diff options
Diffstat (limited to 'arch/tile/lib/memcpy_32.S')
| -rw-r--r-- | arch/tile/lib/memcpy_32.S | 206 |
1 files changed, 104 insertions, 102 deletions
diff --git a/arch/tile/lib/memcpy_32.S b/arch/tile/lib/memcpy_32.S index 30c3b7ebb55d..2a419a6122db 100644 --- a/arch/tile/lib/memcpy_32.S +++ b/arch/tile/lib/memcpy_32.S @@ -10,14 +10,16 @@ * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or * NON INFRINGEMENT. See the GNU General Public License for * more details. - * - * This file shares the implementation of the userspace memcpy and - * the kernel's memcpy, copy_to_user and copy_from_user. */ #include <arch/chip.h> +/* + * This file shares the implementation of the userspace memcpy and + * the kernel's memcpy, copy_to_user and copy_from_user. + */ + #include <linux/linkage.h> /* On TILE64, we wrap these functions via arch/tile/lib/memcpy_tile64.c */ @@ -53,9 +55,9 @@ */ ENTRY(__copy_from_user_inatomic) .type __copy_from_user_inatomic, @function - FEEDBACK_ENTER_EXPLICIT(__copy_from_user_inatomic, \ + FEEDBACK_ENTER_EXPLICIT(__copy_from_user_inatomic, \ .text.memcpy_common, \ - .Lend_memcpy_common - __copy_from_user_inatomic) + .Lend_memcpy_common - __copy_from_user_inatomic) { movei r29, IS_COPY_FROM_USER; j memcpy_common } .size __copy_from_user_inatomic, . - __copy_from_user_inatomic @@ -64,7 +66,7 @@ ENTRY(__copy_from_user_inatomic) */ ENTRY(__copy_from_user_zeroing) .type __copy_from_user_zeroing, @function - FEEDBACK_REENTER(__copy_from_user_inatomic) + FEEDBACK_REENTER(__copy_from_user_inatomic) { movei r29, IS_COPY_FROM_USER_ZEROING; j memcpy_common } .size __copy_from_user_zeroing, . - __copy_from_user_zeroing @@ -74,13 +76,13 @@ ENTRY(__copy_from_user_zeroing) */ ENTRY(__copy_to_user_inatomic) .type __copy_to_user_inatomic, @function - FEEDBACK_REENTER(__copy_from_user_inatomic) + FEEDBACK_REENTER(__copy_from_user_inatomic) { movei r29, IS_COPY_TO_USER; j memcpy_common } .size __copy_to_user_inatomic, . - __copy_to_user_inatomic ENTRY(memcpy) .type memcpy, @function - FEEDBACK_REENTER(__copy_from_user_inatomic) + FEEDBACK_REENTER(__copy_from_user_inatomic) { movei r29, IS_MEMCPY } .size memcpy, . - memcpy /* Fall through */ @@ -157,35 +159,35 @@ EX: { sw r0, r3; addi r0, r0, 4; addi r2, r2, -4 } { addi r3, r1, 60; andi r9, r9, -64 } #if CHIP_HAS_WH64() - /* No need to prefetch dst, we'll just do the wh64 - * right before we copy a line. + /* No need to prefetch dst, we'll just do the wh64 + * right before we copy a line. */ #endif EX: { lw r5, r3; addi r3, r3, 64; movei r4, 1 } - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, .; move r27, lr } + /* Intentionally stall for a few cycles to leave L2 cache alone. */ + { bnzt zero, .; move r27, lr } EX: { lw r6, r3; addi r3, r3, 64 } - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } + /* Intentionally stall for a few cycles to leave L2 cache alone. */ + { bnzt zero, . } EX: { lw r7, r3; addi r3, r3, 64 } #if !CHIP_HAS_WH64() - /* Prefetch the dest */ - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } - /* Use a real load to cause a TLB miss if necessary. We aren't using - * r28, so this should be fine. - */ + /* Prefetch the dest */ + /* Intentionally stall for a few cycles to leave L2 cache alone. */ + { bnzt zero, . } + /* Use a real load to cause a TLB miss if necessary. We aren't using + * r28, so this should be fine. + */ EX: { lw r28, r9; addi r9, r9, 64 } - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } - { prefetch r9; addi r9, r9, 64 } - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } - { prefetch r9; addi r9, r9, 64 } + /* Intentionally stall for a few cycles to leave L2 cache alone. */ + { bnzt zero, . } + { prefetch r9; addi r9, r9, 64 } + /* Intentionally stall for a few cycles to leave L2 cache alone. */ + { bnzt zero, . } + { prefetch r9; addi r9, r9, 64 } #endif - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bz zero, .Lbig_loop2 } + /* Intentionally stall for a few cycles to leave L2 cache alone. */ + { bz zero, .Lbig_loop2 } /* On entry to this loop: * - r0 points to the start of dst line 0 @@ -197,7 +199,7 @@ EX: { lw r28, r9; addi r9, r9, 64 } * to some "safe" recently loaded address. * - r5 contains *(r1 + 60) [i.e. last word of source line 0] * - r6 contains *(r1 + 64 + 60) [i.e. last word of source line 1] - * - r9 contains ((r0 + 63) & -64) + * - r9 contains ((r0 + 63) & -64) * [start of next dst cache line.] */ @@ -208,137 +210,137 @@ EX: { lw r28, r9; addi r9, r9, 64 } /* Copy line 0, first stalling until r5 is ready. */ EX: { move r12, r5; lw r16, r1 } { bz r4, .Lcopy_8_check; slti_u r8, r2, 8 } - /* Prefetch several lines ahead. */ + /* Prefetch several lines ahead. */ EX: { lw r5, r3; addi r3, r3, 64 } - { jal .Lcopy_line } + { jal .Lcopy_line } /* Copy line 1, first stalling until r6 is ready. */ EX: { move r12, r6; lw r16, r1 } { bz r4, .Lcopy_8_check; slti_u r8, r2, 8 } - /* Prefetch several lines ahead. */ + /* Prefetch several lines ahead. */ EX: { lw r6, r3; addi r3, r3, 64 } { jal .Lcopy_line } /* Copy line 2, first stalling until r7 is ready. */ EX: { move r12, r7; lw r16, r1 } { bz r4, .Lcopy_8_check; slti_u r8, r2, 8 } - /* Prefetch several lines ahead. */ + /* Prefetch several lines ahead. */ EX: { lw r7, r3; addi r3, r3, 64 } - /* Use up a caches-busy cycle by jumping back to the top of the - * loop. Might as well get it out of the way now. - */ - { j .Lbig_loop } + /* Use up a caches-busy cycle by jumping back to the top of the + * loop. Might as well get it out of the way now. + */ + { j .Lbig_loop } /* On entry: * - r0 points to the destination line. * - r1 points to the source line. - * - r3 is the next prefetch address. + * - r3 is the next prefetch address. * - r9 holds the last address used for wh64. * - r12 = WORD_15 - * - r16 = WORD_0. - * - r17 == r1 + 16. - * - r27 holds saved lr to restore. + * - r16 = WORD_0. + * - r17 == r1 + 16. + * - r27 holds saved lr to restore. * * On exit: * - r0 is incremented by 64. * - r1 is incremented by 64, unless that would point to a word - * beyond the end of the source array, in which case it is redirected - * to point to an arbitrary word already in the cache. + * beyond the end of the source array, in which case it is redirected + * to point to an arbitrary word already in the cache. * - r2 is decremented by 64. - * - r3 is unchanged, unless it points to a word beyond the - * end of the source array, in which case it is redirected - * to point to an arbitrary word already in the cache. - * Redirecting is OK since if we are that close to the end - * of the array we will not come back to this subroutine - * and use the contents of the prefetched address. + * - r3 is unchanged, unless it points to a word beyond the + * end of the source array, in which case it is redirected + * to point to an arbitrary word already in the cache. + * Redirecting is OK since if we are that close to the end + * of the array we will not come back to this subroutine + * and use the contents of the prefetched address. * - r4 is nonzero iff r2 >= 64. - * - r9 is incremented by 64, unless it points beyond the - * end of the last full destination cache line, in which - * case it is redirected to a "safe address" that can be - * clobbered (sp - 64) + * - r9 is incremented by 64, unless it points beyond the + * end of the last full destination cache line, in which + * case it is redirected to a "safe address" that can be + * clobbered (sp - 64) * - lr contains the value in r27. */ /* r26 unused */ .Lcopy_line: - /* TODO: when r3 goes past the end, we would like to redirect it - * to prefetch the last partial cache line (if any) just once, for the - * benefit of the final cleanup loop. But we don't want to - * prefetch that line more than once, or subsequent prefetches - * will go into the RTF. But then .Lbig_loop should unconditionally - * branch to top of loop to execute final prefetch, and its - * nop should become a conditional branch. - */ - - /* We need two non-memory cycles here to cover the resources - * used by the loads initiated by the caller. - */ - { add r15, r1, r2 } + /* TODO: when r3 goes past the end, we would like to redirect it + * to prefetch the last partial cache line (if any) just once, for the + * benefit of the final cleanup loop. But we don't want to + * prefetch that line more than once, or subsequent prefetches + * will go into the RTF. But then .Lbig_loop should unconditionally + * branch to top of loop to execute final prefetch, and its + * nop should become a conditional branch. + */ + + /* We need two non-memory cycles here to cover the resources + * used by the loads initiated by the caller. + */ + { add r15, r1, r2 } .Lcopy_line2: - { slt_u r13, r3, r15; addi r17, r1, 16 } + { slt_u r13, r3, r15; addi r17, r1, 16 } - /* NOTE: this will stall for one cycle as L1 is busy. */ + /* NOTE: this will stall for one cycle as L1 is busy. */ - /* Fill second L1D line. */ + /* Fill second L1D line. */ EX: { lw r17, r17; addi r1, r1, 48; mvz r3, r13, r1 } /* r17 = WORD_4 */ #if CHIP_HAS_WH64() - /* Prepare destination line for writing. */ + /* Prepare destination line for writing. */ EX: { wh64 r9; addi r9, r9, 64 } #else - /* Prefetch dest line */ + /* Prefetch dest line */ { prefetch r9; addi r9, r9, 64 } #endif - /* Load seven words that are L1D hits to cover wh64 L2 usage. */ + /* Load seven words that are L1D hits to cover wh64 L2 usage. */ - /* Load the three remaining words from the last L1D line, which - * we know has already filled the L1D. - */ + /* Load the three remaining words from the last L1D line, which + * we know has already filled the L1D. + */ EX: { lw r4, r1; addi r1, r1, 4; addi r20, r1, 16 } /* r4 = WORD_12 */ EX: { lw r8, r1; addi r1, r1, 4; slt_u r13, r20, r15 }/* r8 = WORD_13 */ EX: { lw r11, r1; addi r1, r1, -52; mvz r20, r13, r1 } /* r11 = WORD_14 */ - /* Load the three remaining words from the first L1D line, first - * stalling until it has filled by "looking at" r16. - */ + /* Load the three remaining words from the first L1D line, first + * stalling until it has filled by "looking at" r16. + */ EX: { lw r13, r1; addi r1, r1, 4; move zero, r16 } /* r13 = WORD_1 */ EX: { lw r14, r1; addi r1, r1, 4 } /* r14 = WORD_2 */ EX: { lw r15, r1; addi r1, r1, 8; addi r10, r0, 60 } /* r15 = WORD_3 */ - /* Load second word from the second L1D line, first - * stalling until it has filled by "looking at" r17. - */ + /* Load second word from the second L1D line, first + * stalling until it has filled by "looking at" r17. + */ EX: { lw r19, r1; addi r1, r1, 4; move zero, r17 } /* r19 = WORD_5 */ - /* Store last word to the destination line, potentially dirtying it - * for the first time, which keeps the L2 busy for two cycles. - */ + /* Store last word to the destination line, potentially dirtying it + * for the first time, which keeps the L2 busy for two cycles. + */ EX: { sw r10, r12 } /* store(WORD_15) */ - /* Use two L1D hits to cover the sw L2 access above. */ + /* Use two L1D hits to cover the sw L2 access above. */ EX: { lw r10, r1; addi r1, r1, 4 } /* r10 = WORD_6 */ EX: { lw r12, r1; addi r1, r1, 4 } /* r12 = WORD_7 */ - /* Fill third L1D line. */ + /* Fill third L1D line. */ EX: { lw r18, r1; addi r1, r1, 4 } /* r18 = WORD_8 */ - /* Store first L1D line. */ + /* Store first L1D line. */ EX: { sw r0, r16; addi r0, r0, 4; add r16, r0, r2 } /* store(WORD_0) */ EX: { sw r0, r13; addi r0, r0, 4; andi r16, r16, -64 } /* store(WORD_1) */ EX: { sw r0, r14; addi r0, r0, 4; slt_u r16, r9, r16 } /* store(WORD_2) */ #if CHIP_HAS_WH64() EX: { sw r0, r15; addi r0, r0, 4; addi r13, sp, -64 } /* store(WORD_3) */ #else - /* Back up the r9 to a cache line we are already storing to + /* Back up the r9 to a cache line we are already storing to * if it gets past the end of the dest vector. Strictly speaking, * we don't need to back up to the start of a cache line, but it's free * and tidy, so why not? - */ + */ EX: { sw r0, r15; addi r0, r0, 4; andi r13, r0, -64 } /* store(WORD_3) */ #endif - /* Store second L1D line. */ + /* Store second L1D line. */ EX: { sw r0, r17; addi r0, r0, 4; mvz r9, r16, r13 }/* store(WORD_4) */ EX: { sw r0, r19; addi r0, r0, 4 } /* store(WORD_5) */ EX: { sw r0, r10; addi r0, r0, 4 } /* store(WORD_6) */ @@ -348,30 +350,30 @@ EX: { lw r13, r1; addi r1, r1, 4; move zero, r18 } /* r13 = WORD_9 */ EX: { lw r14, r1; addi r1, r1, 4 } /* r14 = WORD_10 */ EX: { lw r15, r1; move r1, r20 } /* r15 = WORD_11 */ - /* Store third L1D line. */ + /* Store third L1D line. */ EX: { sw r0, r18; addi r0, r0, 4 } /* store(WORD_8) */ EX: { sw r0, r13; addi r0, r0, 4 } /* store(WORD_9) */ EX: { sw r0, r14; addi r0, r0, 4 } /* store(WORD_10) */ EX: { sw r0, r15; addi r0, r0, 4 } /* store(WORD_11) */ - /* Store rest of fourth L1D line. */ + /* Store rest of fourth L1D line. */ EX: { sw r0, r4; addi r0, r0, 4 } /* store(WORD_12) */ - { + { EX: sw r0, r8 /* store(WORD_13) */ - addi r0, r0, 4 + addi r0, r0, 4 /* Will r2 be > 64 after we subtract 64 below? */ - shri r4, r2, 7 - } - { + shri r4, r2, 7 + } + { EX: sw r0, r11 /* store(WORD_14) */ - addi r0, r0, 8 - /* Record 64 bytes successfully copied. */ - addi r2, r2, -64 - } + addi r0, r0, 8 + /* Record 64 bytes successfully copied. */ + addi r2, r2, -64 + } { jrp lr; move lr, r27 } - /* Convey to the backtrace library that the stack frame is size + /* Convey to the backtrace library that the stack frame is size * zero, and the real return address is on the stack rather than * in 'lr'. */ |