powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()

Currently memcmp() 64bytes version in powerpc will fall back to .Lshort
(compare per byte mode) if either src or dst address is not 8 bytes aligned.
It can be opmitized in 2 situations:

1) if both addresses are with the same offset with 8 bytes boundary:
memcmp() can compare the unaligned bytes within 8 bytes boundary firstly
and then compare the rest 8-bytes-aligned content with .Llong mode.

2)  If src/dst addrs are not with the same offset of 8 bytes boundary:
memcmp() can align src addr with 8 bytes, increment dst addr accordingly,
 then load src with aligned mode and load dst with unaligned mode.

This patch optmizes memcmp() behavior in the above 2 situations.

Tested with both little/big endian. Performance result below is based on
little endian.

Following is the test result with src/dst having the same offset case:
(a similar result was observed when src/dst having different offset):
(1) 256 bytes
Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp:
- without patch
	29.773018302 seconds time elapsed                                          ( +- 0.09% )
- with patch
	16.485568173 seconds time elapsed                                          ( +-  0.02% )
		-> There is ~+80% percent improvement

(2) 32 bytes
To observe performance impact on < 32 bytes, modify
tools/testing/selftests/powerpc/stringloops/memcmp.c with following:
-------
 #include <string.h>
 #include "utils.h"

-#define SIZE 256
+#define SIZE 32
 #define ITERATIONS 10000

 int test_memcmp(const void *s1, const void *s2, size_t n);
--------

- Without patch
	0.244746482 seconds time elapsed                                          ( +-  0.36%)
- with patch
	0.215069477 seconds time elapsed                                          ( +-  0.51%)
		-> There is ～+13% improvement

(3) 0~8 bytes
To observe <8 bytes performance impact, modify
tools/testing/selftests/powerpc/stringloops/memcmp.c with following:
-------
 #include <string.h>
 #include "utils.h"

-#define SIZE 256
-#define ITERATIONS 10000
+#define SIZE 8
+#define ITERATIONS 1000000

 int test_memcmp(const void *s1, const void *s2, size_t n);
-------
- Without patch
       1.845642503 seconds time elapsed                                          ( +- 0.12% )
- With patch
       1.849767135 seconds time elapsed                                          ( +- 0.26% )
		-> They are nearly the same. (-0.2%)

Signed-off-by: Simon Guo <wei.guo.simon@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

arch/powerpc/lib/memcmp_64.S

@@ -24,28 +24,41 @@
 #define rH	r31
 
 #ifdef __LITTLE_ENDIAN__
+#define LH	lhbrx
+#define LW	lwbrx
 #define LD	ldbrx
 #else
+#define LH	lhzx
+#define LW	lwzx
 #define LD	ldx
 #endif
 
+/*
+ * There are 2 categories for memcmp:
+ * 1) src/dst has the same offset to the 8 bytes boundary. The handlers
+ * are named like .Lsameoffset_xxxx
+ * 2) src/dst has different offset to the 8 bytes boundary. The handlers
+ * are named like .Ldiffoffset_xxxx
+ */
 _GLOBAL(memcmp)
 	cmpdi	cr1,r5,0
 
-	/* Use the short loop if both strings are not 8B aligned */
-	or	r6,r3,r4
+	/* Use the short loop if the src/dst addresses are not
+	 * with the same offset of 8 bytes align boundary.
+	 */
+	xor	r6,r3,r4
 	andi.	r6,r6,7
 
-	/* Use the short loop if length is less than 32B */
-	cmpdi	cr6,r5,31
+	/* Fall back to short loop if compare at aligned addrs
+	 * with less than 8 bytes.
+	 */
+	cmpdi   cr6,r5,7
 
 	beq	cr1,.Lzero
-	bne	.Lshort
-	bgt	cr6,.Llong
+	bgt	cr6,.Lno_short
 
 .Lshort:
 	mtctr	r5
-
 1:	lbz	rA,0(r3)
 	lbz	rB,0(r4)
 	subf.	rC,rB,rA
@@ -78,11 +91,89 @@ _GLOBAL(memcmp)
 	li	r3,0
 	blr
 
+.Lno_short:
+	dcbt	0,r3
+	dcbt	0,r4
+	bne	.Ldiffoffset_8bytes_make_align_start
+
+
+.Lsameoffset_8bytes_make_align_start:
+	/* attempt to compare bytes not aligned with 8 bytes so that
+	 * rest comparison can run based on 8 bytes alignment.
+	 */
+	andi.   r6,r3,7
+
+	/* Try to compare the first double word which is not 8 bytes aligned:
+	 * load the first double word at (src & ~7UL) and shift left appropriate
+	 * bits before comparision.
+	 */
+	rlwinm  r6,r3,3,26,28
+	beq     .Lsameoffset_8bytes_aligned
+	clrrdi	r3,r3,3
+	clrrdi	r4,r4,3
+	LD	rA,0,r3
+	LD	rB,0,r4
+	sld	rA,rA,r6
+	sld	rB,rB,r6
+	cmpld	cr0,rA,rB
+	srwi	r6,r6,3
+	bne	cr0,.LcmpAB_lightweight
+	subfic  r6,r6,8
+	subf.	r5,r6,r5
+	addi	r3,r3,8
+	addi	r4,r4,8
+	beq	.Lzero
+
+.Lsameoffset_8bytes_aligned:
+	/* now we are aligned with 8 bytes.
+	 * Use .Llong loop if left cmp bytes are equal or greater than 32B.
+	 */
+	cmpdi   cr6,r5,31
+	bgt	cr6,.Llong
+
+.Lcmp_lt32bytes:
+	/* compare 1 ~ 32 bytes, at least r3 addr is 8 bytes aligned now */
+	cmpdi   cr5,r5,7
+	srdi    r0,r5,3
+	ble	cr5,.Lcmp_rest_lt8bytes
+
+	/* handle 8 ~ 31 bytes */
+	clrldi  r5,r5,61
+	mtctr   r0
+2:
+	LD	rA,0,r3
+	LD	rB,0,r4
+	cmpld	cr0,rA,rB
+	addi	r3,r3,8
+	addi	r4,r4,8
+	bne	cr0,.LcmpAB_lightweight
+	bdnz	2b
+
+	cmpwi   r5,0
+	beq	.Lzero
+
+.Lcmp_rest_lt8bytes:
+	/* Here we have only less than 8 bytes to compare with. at least s1
+	 * Address is aligned with 8 bytes.
+	 * The next double words are load and shift right with appropriate
+	 * bits.
+	 */
+	subfic  r6,r5,8
+	slwi	r6,r6,3
+	LD	rA,0,r3
+	LD	rB,0,r4
+	srd	rA,rA,r6
+	srd	rB,rB,r6
+	cmpld	cr0,rA,rB
+	bne	cr0,.LcmpAB_lightweight
+	b	.Lzero
+
 .Lnon_zero:
 	mr	r3,rC
 	blr
 
 .Llong:
+	/* At least s1 addr is aligned with 8 bytes */
 	li	off8,8
 	li	off16,16
 	li	off24,24
@@ -232,4 +323,39 @@ _GLOBAL(memcmp)
 	ld	r28,-32(r1)
 	ld	r27,-40(r1)
 	blr
+
+.LcmpAB_lightweight:   /* skip NV GPRS restore */
+	li	r3,1
+	bgtlr
+	li	r3,-1
+	blr
+
+.Ldiffoffset_8bytes_make_align_start:
+	/* now try to align s1 with 8 bytes */
+	rlwinm  r6,r3,3,26,28
+	beq     .Ldiffoffset_align_s1_8bytes
+
+	clrrdi	r3,r3,3
+	LD	rA,0,r3
+	LD	rB,0,r4  /* unaligned load */
+	sld	rA,rA,r6
+	srd	rA,rA,r6
+	srd	rB,rB,r6
+	cmpld	cr0,rA,rB
+	srwi	r6,r6,3
+	bne	cr0,.LcmpAB_lightweight
+
+	subfic  r6,r6,8
+	subf.	r5,r6,r5
+	addi	r3,r3,8
+	add	r4,r4,r6
+
+	beq	.Lzero
+
+.Ldiffoffset_align_s1_8bytes:
+	/* now s1 is aligned with 8 bytes. */
+	cmpdi   cr5,r5,31
+	ble	cr5,.Lcmp_lt32bytes
+	b	.Llong
+
 EXPORT_SYMBOL(memcmp)