diff options
Diffstat (limited to 'src/string/armel')
| -rw-r--r-- | src/string/armel/memcpy.s | 558 |
1 files changed, 279 insertions, 279 deletions
diff --git a/src/string/armel/memcpy.s b/src/string/armel/memcpy.s index 656cad95..9c5e38d3 100644 --- a/src/string/armel/memcpy.s +++ b/src/string/armel/memcpy.s @@ -49,113 +49,113 @@ memcpy: * ARM ABI. Since we have to save R0, we might as well save R4 * which we can use for better pipelining of the reads below */ - .fnstart - .save {r0, r4, lr} - stmfd sp!, {r0, r4, lr} - /* Making room for r5-r11 which will be spilled later */ - .pad #28 - sub sp, sp, #28 - - /* it simplifies things to take care of len<4 early */ - cmp r2, #4 - blo copy_last_3_and_return - - /* compute the offset to align the source - * offset = (4-(src&3))&3 = -src & 3 - */ - rsb r3, r1, #0 - ands r3, r3, #3 - beq src_aligned - - /* align source to 32 bits. We need to insert 2 instructions between - * a ldr[b|h] and str[b|h] because byte and half-word instructions - * stall 2 cycles. - */ - movs r12, r3, lsl #31 - sub r2, r2, r3 /* we know that r3 <= r2 because r2 >= 4 */ - .word 0x44d13001 /* ldrbmi r3, [r1], #1 */ - .word 0x24d14001 /* ldrbcs r4, [r1], #1 */ - .word 0x24d1c001 /* ldrbcs r12,[r1], #1 */ - .word 0x44c03001 /* strbmi r3, [r0], #1 */ - .word 0x24c04001 /* strbcs r4, [r0], #1 */ - .word 0x24c0c001 /* strbcs r12,[r0], #1 */ + .fnstart + .save {r0, r4, lr} + stmfd sp!, {r0, r4, lr} + /* Making room for r5-r11 which will be spilled later */ + .pad #28 + sub sp, sp, #28 + + /* it simplifies things to take care of len<4 early */ + cmp r2, #4 + blo copy_last_3_and_return + + /* compute the offset to align the source + * offset = (4-(src&3))&3 = -src & 3 + */ + rsb r3, r1, #0 + ands r3, r3, #3 + beq src_aligned + + /* align source to 32 bits. We need to insert 2 instructions between + * a ldr[b|h] and str[b|h] because byte and half-word instructions + * stall 2 cycles. + */ + movs r12, r3, lsl #31 + sub r2, r2, r3 /* we know that r3 <= r2 because r2 >= 4 */ + .word 0x44d13001 /* ldrbmi r3, [r1], #1 */ + .word 0x24d14001 /* ldrbcs r4, [r1], #1 */ + .word 0x24d1c001 /* ldrbcs r12,[r1], #1 */ + .word 0x44c03001 /* strbmi r3, [r0], #1 */ + .word 0x24c04001 /* strbcs r4, [r0], #1 */ + .word 0x24c0c001 /* strbcs r12,[r0], #1 */ src_aligned: /* see if src and dst are aligned together (congruent) */ - eor r12, r0, r1 - tst r12, #3 - bne non_congruent - - /* Use post-incriment mode for stm to spill r5-r11 to reserved stack - * frame. Don't update sp. - */ - stmea sp, {r5-r11} - - /* align the destination to a cache-line */ - rsb r3, r0, #0 - ands r3, r3, #0x1C - beq congruent_aligned32 - cmp r3, r2 - andhi r3, r2, #0x1C - - /* conditionnaly copies 0 to 7 words (length in r3) */ - movs r12, r3, lsl #28 - ldmcs r1!, {r4, r5, r6, r7} /* 16 bytes */ - ldmmi r1!, {r8, r9} /* 8 bytes */ - stmcs r0!, {r4, r5, r6, r7} - stmmi r0!, {r8, r9} - tst r3, #0x4 - ldrne r10,[r1], #4 /* 4 bytes */ - strne r10,[r0], #4 - sub r2, r2, r3 + eor r12, r0, r1 + tst r12, #3 + bne non_congruent + + /* Use post-incriment mode for stm to spill r5-r11 to reserved stack + * frame. Don't update sp. + */ + stmea sp, {r5-r11} + + /* align the destination to a cache-line */ + rsb r3, r0, #0 + ands r3, r3, #0x1C + beq congruent_aligned32 + cmp r3, r2 + andhi r3, r2, #0x1C + + /* conditionnaly copies 0 to 7 words (length in r3) */ + movs r12, r3, lsl #28 + ldmcs r1!, {r4, r5, r6, r7} /* 16 bytes */ + ldmmi r1!, {r8, r9} /* 8 bytes */ + stmcs r0!, {r4, r5, r6, r7} + stmmi r0!, {r8, r9} + tst r3, #0x4 + ldrne r10,[r1], #4 /* 4 bytes */ + strne r10,[r0], #4 + sub r2, r2, r3 congruent_aligned32: /* - * here source is aligned to 32 bytes. - */ + * here source is aligned to 32 bytes. + */ cached_aligned32: - subs r2, r2, #32 - blo less_than_32_left - - /* - * We preload a cache-line up to 64 bytes ahead. On the 926, this will - * stall only until the requested world is fetched, but the linefill - * continues in the the background. - * While the linefill is going, we write our previous cache-line - * into the write-buffer (which should have some free space). - * When the linefill is done, the writebuffer will - * start dumping its content into memory - * - * While all this is going, we then load a full cache line into - * 8 registers, this cache line should be in the cache by now - * (or partly in the cache). - * - * This code should work well regardless of the source/dest alignment. - * - */ - - /* Align the preload register to a cache-line because the cpu does - * "critical word first" (the first word requested is loaded first). - */ - @ bic r12, r1, #0x1F - @ add r12, r12, #64 - -1: ldmia r1!, { r4-r11 } - subs r2, r2, #32 - - /* - * NOTE: if r12 is more than 64 ahead of r1, the following ldrhi - * for ARM9 preload will not be safely guarded by the preceding subs. - * When it is safely guarded the only possibility to have SIGSEGV here - * is because the caller overstates the length. - */ - @ ldrhi r3, [r12], #32 /* cheap ARM9 preload */ - stmia r0!, { r4-r11 } - bhs 1b - - add r2, r2, #32 + subs r2, r2, #32 + blo less_than_32_left + + /* + * We preload a cache-line up to 64 bytes ahead. On the 926, this will + * stall only until the requested world is fetched, but the linefill + * continues in the the background. + * While the linefill is going, we write our previous cache-line + * into the write-buffer (which should have some free space). + * When the linefill is done, the writebuffer will + * start dumping its content into memory + * + * While all this is going, we then load a full cache line into + * 8 registers, this cache line should be in the cache by now + * (or partly in the cache). + * + * This code should work well regardless of the source/dest alignment. + * + */ + + /* Align the preload register to a cache-line because the cpu does + * "critical word first" (the first word requested is loaded first). + */ + @ bic r12, r1, #0x1F + @ add r12, r12, #64 + +1: ldmia r1!, { r4-r11 } + subs r2, r2, #32 + + /* + * NOTE: if r12 is more than 64 ahead of r1, the following ldrhi + * for ARM9 preload will not be safely guarded by the preceding subs. + * When it is safely guarded the only possibility to have SIGSEGV here + * is because the caller overstates the length. + */ + @ ldrhi r3, [r12], #32 /* cheap ARM9 preload */ + stmia r0!, { r4-r11 } + bhs 1b + + add r2, r2, #32 less_than_32_left: /* @@ -166,30 +166,30 @@ less_than_32_left: * be a common case (if not executed the code below takes * about 16 cycles) */ - tst r2, #0x1F - beq 1f - - /* conditionnaly copies 0 to 31 bytes */ - movs r12, r2, lsl #28 - ldmcs r1!, {r4, r5, r6, r7} /* 16 bytes */ - ldmmi r1!, {r8, r9} /* 8 bytes */ - stmcs r0!, {r4, r5, r6, r7} - stmmi r0!, {r8, r9} - movs r12, r2, lsl #30 - ldrcs r3, [r1], #4 /* 4 bytes */ - .word 0x40d140b2 /* ldrhmi r4, [r1], #2 */ /* 2 bytes */ - strcs r3, [r0], #4 - .word 0x40c040b2 /* strhmi r4, [r0], #2 */ - tst r2, #0x1 - .word 0x15d13000 /* ldrbne r3, [r1] */ /* last byte */ - .word 0x15c03000 /* strbne r3, [r0] */ - - /* we're done! restore everything and return */ -1: ldmfd sp!, {r5-r11} - ldmfd sp!, {r0, r4, lr} - tst lr, #1 - moveq pc, lr - bx lr + tst r2, #0x1F + beq 1f + + /* conditionnaly copies 0 to 31 bytes */ + movs r12, r2, lsl #28 + ldmcs r1!, {r4, r5, r6, r7} /* 16 bytes */ + ldmmi r1!, {r8, r9} /* 8 bytes */ + stmcs r0!, {r4, r5, r6, r7} + stmmi r0!, {r8, r9} + movs r12, r2, lsl #30 + ldrcs r3, [r1], #4 /* 4 bytes */ + .word 0x40d140b2 /* ldrhmi r4, [r1], #2 */ /* 2 bytes */ + strcs r3, [r0], #4 + .word 0x40c040b2 /* strhmi r4, [r0], #2 */ + tst r2, #0x1 + .word 0x15d13000 /* ldrbne r3, [r1] */ /* last byte */ + .word 0x15c03000 /* strbne r3, [r0] */ + + /* we're done! restore everything and return */ +1: ldmfd sp!, {r5-r11} + ldmfd sp!, {r0, r4, lr} + tst lr, #1 + moveq pc, lr + bx lr /********************************************************************/ @@ -202,180 +202,180 @@ non_congruent: * (the number of bytes written is always smaller, because we have * partial words in the shift queue) */ - cmp r2, #4 - blo copy_last_3_and_return - - /* Use post-incriment mode for stm to spill r5-r11 to reserved stack - * frame. Don't update sp. - */ - stmea sp, {r5-r11} - - /* compute shifts needed to align src to dest */ - rsb r5, r0, #0 - and r5, r5, #3 /* r5 = # bytes in partial words */ - mov r12, r5, lsl #3 /* r12 = right */ - rsb lr, r12, #32 /* lr = left */ - - /* read the first word */ - ldr r3, [r1], #4 - sub r2, r2, #4 - - /* write a partial word (0 to 3 bytes), such that destination - * becomes aligned to 32 bits (r5 = nb of words to copy for alignment) - */ - movs r5, r5, lsl #31 - .word 0x44c03001 /* strbmi r3, [r0], #1 */ - movmi r3, r3, lsr #8 - .word 0x24c03001 /* strbcs r3, [r0], #1 */ - movcs r3, r3, lsr #8 - .word 0x24c03001 /* strbcs r3, [r0], #1 */ - movcs r3, r3, lsr #8 - - cmp r2, #4 - blo partial_word_tail - - /* Align destination to 32 bytes (cache line boundary) */ -1: tst r0, #0x1c - beq 2f - ldr r5, [r1], #4 - sub r2, r2, #4 - orr r4, r3, r5, lsl lr - mov r3, r5, lsr r12 - str r4, [r0], #4 - cmp r2, #4 - bhs 1b - blo partial_word_tail + cmp r2, #4 + blo copy_last_3_and_return + + /* Use post-incriment mode for stm to spill r5-r11 to reserved stack + * frame. Don't update sp. + */ + stmea sp, {r5-r11} + + /* compute shifts needed to align src to dest */ + rsb r5, r0, #0 + and r5, r5, #3 /* r5 = # bytes in partial words */ + mov r12, r5, lsl #3 /* r12 = right */ + rsb lr, r12, #32 /* lr = left */ + + /* read the first word */ + ldr r3, [r1], #4 + sub r2, r2, #4 + + /* write a partial word (0 to 3 bytes), such that destination + * becomes aligned to 32 bits (r5 = nb of words to copy for alignment) + */ + movs r5, r5, lsl #31 + .word 0x44c03001 /* strbmi r3, [r0], #1 */ + movmi r3, r3, lsr #8 + .word 0x24c03001 /* strbcs r3, [r0], #1 */ + movcs r3, r3, lsr #8 + .word 0x24c03001 /* strbcs r3, [r0], #1 */ + movcs r3, r3, lsr #8 + + cmp r2, #4 + blo partial_word_tail + + /* Align destination to 32 bytes (cache line boundary) */ +1: tst r0, #0x1c + beq 2f + ldr r5, [r1], #4 + sub r2, r2, #4 + orr r4, r3, r5, lsl lr + mov r3, r5, lsr r12 + str r4, [r0], #4 + cmp r2, #4 + bhs 1b + blo partial_word_tail /* copy 32 bytes at a time */ -2: subs r2, r2, #32 - blo less_than_thirtytwo +2: subs r2, r2, #32 + blo less_than_thirtytwo - /* Use immediate mode for the shifts, because there is an extra cycle - * for register shifts, which could account for up to 50% of - * performance hit. - */ + /* Use immediate mode for the shifts, because there is an extra cycle + * for register shifts, which could account for up to 50% of + * performance hit. + */ - cmp r12, #24 - beq loop24 - cmp r12, #8 - beq loop8 + cmp r12, #24 + beq loop24 + cmp r12, #8 + beq loop8 loop16: - ldr r12, [r1], #4 -1: mov r4, r12 - ldmia r1!, { r5,r6,r7, r8,r9,r10,r11} - subs r2, r2, #32 - ldrhs r12, [r1], #4 - orr r3, r3, r4, lsl #16 - mov r4, r4, lsr #16 - orr r4, r4, r5, lsl #16 - mov r5, r5, lsr #16 - orr r5, r5, r6, lsl #16 - mov r6, r6, lsr #16 - orr r6, r6, r7, lsl #16 - mov r7, r7, lsr #16 - orr r7, r7, r8, lsl #16 - mov r8, r8, lsr #16 - orr r8, r8, r9, lsl #16 - mov r9, r9, lsr #16 - orr r9, r9, r10, lsl #16 - mov r10, r10, lsr #16 - orr r10, r10, r11, lsl #16 - stmia r0!, {r3,r4,r5,r6, r7,r8,r9,r10} - mov r3, r11, lsr #16 - bhs 1b - b less_than_thirtytwo + ldr r12, [r1], #4 +1: mov r4, r12 + ldmia r1!, { r5,r6,r7, r8,r9,r10,r11} + subs r2, r2, #32 + ldrhs r12, [r1], #4 + orr r3, r3, r4, lsl #16 + mov r4, r4, lsr #16 + orr r4, r4, r5, lsl #16 + mov r5, r5, lsr #16 + orr r5, r5, r6, lsl #16 + mov r6, r6, lsr #16 + orr r6, r6, r7, lsl #16 + mov r7, r7, lsr #16 + orr r7, r7, r8, lsl #16 + mov r8, r8, lsr #16 + orr r8, r8, r9, lsl #16 + mov r9, r9, lsr #16 + orr r9, r9, r10, lsl #16 + mov r10, r10, lsr #16 + orr r10, r10, r11, lsl #16 + stmia r0!, {r3,r4,r5,r6, r7,r8,r9,r10} + mov r3, r11, lsr #16 + bhs 1b + b less_than_thirtytwo loop8: - ldr r12, [r1], #4 -1: mov r4, r12 - ldmia r1!, { r5,r6,r7, r8,r9,r10,r11} - subs r2, r2, #32 - ldrhs r12, [r1], #4 - orr r3, r3, r4, lsl #24 - mov r4, r4, lsr #8 - orr r4, r4, r5, lsl #24 - mov r5, r5, lsr #8 - orr r5, r5, r6, lsl #24 - mov r6, r6, lsr #8 - orr r6, r6, r7, lsl #24 - mov r7, r7, lsr #8 - orr r7, r7, r8, lsl #24 - mov r8, r8, lsr #8 - orr r8, r8, r9, lsl #24 - mov r9, r9, lsr #8 - orr r9, r9, r10, lsl #24 - mov r10, r10, lsr #8 - orr r10, r10, r11, lsl #24 - stmia r0!, {r3,r4,r5,r6, r7,r8,r9,r10} - mov r3, r11, lsr #8 - bhs 1b - b less_than_thirtytwo + ldr r12, [r1], #4 +1: mov r4, r12 + ldmia r1!, { r5,r6,r7, r8,r9,r10,r11} + subs r2, r2, #32 + ldrhs r12, [r1], #4 + orr r3, r3, r4, lsl #24 + mov r4, r4, lsr #8 + orr r4, r4, r5, lsl #24 + mov r5, r5, lsr #8 + orr r5, r5, r6, lsl #24 + mov r6, r6, lsr #8 + orr r6, r6, r7, lsl #24 + mov r7, r7, lsr #8 + orr r7, r7, r8, lsl #24 + mov r8, r8, lsr #8 + orr r8, r8, r9, lsl #24 + mov r9, r9, lsr #8 + orr r9, r9, r10, lsl #24 + mov r10, r10, lsr #8 + orr r10, r10, r11, lsl #24 + stmia r0!, {r3,r4,r5,r6, r7,r8,r9,r10} + mov r3, r11, lsr #8 + bhs 1b + b less_than_thirtytwo loop24: - ldr r12, [r1], #4 -1: mov r4, r12 - ldmia r1!, { r5,r6,r7, r8,r9,r10,r11} - subs r2, r2, #32 - ldrhs r12, [r1], #4 - orr r3, r3, r4, lsl #8 - mov r4, r4, lsr #24 - orr r4, r4, r5, lsl #8 - mov r5, r5, lsr #24 - orr r5, r5, r6, lsl #8 - mov r6, r6, lsr #24 - orr r6, r6, r7, lsl #8 - mov r7, r7, lsr #24 - orr r7, r7, r8, lsl #8 - mov r8, r8, lsr #24 - orr r8, r8, r9, lsl #8 - mov r9, r9, lsr #24 - orr r9, r9, r10, lsl #8 - mov r10, r10, lsr #24 - orr r10, r10, r11, lsl #8 - stmia r0!, {r3,r4,r5,r6, r7,r8,r9,r10} - mov r3, r11, lsr #24 - bhs 1b + ldr r12, [r1], #4 +1: mov r4, r12 + ldmia r1!, { r5,r6,r7, r8,r9,r10,r11} + subs r2, r2, #32 + ldrhs r12, [r1], #4 + orr r3, r3, r4, lsl #8 + mov r4, r4, lsr #24 + orr r4, r4, r5, lsl #8 + mov r5, r5, lsr #24 + orr r5, r5, r6, lsl #8 + mov r6, r6, lsr #24 + orr r6, r6, r7, lsl #8 + mov r7, r7, lsr #24 + orr r7, r7, r8, lsl #8 + mov r8, r8, lsr #24 + orr r8, r8, r9, lsl #8 + mov r9, r9, lsr #24 + orr r9, r9, r10, lsl #8 + mov r10, r10, lsr #24 + orr r10, r10, r11, lsl #8 + stmia r0!, {r3,r4,r5,r6, r7,r8,r9,r10} + mov r3, r11, lsr #24 + bhs 1b less_than_thirtytwo: /* copy the last 0 to 31 bytes of the source */ - rsb r12, lr, #32 /* we corrupted r12, recompute it */ - add r2, r2, #32 - cmp r2, #4 - blo partial_word_tail - -1: ldr r5, [r1], #4 - sub r2, r2, #4 - orr r4, r3, r5, lsl lr - mov r3, r5, lsr r12 - str r4, [r0], #4 - cmp r2, #4 - bhs 1b + rsb r12, lr, #32 /* we corrupted r12, recompute it */ + add r2, r2, #32 + cmp r2, #4 + blo partial_word_tail + +1: ldr r5, [r1], #4 + sub r2, r2, #4 + orr r4, r3, r5, lsl lr + mov r3, r5, lsr r12 + str r4, [r0], #4 + cmp r2, #4 + bhs 1b partial_word_tail: /* we have a partial word in the input buffer */ - movs r5, lr, lsl #(31-3) + movs r5, lr, lsl #(31-3) .word 0x44c03001 /* strbmi r3, [r0], #1 */ - movmi r3, r3, lsr #8 - .word 0x24c03001 /* strbcs r3, [r0], #1 */ - movcs r3, r3, lsr #8 - .word 0x24c03001 /* strbcs r3, [r0], #1 */ + movmi r3, r3, lsr #8 + .word 0x24c03001 /* strbcs r3, [r0], #1 */ + movcs r3, r3, lsr #8 + .word 0x24c03001 /* strbcs r3, [r0], #1 */ - /* Refill spilled registers from the stack. Don't update sp. */ - ldmfd sp, {r5-r11} + /* Refill spilled registers from the stack. Don't update sp. */ + ldmfd sp, {r5-r11} copy_last_3_and_return: - movs r2, r2, lsl #31 /* copy remaining 0, 1, 2 or 3 bytes */ - .word 0x44d12001 /* ldrbmi r2, [r1], #1 */ - .word 0x24d13001 /* ldrbcs r3, [r1], #1 */ - .word 0x25d1c000 /* ldrbcs r12,[r1] */ - .word 0x44c02001 /* strbmi r2, [r0], #1 */ - .word 0x24c03001 /* strbcs r3, [r0], #1 */ - .word 0x25c0c000 /* strbcs r12,[r0] */ - - /* we're done! restore sp and spilled registers and return */ - add sp, sp, #28 - ldmfd sp!, {r0, r4, lr} - tst lr, #1 - moveq pc, lr - bx lr + movs r2, r2, lsl #31 /* copy remaining 0, 1, 2 or 3 bytes */ + .word 0x44d12001 /* ldrbmi r2, [r1], #1 */ + .word 0x24d13001 /* ldrbcs r3, [r1], #1 */ + .word 0x25d1c000 /* ldrbcs r12,[r1] */ + .word 0x44c02001 /* strbmi r2, [r0], #1 */ + .word 0x24c03001 /* strbcs r3, [r0], #1 */ + .word 0x25c0c000 /* strbcs r12,[r0] */ + + /* we're done! restore sp and spilled registers and return */ + add sp, sp, #28 + ldmfd sp!, {r0, r4, lr} + tst lr, #1 + moveq pc, lr + bx lr |