diff options
Diffstat (limited to 'libc/string/sh/sh4/memcpy.S')
| -rw-r--r-- | libc/string/sh/sh4/memcpy.S | 107 |
1 files changed, 53 insertions, 54 deletions
diff --git a/libc/string/sh/sh4/memcpy.S b/libc/string/sh/sh4/memcpy.S index 8621e4a5c..b02d614df 100644 --- a/libc/string/sh/sh4/memcpy.S +++ b/libc/string/sh/sh4/memcpy.S @@ -21,7 +21,7 @@ ! ! GHIJ KLMN OPQR --> ...G HIJK LMNO PQR. ! - + ! Size is 16 or greater, and may have trailing bytes .balign 32 @@ -40,7 +40,7 @@ ! 6 cycles, 4 bytes per iteration 3: mov.l @(r0,r5),r1 ! 21 LS (latency=2) ! NMLK mov r7, r3 ! 5 MT (latency=0) ! RQPO - + cmp/hi r2,r0 ! 57 MT shll16 r3 ! 103 EX @@ -49,7 +49,7 @@ shlr8 r6 ! 106 EX ! xNML mov r1, r7 ! 5 MT (latency=0) - + or r6,r3 ! 82 EX ! ONML bt/s 3b ! 109 BR @@ -82,7 +82,7 @@ 8: cmp/hi r2,r0 ! 57 MT mov.b @(r0,r5),r1 ! 20 LS (latency=2) - + bt/s 8b ! 109 BR mov.b r1,@-r0 ! 29 LS @@ -90,11 +90,11 @@ 9: rts nop - + ! ! GHIJ KLMN OPQR --> .GHI JKLM NOPQ R... ! - + ! Size is 16 or greater, and may have trailing bytes .balign 32 @@ -113,7 +113,7 @@ ! 6 cycles, 4 bytes per iteration 3: mov.l @(r0,r5),r1 ! 21 LS (latency=2) ! NMLK mov r7, r3 ! 5 MT (latency=0) ! RQPO - + cmp/hi r2,r0 ! 57 MT shll8 r3 ! 102 EX ! QPOx @@ -122,7 +122,7 @@ shlr8 r6 ! 106 EX ! xxxN mov r1, r7 ! 5 MT (latency=0) - + or r6,r3 ! 82 EX ! QPON bt/s 3b ! 109 BR @@ -150,14 +150,13 @@ 8: cmp/hi r2,r0 ! 57 MT mov.b @(r0,r5),r1 ! 20 LS (latency=2) - + bt/s 8b ! 109 BR mov.b r1,@-r0 ! 29 LS 9: rts nop - ENTRY(memcpy) @@ -170,7 +169,7 @@ ENTRY(memcpy) ! r0 --> [ ... ] r0+r5 --> [ ... ] ! ! - + ! Short circuit the common case of src, dst and len being 32 bit aligned ! and test for zero length move @@ -203,7 +202,7 @@ ENTRY(memcpy) ! 36 82 49-50 66-70 80-85 ! However the penalty for getting it 'wrong' is much higher for long word ! aligned data (and this is more common), so use a value of 16. - + cmp/gt r6,r1 ! 56 MT add #-1,r5 ! 50 EX @@ -214,7 +213,7 @@ ENTRY(memcpy) mov.b @(r0,r5),r1 ! 20 LS (latency=2) bf/s 4f ! 111 BR - + add #-1,r3 ! 50 EX tst r6, r6 ! 86 MT @@ -249,7 +248,7 @@ ENTRY(memcpy) bt/s 2f ! 111 BR and r0,r3 ! 78 EX - ! 3 cycles, 1 byte per iteration + ! 3 cycles, 1 byte per iteration 1: dt r3 ! 67 EX mov.b @(r0,r5),r1 ! 19 LS (latency=2) @@ -259,10 +258,10 @@ ENTRY(memcpy) mov.b r1,@-r0 ! 28 LS 2: add #1, r5 ! 79 EX - + ! Now select the appropriate bulk transfer code based on relative ! alignment of src and dst. - + mov r0, r3 ! 5 MT (latency=0) mov r5, r0 ! 5 MT (latency=0) @@ -272,7 +271,7 @@ ENTRY(memcpy) mov #64, r7 ! 6 EX ! bit 0 clear - + cmp/ge r7, r6 ! 55 MT bt/s 2f ! 111 BR @@ -291,7 +290,7 @@ ENTRY(memcpy) bra .Lcase2b nop - + ! bit 0 set 1: tst #2, r0 ! 87 MT @@ -300,7 +299,7 @@ ENTRY(memcpy) bra .Lcase3 nop - + ! ! GHIJ KLMN OPQR --> GHIJ KLMN OPQR @@ -308,7 +307,7 @@ ENTRY(memcpy) ! src, dst and size are all long word aligned ! size is non-zero - + .balign 32 .Lcase00: mov #64, r1 ! 6 EX @@ -321,7 +320,7 @@ ENTRY(memcpy) shlr2 r6 ! 105 EX shlr r6 ! 104 EX - mov.l @(r0, r5), r1 ! 21 LS (latency=2) + mov.l @(r0, r5), r1 ! 21 LS (latency=2) bf/s 4f ! 111 BR add #-8, r3 ! 50 EX @@ -345,7 +344,7 @@ ENTRY(memcpy) 5: rts nop - + ! Size is 16 or greater and less than 64, but may have trailing bytes .balign 32 @@ -353,7 +352,7 @@ ENTRY(memcpy) add #-4, r5 ! 50 EX mov r4, r7 ! 5 MT (latency=0) - mov.l @(r0, r5), r1 ! 21 LS (latency=2) + mov.l @(r0, r5), r1 ! 21 LS (latency=2) mov #4, r2 ! 6 EX add #11, r7 ! 50 EX @@ -379,7 +378,7 @@ ENTRY(memcpy) ! Copy the final 0-3 bytes add #3,r5 ! 50 EX - + cmp/eq r0, r4 ! 54 MT add #-10, r7 ! 50 EX @@ -388,7 +387,7 @@ ENTRY(memcpy) ! 3 cycles, 1 byte per iteration 1: mov.b @(r0,r5),r1 ! 19 LS cmp/hi r7,r0 ! 57 MT - + bt/s 1b ! 111 BR mov.b r1,@-r0 ! 28 LS @@ -429,10 +428,10 @@ ENTRY(memcpy) add #8, r3 ! 50 EX tst #0x18, r0 ! 87 MT - + bt/s 1f ! 109 BR mov.l r1,@-r0 ! 30 LS - + ! 4 cycles, 2 long words per iteration 3: mov.l @(r0, r5), r1 ! 21 LS (latency=2) @@ -452,7 +451,7 @@ ENTRY(memcpy) ! We could do this with the four scratch registers, but if src ! and dest hit the same cache line, this will thrash, so make ! use of additional registers. - ! + ! ! We also need r0 as a temporary (for movca), so 'undo' the invariant: ! r5: src (was r0+r5) ! r1: dest (was r0) @@ -461,14 +460,14 @@ ENTRY(memcpy) ! 1: mov.l r8, @-r15 ! 30 LS add r0, r5 ! 49 EX - + mov.l r9, @-r15 ! 30 LS mov r0, r1 ! 5 MT (latency=0) - + mov.l r10, @-r15 ! 30 LS add #-0x1c, r5 ! 50 EX - - mov.l r11, @-r15 ! 30 LS + + mov.l r11, @-r15 ! 30 LS ! 16 cycles, 32 bytes per iteration 2: mov.l @(0x00,r5),r0 ! 18 LS (latency=2) @@ -512,10 +511,10 @@ ENTRY(memcpy) sub r4, r1 ! 75 EX (len remaining) ! number of trailing bytes is non-zero - ! + ! ! invariants restored (r5 already decremented by 4) ! also r1=num bytes remaining - + mov #4, r2 ! 6 EX mov r4, r7 ! 5 MT (latency=0) @@ -525,7 +524,7 @@ ENTRY(memcpy) bf/s 5f ! 108 BR add #11, r7 ! 50 EX - mov.l @(r0, r5), r6 ! 21 LS (latency=2) + mov.l @(r0, r5), r6 ! 21 LS (latency=2) tst r2, r1 ! 86 MT mov r5, r3 ! 5 MT (latency=0) @@ -555,11 +554,11 @@ ENTRY(memcpy) bt 9f ! 110 BR add #3,r5 ! 50 EX - + ! 3 cycles, 1 byte per iteration 1: mov.b @(r0,r5),r1 ! 19 LS cmp/hi r7,r0 ! 57 MT - + bt/s 1b ! 111 BR mov.b r1,@-r0 ! 28 LS @@ -569,7 +568,7 @@ ENTRY(memcpy) ! ! GHIJ KLMN OPQR --> ..GH IJKL MNOP QR.. ! - + .balign 32 .Lcase2: ! Size is 16 or greater and less then 64, but may have trailing bytes @@ -610,21 +609,21 @@ ENTRY(memcpy) cmp/eq r3, r0 ! 54 MT add #0x1f, r2 ! 50 EX - + add #-2, r5 ! 50 EX bt/s 1f ! 110 BR and r1, r2 ! 78 EX - + ! Copy a short word one at a time until we are cache line aligned ! Normal values: r0, r2, r3, r4 ! Unused: r1, r6, r7 ! Mod: r5 (=r5-2) ! add #2, r3 ! 50 EX - + 2: mov.w @(r0,r5),r1 ! 20 LS (latency=2) cmp/eq r3,r0 ! 54 MT - + bf/s 2b ! 111 BR mov.w r1,@-r0 ! 29 LS @@ -637,7 +636,7 @@ ENTRY(memcpy) ! We could do this with the four scratch registers, but if src ! and dest hit the same cache line, this will thrash, so make ! use of additional registers. - ! + ! ! We also need r0 as a temporary (for movca), so 'undo' the invariant: ! r5: src (was r0+r5) ! r1: dest (was r0) @@ -646,16 +645,16 @@ ENTRY(memcpy) ! 1: mov.l r8, @-r15 ! 30 LS add r0, r5 ! 49 EX - + mov.l r9, @-r15 ! 30 LS mov r0, r1 ! 5 MT (latency=0) - + mov.l r10, @-r15 ! 30 LS add #-0x1e, r5 ! 50 EX - - mov.l r11, @-r15 ! 30 LS - - mov.l r12, @-r15 ! 30 LS + + mov.l r11, @-r15 ! 30 LS + + mov.l r12, @-r15 ! 30 LS ! 17 cycles, 32 bytes per iteration #ifdef __LITTLE_ENDIAN__ @@ -692,7 +691,7 @@ ENTRY(memcpy) xtrct r12, r11 ! 48 EX mov.l r6, @(0x08,r1) ! 33 LS - + mov.l r7, @(0x0c,r1) ! 33 LS mov.l r8, @(0x10,r1) ! 33 LS @@ -741,7 +740,7 @@ ENTRY(memcpy) mov.l r6, @(0x18,r1) ! 33 LS xtrct r12, r11 ! 48 EX - + mov.l r7, @(0x14,r1) ! 33 LS mov.l r8, @(0x10,r1) ! 33 LS @@ -772,7 +771,7 @@ ENTRY(memcpy) 1: mov.l @r15+, r8 ! 15 LS add #0x1e, r5 ! 50 EX - + ! Finish off a short word at a time ! r5 must be invariant - 2 10: mov r4,r2 ! 5 MT (latency=0) @@ -782,7 +781,7 @@ ENTRY(memcpy) bf/s 1f ! 109 BR add #2, r2 ! 50 EX - + 3: mov.w @(r0,r5),r1 ! 20 LS cmp/hi r2,r0 ! 57 MT @@ -790,7 +789,7 @@ ENTRY(memcpy) mov.w r1,@-r0 ! 29 LS 1: - + ! ! Finally, copy the last byte if necessary cmp/eq r4,r0 ! 54 MT |