diff options
Diffstat (limited to 'sysdeps/powerpc/powerpc32/power4/memcpy.S')
| -rw-r--r-- | sysdeps/powerpc/powerpc32/power4/memcpy.S | 96 |
1 files changed, 48 insertions, 48 deletions
diff --git a/sysdeps/powerpc/powerpc32/power4/memcpy.S b/sysdeps/powerpc/powerpc32/power4/memcpy.S index a11407c3d4..d6c6e6851a 100644 --- a/sysdeps/powerpc/powerpc32/power4/memcpy.S +++ b/sysdeps/powerpc/powerpc32/power4/memcpy.S @@ -24,10 +24,10 @@ /* __ptr_t [r3] memcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]); Returns 'dst'. - Memcpy handles short copies (< 32-bytes) using a binary move blocks - (no loops) of lwz/stw. The tail (remaining 1-3) bytes is handled - with the appropriate combination of byte and halfword load/stores. - There is minimal effort to optimize the alignment of short moves. + Memcpy handles short copies (< 32-bytes) using a binary move blocks + (no loops) of lwz/stw. The tail (remaining 1-3) bytes is handled + with the appropriate combination of byte and halfword load/stores. + There is minimal effort to optimize the alignment of short moves. Longer moves (>= 32-bytes) justify the effort to get at least the destination word (4-byte) aligned. Further optimization is @@ -43,7 +43,7 @@ EALIGN (BP_SYM (memcpy), 5, 0) stw 30,20(1) cfi_offset(30,(20-32)) mr 30,3 - cmplwi cr1,5,31 + cmplwi cr1,5,31 stw 31,24(1) cfi_offset(31,(24-32)) neg 0,3 @@ -52,13 +52,13 @@ EALIGN (BP_SYM (memcpy), 5, 0) clrlwi 10,4,30 /* check alignment of src. */ cmplwi cr6,5,8 ble- cr1,.L2 /* If move < 32 bytes use short move code. */ - cmplw cr6,10,11 + cmplw cr6,10,11 mr 12,4 srwi 9,5,2 /* Number of full words remaining. */ mtcrf 0x01,0 mr 31,5 beq .L0 - + subf 31,0,5 /* Move 0-3 bytes as needed to get the destination word aligned. */ 1: bf 31,2f @@ -72,17 +72,17 @@ EALIGN (BP_SYM (memcpy), 5, 0) sth 6,0(3) addi 3,3,2 0: - clrlwi 10,12,30 /* check alignment of src again. */ + clrlwi 10,12,30 /* check alignment of src again. */ srwi 9,31,2 /* Number of full words remaining. */ - - /* Copy words from source to destination, assuming the destination is + + /* Copy words from source to destination, assuming the destination is aligned on a word boundary. At this point we know there are at least 25 bytes left (32-7) to copy. - The next step is to determine if the source is also word aligned. + The next step is to determine if the source is also word aligned. If not branch to the unaligned move code at .L6. which uses a load, shift, store strategy. - + Otherwise source and destination are word aligned, and we can use the optimized word copy loop. */ .L0: @@ -92,16 +92,16 @@ EALIGN (BP_SYM (memcpy), 5, 0) /* Move words where destination and source are word aligned. Use an unrolled loop to copy 4 words (16-bytes) per iteration. - If the copy is not an exact multiple of 16 bytes, 1-3 + If the copy is not an exact multiple of 16 bytes, 1-3 words are copied as needed to set up the main loop. After - the main loop exits there may be a tail of 1-3 bytes. These bytes are + the main loop exits there may be a tail of 1-3 bytes. These bytes are copied a halfword/byte at a time as needed to preserve alignment. */ srwi 8,31,4 /* calculate the 16 byte loop count */ cmplwi cr1,9,4 cmplwi cr6,11,0 mr 11,12 - + bf 30,1f lwz 6,0(12) lwz 7,4(12) @@ -112,7 +112,7 @@ EALIGN (BP_SYM (memcpy), 5, 0) addi 10,3,8 bf 31,4f lwz 0,8(12) - stw 0,8(3) + stw 0,8(3) blt cr1,3f addi 11,12,12 addi 10,3,12 @@ -126,7 +126,7 @@ EALIGN (BP_SYM (memcpy), 5, 0) addi 11,12,4 stw 6,0(3) addi 10,3,4 - + .align 4 4: lwz 6,0(11) @@ -140,14 +140,14 @@ EALIGN (BP_SYM (memcpy), 5, 0) addi 11,11,16 addi 10,10,16 bdnz 4b -3: +3: clrrwi 0,31,2 mtcrf 0x01,31 beq cr6,0f .L9: add 3,3,0 add 12,12,0 - + /* At this point we have a tail of 0-3 bytes and we know that the destination is word aligned. */ 2: bf 30,1f @@ -165,27 +165,27 @@ EALIGN (BP_SYM (memcpy), 5, 0) lwz 31,24(1) addi 1,1,32 blr - -/* Copy up to 31 bytes. This is divided into two cases 0-8 bytes and - 9-31 bytes. Each case is handled without loops, using binary - (1,2,4,8) tests. - + +/* Copy up to 31 bytes. This is divided into two cases 0-8 bytes and + 9-31 bytes. Each case is handled without loops, using binary + (1,2,4,8) tests. + In the short (0-8 byte) case no attempt is made to force alignment - of either source or destination. The hardware will handle the - unaligned load/stores with small delays for crossing 32- 64-byte, and + of either source or destination. The hardware will handle the + unaligned load/stores with small delays for crossing 32- 64-byte, and 4096-byte boundaries. Since these short moves are unlikely to be - unaligned or cross these boundaries, the overhead to force + unaligned or cross these boundaries, the overhead to force alignment is not justified. - + The longer (9-31 byte) move is more likely to cross 32- or 64-byte boundaries. Since only loads are sensitive to the 32-/64-byte - boundaries it is more important to align the source than the + boundaries it is more important to align the source than the destination. If the source is not already word aligned, we first - move 1-3 bytes as needed. While the destination and stores may + move 1-3 bytes as needed. While the destination and stores may still be unaligned, this is only an issue for page (4096 byte - boundary) crossing, which should be rare for these short moves. - The hardware handles this case automatically with a small delay. */ - + boundary) crossing, which should be rare for these short moves. + The hardware handles this case automatically with a small delay. */ + .align 4 .L2: mtcrf 0x01,5 @@ -248,11 +248,11 @@ EALIGN (BP_SYM (memcpy), 5, 0) lwz 6,0(12) addi 12,12,4 stw 6,0(3) - addi 3,3,4 + addi 3,3,4 2: /* Move 2-3 bytes. */ bf 30,1f lhz 6,0(12) - sth 6,0(3) + sth 6,0(3) bf 31,0f lbz 7,2(12) stb 7,2(3) @@ -292,7 +292,7 @@ EALIGN (BP_SYM (memcpy), 5, 0) 6: bf 30,5f lhz 7,4(4) - sth 7,4(3) + sth 7,4(3) bf 31,0f lbz 8,6(4) stb 8,6(3) @@ -301,7 +301,7 @@ EALIGN (BP_SYM (memcpy), 5, 0) addi 1,1,32 blr .align 4 -5: +5: bf 31,0f lbz 6,4(4) stb 6,4(3) @@ -318,15 +318,15 @@ EALIGN (BP_SYM (memcpy), 5, 0) /* Copy words where the destination is aligned but the source is not. Use aligned word loads from the source, shifted to realign - the data, to allow aligned destination stores. + the data, to allow aligned destination stores. Use an unrolled loop to copy 4 words (16-bytes) per iteration. A single word is retained for storing at loop exit to avoid walking off the end of a page within the loop. - If the copy is not an exact multiple of 16 bytes, 1-3 + If the copy is not an exact multiple of 16 bytes, 1-3 words are copied as needed to set up the main loop. After - the main loop exits there may be a tail of 1-3 bytes. These bytes are + the main loop exits there may be a tail of 1-3 bytes. These bytes are copied a halfword/byte at a time as needed to preserve alignment. */ - + cmplwi cr6,11,0 /* are there tail bytes left ? */ subf 5,10,12 /* back up src pointer to prev word alignment */ @@ -381,8 +381,8 @@ EALIGN (BP_SYM (memcpy), 5, 0) .align 4 4: /* copy 16 bytes at a time */ - slw 0,6,10 - srw 8,7,9 + slw 0,6,10 + srw 8,7,9 or 0,0,8 lwz 6,0(5) stw 0,0(4) @@ -391,13 +391,13 @@ EALIGN (BP_SYM (memcpy), 5, 0) or 0,0,8 lwz 7,4(5) stw 0,4(4) - slw 0,6,10 - srw 8,7,9 + slw 0,6,10 + srw 8,7,9 or 0,0,8 lwz 6,8(5) stw 0,8(4) slw 0,7,10 - srw 8,6,9 + srw 8,6,9 or 0,0,8 lwz 7,12(5) stw 0,12(4) @@ -406,8 +406,8 @@ EALIGN (BP_SYM (memcpy), 5, 0) bdnz+ 4b 8: /* calculate and store the final word */ - slw 0,6,10 - srw 8,7,9 + slw 0,6,10 + srw 8,7,9 or 0,0,8 stw 0,0(4) 3: |