Module Name: src
Committed By: matt
Date: Sun Dec 23 03:44:24 UTC 2012
Modified Files:
src/sys/arch/arm/arm: cpu_in_cksum_buffer.S
Log Message:
Generate the trailing mask at the start and put it and the starting address
in r11/r10 and use them as need. Always round the length endian address to
a word boundary. Unconditionally apply the trailing mask at the end since
it's a cheap op.
To generate a diff of this commit:
cvs rdiff -u -r1.6 -r1.7 src/sys/arch/arm/arm/cpu_in_cksum_buffer.S
Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.
Modified files:
Index: src/sys/arch/arm/arm/cpu_in_cksum_buffer.S
diff -u src/sys/arch/arm/arm/cpu_in_cksum_buffer.S:1.6 src/sys/arch/arm/arm/cpu_in_cksum_buffer.S:1.7
--- src/sys/arch/arm/arm/cpu_in_cksum_buffer.S:1.6 Sat Dec 22 08:12:26 2012
+++ src/sys/arch/arm/arm/cpu_in_cksum_buffer.S Sun Dec 23 03:44:24 2012
@@ -29,7 +29,7 @@
#include <machine/asm.h>
-RCSID("$NetBSD: cpu_in_cksum_buffer.S,v 1.6 2012/12/22 08:12:26 matt Exp $")
+RCSID("$NetBSD: cpu_in_cksum_buffer.S,v 1.7 2012/12/23 03:44:24 matt Exp $")
/*
* Special note:
@@ -67,27 +67,47 @@ ENTRY(cpu_in_cksum_buffer)
#if defined(_ARM_ARCH_DWORD_OK) && !defined(__OPTIMIZE_SIZE__)
pld [r0] /* prefetch the first data */
#endif
- mov ip, r2 /* initialize accumulator */
- adds ip, ip, #0 /* clear carry */
+ adds ip, r2, #0 /* initialize accumulator/clear carry */
teq r1, #0 /* did we get passed a zero length? */
- beq .Lfold_nopop /* fold the checksum */
- ands r2, r0, #7 /* test for dword alignment */
+ beq .Lfold /* fold the checksum */
+ add r2, r0, r1 /* point r2 just past end */
+ push {r4-r5,r10-r11} /* save registers */
+ mvn r11, #0 /* initialize trailing mask */
+ ands r3, r2, #3 /* limit to a word */
+ beq 1f /* no trailing bytes? */
+ /*
+ * This buffer doesn't end on a word boundary so create a mask
+ * to discard the unneeded bytes in the last word and then round
+ * up the length and ending address to a word boundary.
+ */
+ rsb r3, r3, #4 /* find out how many bytes to clear */
+ add r2, r2, r3 /* align to word boundary */
+ add r1, r1, r3 /* align to word boundary */
+ mov r3, r3, lsl #3 /* bytes -> bits */
+#ifdef __ARMEL__
+ mov r11, r11, lsr r3 /* replace with zero bits */
+#else
+ mov r11, r11, lsl r3 /* replace with zero bits */
+#endif
+1:
+ ands r10, r0, #7 /* test for dword alignment */
bne .Ldword_misaligned /* no, fixup non dword aligned */
- push {r4-r5} /* save temporaries */
- sub RLO, r1, #1 /* subtract 1 from length */
- bics RLO, RLO, #3 /* more than 1 word? */
- beq .Lfinal_word /* no, just load final word */
- add r2, r1, r0 /* point r2 just past end */
+ /*
+ * If the (now rounded up) length is 4, then only bit 2 will be set.
+ * So if we clear that bit and the result is 0, then the length must
+ * have been 4.
+ */
+ bics RLO, r1, #4 /* more than 1 word? */
+ beq .Lfinal_word_load /* no, just load final word */
LOAD_DWORD_INTO_R4(r0) /* load first dword */
- sub r1, r1, #8 /* we've read one dword */
#if defined(_ARM_ARCH_DWORD_OK) && !defined(__OPTIMIZE_SIZE__)
pld [r0, #32] /* prefetch data */
#endif
.p2align 3
.Ldword_aligned_noload:
- add r1, r1, #3 /* round up word length */
+ sub r1, r2, r0 /* how much is remaining? */
bics r3, r1, #15 /* at least 16 bytes to do? */
- beq 3f
+ beq .Lfinal_words /* no, but we have at least 1 word */
push {r6-r7}
#if !defined(__OPTIMIZE_SIZE__)
bics r3, r1, #63 /* at least 64 bytes to do? */
@@ -125,7 +145,6 @@ ENTRY(cpu_in_cksum_buffer)
adcs ip, ip, r7
sub r1, r2, r0 /* find how much is left */
- add r1, r1, #3 /* round up word length */
#if !defined(__OPTIMIZE_SIZE__)
bics r3, r1, #63 /* at least 64 bytes to do? */
bne .Lloop64 /* yes, run the loop again */
@@ -138,69 +157,47 @@ ENTRY(cpu_in_cksum_buffer)
pop {r6-r7} /* done with these so restore them */
-3: sub r1, r2, r0 /* find how much is left */
teq r1, #0 /* how much left?? */
- beq .Lfinal_add_one_dword /* = 0? do the final add */
- bmi .Lfinal_dword_noload /* < 0? trim last word */
+ beq .Ladd_final_dword /* = 0? do the final add */
+.Lfinal_words:
/*
- * We have from 1-12 bytes left to do.
+ * We have 1 to 3 words left to load.
*/
- add r3, r1, #3 /* round up word length */
- tst r3, #8 /* at least one dword (5+ bytes)? */
- beq .Lfinal_word /* no, deal with the final word. */
+ tst r1, #8 /* at least one dword (5+ bytes)? */
+ beq .Lfinal_word_load /* no, deal with the final word. */
/*
- * We have at least 5 bytes so we need to load at least 8 (maybe 12)
- * so load 8.
+ * We have at least 8 bytes left so accumulate the pending dword
+ * and then load the next dword.
*/
adcs ip, ip, r4
adcs ip, ip, r5
LOAD_DWORD_INTO_R4(r0)
- sub r1, r1, #8 /* subtract dword from length */
- teq r1, #0 /* how much left?? */
- beq .Lfinal_add_one_dword /* = 0? do the final add */
- bmi .Lfinal_dword_noload /* < 0? trim last word */
-.Lfinal_word:
+ /*
+ * At this point r1 is either 8 or 12 so we can just clear bit 3
+ * to see if we have one more word to read.
+ */
+ bics r1, r1, #8 /* subtract dword from length */
+ beq .Ladd_final_dword /* = 0? do the final add */
+.Lfinal_word_load:
/*
* Finally we are at the word to load.
*/
adcs ip, ip, RHI /* accumulate RHI */
- ldr RHI, [r0] /* load last word */
- tst r1, #3 /* are we word aligned */
- beq .Lfinal_add_one_dword /* yes, accumulate last dword */
-
-.Lfinal_dword_noload:
- rsb r1, r1, #4 /* find out many bytes to discard */
- and r1, r1, #3 /* limit to a single word length */
- mov r1, r1, lsl #3 /* bytes -> bits */
-#ifdef __ARMEL__
- mov RHI, RHI, lsl r1 /* discard unneeded bits */
- mov RHI, RHI, lsr r1 /* replace with zero bits */
-#else
- mov RHI, RHI, lsr r1 /* discard unneeded bits */
- mov RHI, RHI, lsl r1 /* replace with zero bits */
-#endif
-#if 0
- tst r1, #2 /* discard at least 2? */
-#ifdef __ARMEL__
- movne RHI, RHI, lsl #16 /* yes, discard upper halfword */
-#else
- movne RHI, RHI, lsr #16 /* yes, discard lower halfword */
-#endif
- tst r1, #1 /* discard odd? */
- bicne RHI, RHI, #BYTE3 /* yes, discard odd byte */
-#endif
-.Lfinal_add_one_dword:
- adcs ip, ip, RLO /* add 1st to accumulator */
-.Lfinal_add_one_word:
- adcs ip, ip, RHI /* add 2nd to accumulator */
+ ldr RHI, [r0] /* load last word into RHI */
+.Ladd_final_dword:
+ adcs ip, ip, RLO /* add RLO to accumulator */
+.Ladd_final_word:
+ and RHI, RHI, r11 /* apply trailing mask to RHI */
+ adcs ip, ip, RHI /* add RHI to accumulator */
/*
* Fall into fold.
*/
+ tst r10, #1 /* was starting address odd? */
+ movne ip, ip, ror #8 /* yes, compensate */
+ pop {r4-r5,r10-r11} /* we don't need these anymore */
.Lfold:
- pop {r4-r5} /* we don't need these anymore */
-.Lfold_nopop:
/*
* We now have the 33-bit result in <carry>, ip. Pull in the
* standard folding code.
@@ -213,14 +210,11 @@ ENTRY(cpu_in_cksum_buffer)
#endif
tst r0, #3 /* are at least word aligned? */
bne .Lword_misaligned /* no, do it the hard way */
- push {r4-r5} /* save temporaries */
ldr RHI, [r0], #4 /* load word here in case of partial */
sub r1, r1, #4 /* subtract length of one word */
teq r1, #0 /* what is length? */
- beq .Lfinal_add_one_word /* = 0? just do the final add */
- mov RLO, #0 /* <= 0? zero this */
- bmi .Lfinal_dword_noload /* < 0? handle final partial dword */
- add r2, r1, r0 /* > 0? point r2 just past end */
+ beq .Ladd_final_word /* <= 0? just do the final add */
+ mov RLO, #0 /* > 0? clear RLO */
b .Ldword_aligned_noload /* > 0? accumulate it and loop */
.Lword_misaligned:
@@ -228,13 +222,9 @@ ENTRY(cpu_in_cksum_buffer)
* If we start on an odd boundary, set up our stack frame so we
* can fixup the return value to be byteswapped.
*/
- tst r0, #1 /* start address odd? */
- strne lr, [sp, #-8]! /* yes, save our return address */
- adrne lr, .Lmisaligned_fixup /* yes, return to fixup code. */
- push {r4-r5} /* save temporaries */
tst r0, #4 /* do we load 1 or 2 words? */
bic r0, r0, #3 /* force word alignment */
- add r1, r1, r2 /* add initial offset to length */
+ add r1, r1, r10 /* add initial offset to length */
sub r1, r1, #8 /* subtract length of one dword */
#ifdef _ARM_ARCH_DWORD_OK
ldreqd r4, [r0], #8 /* load first dword */
@@ -246,7 +236,7 @@ ENTRY(cpu_in_cksum_buffer)
/*
* We are now dword aligned.
*/
- and r3, r2, #3 /* limit to a single word length */
+ and r3, r10, #3 /* limit to a single word length */
mov r3, r3, lsl #3 /* bytes -> bits */
#ifdef __ARMEL__
mov RLO, RLO, lsr r3 /* discard unneeded bits */
@@ -260,34 +250,8 @@ ENTRY(cpu_in_cksum_buffer)
* into the main loop as if we just load a single dword.
*/
teq r1, #0 /* what is length? */
- beq .Lfinal_add_one_dword /* = 0? just do the final add */
- addpl r2, r1, r0 /* > 0? point r2 just past end */
+ beq .Ladd_final_dword /* = 0? just do the final add */
bpl .Ldword_aligned_noload /* > 0? accumulate it and loop */
-
- /*
- * Not a full dword so do the final dword processing to find out
- * bytes to discard. If we only loaded one word, move it to 2nd
- * word since that is what final_dword will be discarding from and
- * clear the 1st word.
- */
- tst r2, #4 /* one or two words? */
- movne RHI, RLO /* one, move lo word to hi word */
- movne RLO, #0 /* and clear lo word */
- b .Lfinal_dword_noload /* handle final dword */
-
- /*
- * If we had an odd address, we have byte swap the return value.
- * instead of testing everywhere, we inserted a fake callframe and
- * set LR to return to do the fixup and return to the caller.
- */
-.Lmisaligned_fixup:
- ldr lr, [sp], #8 /* fetch saved LR */
-#ifdef _ARM_ARCH_6
- rev16 r0, r0 /* byte swap */
-#else
- mov r0, r0, r0, ror #8 /* move 0:7 to 24:31 and 8:15 to 0:7 */
- orr r0, r0, r0, lsl #16 /* move 0:7 to 16:23 */
- mov r0, r0, r0, lsr #16 /* clear 16:31 to 0:15 */
-#endif
- RET
+ movne RHI, RLO /* yes? move RLO to RHI */
+ b .Ladd_final_word /* handle final word */
END(cpu_in_cksum_buffer)
