/* __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
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
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
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:
/* 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)
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
addi 11,12,4
stw 6,0(3)
addi 10,3,4
-
+
.align 4
4:
lwz 6,0(11)
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
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
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)
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)
addi 1,1,32
blr
.align 4
-5:
+5:
bf 31,0f
lbz 6,4(4)
stb 6,4(3)
/* 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 */
.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)
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)
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:
#define rSTR2 r4 /* second string arg */
#define rN r5 /* max string length */
/* Note: The Bounded pointer support in this code is broken. This code
- was inherited from PPC32 and that support was never completed.
+ was inherited from PPC32 and that support was never completed.
Current PPC gcc does not support -fbounds-check or -fbounded-pointers. */
#define rWORD1 r6 /* current word in s1 */
#define rWORD2 r7 /* current word in s2 */
clrlwi rN, rN, 30
addi rFEFE, rFEFE, -0x101
addi r7F7F, r7F7F, 0x7f7f
- cmplwi cr1, rN, 0
+ cmplwi cr1, rN, 0
beq L(unaligned)
mtctr rTMP /* Power4 wants mtctr 1st in dispatch group. */
lwz rWORD2, 0(rSTR2)
b L(g1)
-L(g0):
+L(g0):
lwzu rWORD1, 4(rSTR1)
bne- cr1, L(different)
lwzu rWORD2, 4(rSTR2)
and. rTMP, rTMP, rNEG
cmpw cr1, rWORD1, rWORD2
beq+ L(g0)
-
+
/* OK. We've hit the end of the string. We need to be careful that
we don't compare two strings as different because of gunk beyond
the end of the strings... */
-
+
L(endstring):
and rTMP, r7F7F, rWORD1
beq cr1, L(equal)
b L(u1)
L(u3): sub rRTN, rWORD3, rWORD4
- blr
+ blr
L(u4): sub rRTN, rWORD1, rWORD2
blr
L(ux):
blr
END (BP_SYM (strncmp))
libc_hidden_builtin_def (strncmp)
-
#define rSTR2 r4 /* second string arg */
#define rN r5 /* max string length */
/* Note: The Bounded pointer support in this code is broken. This code
- was inherited from PPC32 and that support was never completed.
+ was inherited from PPC32 and that support was never completed.
Current PPC gcc does not support -fbounds-check or -fbounded-pointers. */
#define rWORD1 r6 /* current word in s1 */
#define rWORD2 r7 /* current word in s2 */
clrlwi rN, rN, 30
addi rFEFE, rFEFE, -0x101
addi r7F7F, r7F7F, 0x7f7f
- cmplwi cr1, rN, 0
+ cmplwi cr1, rN, 0
beq L(unaligned)
mtctr rTMP /* Power4 wants mtctr 1st in dispatch group. */
lwz rWORD2, 0(rSTR2)
b L(g1)
-L(g0):
+L(g0):
lwzu rWORD1, 4(rSTR1)
bne- cr1, L(different)
lwzu rWORD2, 4(rSTR2)
and. rTMP, rTMP, rNEG
cmpw cr1, rWORD1, rWORD2
beq+ L(g0)
-
+
/* OK. We've hit the end of the string. We need to be careful that
we don't compare two strings as different because of gunk beyond
the end of the strings... */
-
+
L(endstring):
and rTMP, r7F7F, rWORD1
beq cr1, L(equal)
lbzu rWORD1, 1(rSTR1)
bne+ cr1, L(u0)
-L(u2): lbzu rWORD1, -1(rSTR1)
+L(u2): lbzu rWORD1, -1(rSTR1)
L(u3): sub rRTN, rWORD1, rWORD2
blr
END (BP_SYM (strncmp))
/* __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.
The 64-bit implementations of POWER3 and POWER4 do a reasonable job
of handling unligned load/stores that do not cross 32-byte boundries.
clrldi 10,4,61 /* check alignement of src. */
cmpldi cr6,5,8
ble- cr1,.L2 /* If move < 32 bytes use short move code. */
- cmpld cr6,10,11
+ cmpld cr6,10,11
mr 12,4
srdi 9,5,3 /* Number of full double words remaining. */
mtcrf 0x01,0
mr 31,5
beq .L0
-
+
subf 31,0,5
/* Move 0-7 bytes as needed to get the destination doubleword alligned. */
1: bf 31,2f
stw 6,0(3)
addi 3,3,4
0:
- clrldi 10,12,61 /* check alignement of src again. */
+ clrldi 10,12,61 /* check alignement of src again. */
srdi 9,31,3 /* Number of full double words remaining. */
-
+
/* Copy doublewords from source to destination, assumpting the
destination is aligned on a doubleword 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 doubleword aligned.
+ The next step is to determine if the source is also doubleword aligned.
If not branch to the unaligned move code at .L6. which uses
a load, shift, store strategy.
-
+
Otherwise source and destination are doubleword aligned, and we can
the optimized doubleword copy loop. */
.L0:
/* Move doublewords where destination and source are DW aligned.
Use a unrolled loop to copy 4 doubleword (32-bytes) per iteration.
- If the copy is not an exact multiple of 32 bytes, 1-3
+ If the copy is not an exact multiple of 32 bytes, 1-3
doublewords are copied as needed to set up the main loop. After
- the main loop exits there may be a tail of 1-7 bytes. These byte are
+ the main loop exits there may be a tail of 1-7 bytes. These byte are
copied a word/halfword/byte at a time as needed to preserve alignment. */
srdi 8,31,5
cmpldi cr1,9,4
cmpldi cr6,11,0
mr 11,12
-
+
bf 30,1f
ld 6,0(12)
ld 7,8(12)
addi 10,3,16
bf 31,4f
ld 0,16(12)
- std 0,16(3)
+ std 0,16(3)
blt cr1,3f
addi 11,12,24
addi 10,3,24
addi 11,12,8
std 6,0(3)
addi 10,3,8
-
+
.align 4
4:
ld 6,0(11)
std 0,24(10)
addi 10,10,32
bdnz 4b
-3:
+3:
rldicr 0,31,0,60
mtcrf 0x01,31
.L9:
add 3,3,0
add 12,12,0
-
+
/* At this point we have a tail of 0-7 bytes and we know that the
destiniation is double word aligned. */
4: bf 29,2f
ld 31,-8(1)
ld 3,-16(1)
blr
-
-/* Copy up to 31 bytes. This 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 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 then the
+ boundaries it is more important to align the source then the
destination. If the source is not already word aligned, we first
- move 1-3 bytes as needed. Since we are only word aligned we don't
- use double word load/stores to insure that all loads are aligned.
+ move 1-3 bytes as needed. Since we are only word aligned we don't
+ use double word load/stores to insure that all loads are aligned.
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. */
-
+ case automatically with a small delay. */
+
.align 4
.L2:
mtcrf 0x01,5
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)
mr 12,4
bne cr6,4f
/* Would have liked to use use ld/std here but the 630 processors are
- slow for load/store doubles that are not at least word aligned.
+ slow for load/store doubles that are not at least word aligned.
Unaligned Load/Store word execute with only a 1 cycle penaltity. */
lwz 6,0(4)
lwz 7,4(4)
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)
ld 3,-16(1)
blr
.align 4
-5:
+5:
bf 31,0f
lbz 6,4(4)
stb 6,4(3)
#define rSTR2 r4 /* second string arg */
#define rN r5 /* max string length */
/* Note: The Bounded pointer support in this code is broken. This code
- was inherited from PPC32 and that support was never completed.
+ was inherited from PPC32 and that support was never completed.
Current PPC gcc does not support -fbounds-check or -fbounded-pointers. */
#define rWORD1 r6 /* current word in s1 */
#define rWORD2 r7 /* current word in s2 */
clrldi rN, rN, 61
addi rFEFE, rFEFE, -0x101
addi r7F7F, r7F7F, 0x7f7f
- cmpldi cr1, rN, 0
+ cmpldi cr1, rN, 0
beq L(unaligned)
mtctr rTMP /* Power4 wants mtctr 1st in dispatch group. */
add rFEFE, rFEFE, rTMP
b L(g1)
-L(g0):
+L(g0):
ldu rWORD1, 8(rSTR1)
bne- cr1, L(different)
ldu rWORD2, 8(rSTR2)
and. rTMP, rTMP, rNEG
cmpd cr1, rWORD1, rWORD2
beq+ L(g0)
-
+
/* OK. We've hit the end of the string. We need to be careful that
we don't compare two strings as different because of gunk beyond
the end of the strings... */
-
+
L(endstring):
and rTMP, r7F7F, rWORD1
beq cr1, L(equal)
b L(u1)
L(u3): sub rRTN, rWORD3, rWORD4
- blr
+ blr
L(u4): sub rRTN, rWORD1, rWORD2
blr
L(ux):
blr
END (BP_SYM (strncmp))
libc_hidden_builtin_def (strncmp)
-
#define rSTR1 r3 /* first string arg */
#define rSTR2 r4 /* second string arg */
/* Note: The Bounded pointer support in this code is broken. This code
- was inherited from PPC32 and that support was never completed.
+ was inherited from PPC32 and that support was never completed.
Current PPC gcc does not support -fbounds-check or -fbounded-pointers.
These artifacts are left in the code as a reminder in case we need
bounded pointer support in the future. */
#define rTMP r0
#define rRTN r3 /* incoming DEST arg preserved as result */
/* Note. The Bounded pointer support in this code is broken. This code
- was inherited from PPC32 and that support was never completed.
+ was inherited from PPC32 and that support was never completed.
Current PPC gcc does not support -fbounds-check or -fbounded-pointers.
These artifacts are left in the code as a reminder in case we need
bounded pointer support in the future. */
git://sourceware.org / glibc.git / commitdiff