stub for returning from overlay functions. |
| * |
| * On entry the four slots of $lr are: |
| * __ovly_return, prev ovl index, caller return addr, undefined. |
| * |
| * Load the previous overlay and jump to the caller return address. |
| * Updates __ovly_current. |
| */ |
| .align 4 |
| .global __ovly_return |
| .type __ovly_return, @function |
| __ovly_return: |
| ila tab1, _ovly_table - 16 # 0,2 0 |
| shlqbyi ovl, $lr, 4 # 1,4 0 |
| #nop |
| shlqbyi target, $lr, 8 # 1,4 1 |
| #nop; lnop |
| #nop; lnop |
| shli off1, ovl, 4 # 0,4 4 |
| #lnop |
| #nop |
| hbr ovly_ret9, target # 1,15 5 |
| #nop; lnop |
| #nop; lnop |
| #nop |
| lqx vma, tab1, off1 # 1,6 8 |
| #ifdef OVLY_IRQ_SAVE |
| nop |
| stqd save4, -64($sp) # 1,6 9 |
| #else |
| #nop; lnop |
| #endif |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| #nop |
| rotqbyi size1, vma, 4 # 1,4 14 |
| #nop |
| stqd save3, -48($sp) # 1,6 15 |
| #nop |
| stqd save2, -32($sp) # 1,6 16 |
| #nop |
| stqd save1, -16($sp) # 1,6 17 |
| andi present1, size1, 1 # 0,2 18 |
| stqr ovl, __ovly_current # 1,6 18 |
| #nop; lnop |
| #nop |
| brz present1, do_load # 1,4 20 |
| ovly_ret9: |
| #nop |
| bi target # 1,4 21 |
| |
| /* |
| * __ovly_load - copy an overlay partion to local store. |
| * |
| * On entry $75 points to a word consisting of the overlay index in |
| * the top 14 bits, and the target address in the bottom 18 bits. |
| * |
| * Sets up $lr to return via __ovly_return. If $lr is already set |
| * to return via __ovly_return, don't change it. In that case we |
| * have a tail call from one overlay function to another. |
| * Updates __ovly_current. |
| */ |
| .align 3 |
| .global __ovly_load |
| .type __ovly_load, @function |
| __ovly_load: |
| #if OVL_STUB_SIZE == 8 |
| ######## |
| #nop |
| lqd target, 0(parm) # 1,6 -11 |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| #nop |
| rotqby target, target, parm # 1,4 -5 |
| ila tab2, _ovly_table - 16 # 0,2 -4 |
| stqd save3, -48($sp) # 1,6 -4 |
| #nop |
| stqd save2, -32($sp) # 1,6 -3 |
| #nop |
| stqd save1, -16($sp) # 1,6 -2 |
| rotmi ovl, target, -18 # 0,4 -1 |
| hbr ovly_load9, target # 1,15 -1 |
| ila rv1, __ovly_return # 0,2 0 |
| #lnop |
| #nop; lnop |
| #nop |
| lqr cur, __ovly_current # 1,6 2 |
| shli off2, ovl, 4 # 0,4 3 |
| stqr ovl, __ovly_current # 1,6 3 |
| ceq rv2, $lr, rv1 # 0,2 4 |
| lqr rv3, __rv_pattern # 1,6 4 |
| #nop; lnop |
| #nop; lnop |
| #nop |
| lqx vma, tab2, off2 # 1,6 7 |
| ######## |
| #else /* OVL_STUB_SIZE == 16 */ |
| ######## |
| ila tab2, _ovly_table - 16 # 0,2 0 |
| stqd save3, -48($sp) # 1,6 0 |
| ila rv1, __ovly_return # 0,2 1 |
| stqd save2, -32($sp) # 1,6 1 |
| shli off2, ovl, 4 # 0,4 2 |
| lqr cur, __ovly_current # 1,6 2 |
| nop |
| stqr ovl, __ovly_current # 1,6 3 |
| ceq rv2, $lr, rv1 # 0,2 4 |
| lqr rv3, __rv_pattern # 1,6 4 |
| #nop |
| hbr ovly_load9, target # 1,15 5 |
| #nop |
| lqx vma, tab2, off2 # 1,6 6 |
| #nop |
| stqd save1, -16($sp) # 1,6 7 |
| ######## |
| #endif |
| |
| #nop; lnop |
| #nop; lnop |
| #nop |
| shufb rv4, rv1, cur, rv3 # 1,4 10 |
| #nop |
| fsmb rv5, rv2 # 1,4 11 |
| #nop |
| rotqmbyi rv6, $lr, -8 # 1,4 12 |
| #nop |
| rotqbyi size2, vma, 4 # 1,4 13 |
| #nop |
| lqd save3, -48($sp) # 1,6 14 |
| #nop; lnop |
| or rv7, rv4, rv6 # 0,2 16 |
| lqd save2, -32($sp) # 1,6 16 |
| andi present2, size2, 1 # 0,2 17 |
| #ifdef OVLY_IRQ_SAVE |
| stqd save4, -64($sp) # 1,6 17 |
| #else |
| lnop # 1,0 17 |
| #endif |
| selb $lr, rv7, $lr, rv5 # 0,2 18 |
| lqd save1, -16($sp) # 1,6 18 |
| #nop |
| brz present2, do_load # 1,4 19 |
| ovly_load9: |
| #nop |
| bi target # 1,4 20 |
| |
| /* If we get here, we are about to load a new overlay. |
| * "vma" contains the relevant entry from _ovly_table[]. |
| * extern struct { |
| * u32 vma; |
| * u32 size; |
| * u32 file_offset; |
| * u32 buf; |
| * } _ovly_table[]; |
| */ |
| .align 3 |
| .global __ovly_load_event |
| .type __ovly_load_event, @function |
| __ovly_load_event: |
| do_load: |
| #ifdef OVLY_IRQ_SAVE |
| ila irqtmp1, do_load10 # 0,2 -5 |
| rotqbyi sz, vma, 8 # 1,4 -5 |
| #nop |
| rdch irq_stat, $SPU_RdMachStat # 1,6 -4 |
| #nop |
| bid irqtmp1 # 1,4 -3 |
| do_load10: |
| nop |
| #else |
| #nop |
| rotqbyi sz, vma, 8 # 1,4 0 |
| #endif |
| rotqbyi osize, vma, 4 # 1,4 1 |
| #nop |
| lqa ea64, _EAR_ # 1,6 2 |
| #nop |
| lqr cgshuf, __cg_pattern # 1,6 3 |
| |
| /* We could predict the branch at the end of this loop by adding a few |
| instructions, and there are plenty of free cycles to do so without |
| impacting loop execution time. However, it doesn't make a great |
| deal of sense since we need to wait for the dma to complete anyway. */ |
| __ovly_xfer_loop: |
| #nop |
| rotqmbyi off64, sz, -4 # 1,4 4 |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| cg cgbits, ea64, off64 # 0,2 8 |
| #lnop |
| #nop; lnop |
| #nop |
| shufb add64, cgbits, cgbits, cgshuf # 1,4 10 |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| addx add64, ea64, off64 # 0,2 14 |
| #lnop |
| ila maxsize, MFC_MAX_DMA_SIZE # 0,2 15 |
| lnop |
| ori ea64, add64, 0 # 0,2 16 |
| rotqbyi ealo, add64, 4 # 1,4 16 |
| cgt cmp, osize, maxsize # 0,2 17 |
| wrch $MFC_LSA, vma # 1,6 17 |
| #nop; lnop |
| selb sz, osize, maxsize, cmp # 0,2 19 |
| wrch $MFC_EAH, ea64 # 1,6 19 |
| ila tagid, MFC_TAG_ID # 0,2 20 |
| wrch $MFC_EAL, ealo # 1,6 20 |
| ila cmd, MFC_GET_CMD # 0,2 21 |
| wrch $MFC_Size, sz # 1,6 21 |
| sf osize, sz, osize # 0,2 22 |
| wrch $MFC_TagId, tagid # 1,6 22 |
| a vma, vma, sz # 0,2 23 |
| wrch $MFC_Cmd, cmd # 1,6 23 |
| #nop |
| brnz osize, __ovly_xfer_loop # 1,4 24 |
| |
| /* Now update our data structions while waiting for DMA to complete. |
| Low bit of .size needs to be cleared on the _ovly_table entry |
| corresponding to the evicted overlay, and set on the entry for the |
| newly loaded overlay. Note that no overlay may in fact be evicted |
| as _ovly_buf_table[] starts with all zeros. Don't zap .size entry |
| for zero index! Also of course update the _ovly_buf_table entry. */ |
| #nop |
| lqr newovl, __ovly_current # 1,6 25 |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| shli off3, newovl, 4 # 0,4 31 |
| #lnop |
| ila tab3, _ovly_table - 16 # 0,2 32 |
| #lnop |
| #nop |
| fsmbi pbyte, 0x100 # 1,4 33 |
| #nop; lnop |
| #nop |
| lqx vma, tab3, off3 # 1,6 35 |
| #nop; lnop |
| andi pbit, pbyte, 1 # 0,2 37 |
| lnop |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| or newvma, vma, pbit # 0,2 41 |
| rotqbyi buf3, vma, 12 # 1,4 41 |
| #nop; lnop |
| #nop |
| stqx newvma, tab3, off3 # 1,6 43 |
| #nop; lnop |
| shli off4, buf3, 2 # 1,4 45 |
| #lnop |
| ila tab4, _ovly_buf_table - 4 # 0,2 46 |
| #lnop |
| #nop; lnop |
| #nop; lnop |
| #nop |
| lqx map, tab4, off4 # 1,6 49 |
| #nop |
| cwx genwi, tab4, off4 # 1,4 50 |
| a addr4, tab4, off4 # 0,2 51 |
| #lnop |
| #nop; lnop |
| #nop; lnop |
| #nop; lnop |
| #nop |
| rotqby oldovl, map, addr4 # 1,4 55 |
| #nop |
| shufb newmap, newovl, map, genwi # 0,4 56 |
| #if MFC_TAG_ID < 16 |
| ila newmask, 1 << MFC_TAG_ID # 0,2 57 |
| #else |
| ilhu newmask, 1 << (MFC_TAG_ID - 16) # 0,2 57 |
| #endif |
| #lnop |
| #nop; lnop |
| #nop; lnop |
| stqd newmap, 0(addr4) # 1,6 60 |
| |
| /* Save app's tagmask, wait for DMA complete, restore mask. */ |
| ila tagstat, MFC_TAG_UPDATE_ALL # 0,2 61 |
| rdch oldmask, $MFC_RdTagMask # 1,6 61 |
| #nop |
| wrch $MFC_WrTagMask, newmask # 1,6 62 |
| #nop |
| wrch $MFC_WrTagUpdate, tagstat # 1,6 63 |
| #nop |
| rdch tagstat, $MFC_RdTagStat # 1,6 64 |
| #nop |
| sync # 1,4 65 |
| /* Any hint prior to the sync is lost. A hint here allows the branch |
| to complete 15 cycles after the hint. With no hint the branch will |
| take 18 or 19 cycles. */ |
| ila tab5, _ovly_table - 16 # 0,2 66 |
| hbr do_load99, target # 1,15 66 |
| shli off5, oldovl, 4 # 0,4 67 |
| wrch $MFC_WrTagMask, oldmask # 1,6 67 |
| ceqi zovl, oldovl, 0 # 0,2 68 |
| #lnop |
| #nop; lnop |
| #nop |
| fsm zovl, zovl # 1,4 70 |
| #nop |
| lqx oldvma, tab5, off5 # 1,6 71 |
| #nop |
| lqd save3, -48($sp) # 1,6 72 |
| #nop; lnop |
| andc pbit, pbit, zovl # 0,2 74 |
| lqd save2, -32($sp) # 1,6 74 |
| #ifdef OVLY_IRQ_SAVE |
| ila irqtmp2, do_load90 # 0,2 75 |
| #lnop |
| andi irq_stat, irq_stat, 1 # 0,2 76 |
| #lnop |
| #else |
| #nop; lnop |
| #nop; lnop |
| #endif |
| andc oldvma, oldvma, pbit # 0,2 77 |
| lqd save1, -16($sp) # 1,6 77 |
| nop # 0,0 78 |
| #lnop |
| #nop |
| stqx oldvma, tab5, off5 # 1,6 79 |
| #nop |
| #ifdef OVLY_IRQ_SAVE |
| binze irq_stat, irqtmp2 # 1,4 80 |
| do_load90: |
| #nop |
| lqd save4, -64($sp) # 1,6 84 |
| #else |
| #nop; lnop |
| #endif |
| |
| .global _ovly_debug_event |
| .type _ovly_debug_event, @function |
| _ovly_debug_event: |
| nop |
| /* Branch to target address. */ |
| do_load99: |
| bi target # 1,4 81/85 |
| |
| .size __ovly_load, . - __ovly_load |
