bca718
    Backport of
bca718
    commit 72607db038df1a1a7987af814aad8d2ed466c45c
bca718
    Author: Rajalakshmi Srinivasaraghavan <raji@linux.vnet.ibm.com>
bca718
    Date:   Fri Jan 9 11:56:35 2015 -0500
bca718
    
bca718
        powerpc: Optimize POWER7 strcmp trailing checks
bca718
    
bca718
        This patch optimized the POWER7 trailing check by avoiding using byte
bca718
        read operations and instead use the doubleword already readed with
bca718
        bitwise operations.
bca718
    
bca718
        ChangeLog:
bca718
    	2015-01-13  Rajalakshmi Srinivasaraghavan  <raji@linux.vnet.ibm.com>
bca718
    		    Adhemerval Zanella  <azanella@linux.vnet.ibm.com>
bca718
    
bca718
    	* sysdeps/powerpc/powerpc64/power7/strcmp.S (strcmp): Optimize
bca718
    	trailing byte check.
bca718
bca718
diff --git a/sysdeps/powerpc/powerpc64/power7/strcmp.S b/sysdeps/powerpc/powerpc64/power7/strcmp.S
bca718
index f16a9d8..ade2811 100644
bca718
--- a/sysdeps/powerpc/powerpc64/power7/strcmp.S
bca718
+++ b/sysdeps/powerpc/powerpc64/power7/strcmp.S
bca718
@@ -25,122 +25,96 @@
bca718
 
bca718
 /* int [r3] strcmp (const char *s1 [r3], const char *s2 [r4])  */
bca718
 
bca718
+	.machine	power7
bca718
 EALIGN (strcmp, 4, 0)
bca718
 	CALL_MCOUNT 2
bca718
 
bca718
 	or r9, r3, r4
bca718
 	rldicl. r10, r9, 0, 61	/* are s1 and s2 8 byte aligned..?  */
bca718
 	bne cr0, L(process_unaligned_bytes)
bca718
+	li	r5, 0
bca718
 
bca718
+	.align 4
bca718
 /* process input parameters on double word aligned boundary  */
bca718
-	ld r9, 0(r4)		/* load s2 at offset=0  */
bca718
-	li r10, 0		/* load mask=0  */
bca718
-	cmpb r10, r9, r10	/* compare bytes at s2 with mask  */
bca718
-	cmpdi cr7, r10, 0	/* is NULL found ..? is end of string HIT  */
bca718
-	bne cr7, L(process_unaligned_bytes)	/* process byte by byte  */
bca718
-
bca718
-	ld r10, 0(r3)		/* load s1 at offset=0  */
bca718
-	li r8, 0		/* load mask=0  */
bca718
-	cmpb r8, r10, r8	/* compare bytes at s1 with mask  */
bca718
-	cmpdi cr7, r8, 0	/* is NULL found ..? is end of string HIT  */
bca718
-	bne cr7, L(process_unaligned_bytes)	/* process byte by byte  */
bca718
-
bca718
-/*s1 and s2 does not contain NULL now , so compare all 8 bytes in a GO  */
bca718
-	cmpb r9, r10, r9	/* compare s1 and s2  */
bca718
-	cmpdi cr7, r9, -1	/* compare result with 0xFFFFFFFFFFFFFFFF  */
bca718
-	bne cr7, L(process_unaligned_bytes)	/* s1,s2 mismatch found  */
bca718
-
bca718
-	addi r5, r3, 8		/* save next offset of s2  */
bca718
-	addi r11, r4, 8		/* save next offset of s1  */
bca718
-	ld r8, 8(r4)		/* load s2 at offset=8  */
bca718
-	li r9, 0		/* load mask=0  */
bca718
-	cmpb r9, r8, r9		/* compare bytes at s2 with mask  */
bca718
-	cmpdi cr7, r9, 0	/* NULL found ..?  */
bca718
-	bne cr7, L(processBytes)/* update input and process bytes one by one  */
bca718
-
bca718
-	mr r9, r4		/* save s2  */
bca718
-	li r10, 0		/* load mask=0  */
bca718
-
bca718
-	ld r7, 8(r3)		/* load s1 at offset=8  */
bca718
-	cmpb r6, r7, r10	/* compare bytes at s1 with mask  */
bca718
-	cmpdi cr7, r6, 0	/* is NULL found  */
bca718
-	bne cr7, L(processBytes)/* mismatch, so process one by one  */
bca718
-
bca718
 L(unrollDword):
bca718
-	cmpb r8, r7, r8		/* compare s1 and s2  */
bca718
-	cmpdi cr7, r8, -1	/* compare result with 0xFFFFFFFFFFFFFFFF  */
bca718
-	bne cr7, L(processBytes)/* mismatch with s1 and s2  */
bca718
-
bca718
-	addi r5, r3, 16		/* save offset=16 of s1  */
bca718
-	addi r4, r9, 16		/* save offset=16 of s2  */
bca718
-	ld r8, 16(r9)		/* load s2 at offset=16  */
bca718
-	cmpb r7, r8, r10	/* compare bytes at s2 with mask  */
bca718
-	cmpdi cr7, r7, 0	/* NULL found  ..?  */
bca718
-	bne cr7, L(update2processBytes)
bca718
-
bca718
-	ld r7, 16(r3)		/* load s1 at offset=16  */
bca718
-	cmpb r6, r7, r10	/* check s1 for end of string  */
bca718
-	cmpdi cr7, r6, 0	/* end of s1 ?,then handle byte by byte  */
bca718
-	bne 7,L(update2processBytes)
bca718
-
bca718
-	cmpb r8, r7, r8		/* compare s1 and s2 double words  */
bca718
-	cmpdi cr7, r8, -1	/* compare results with 0xFFFFFFFFFFFFFFFF  */
bca718
-	bne cr7,L(update2processBytes)
bca718
-
bca718
-	addi r5, r3, 24		/* update s1 to offset=24  */
bca718
-	addi r4, r9, 24		/* update s2 to offset=24  */
bca718
-
bca718
-	ld r8, 24(r9)		/* load s2  */
bca718
-	cmpb r7, r8, r10	/* compare s2 for NULL  */
bca718
-	cmpdi cr7, r7, 0	/* verify if s2 is ending now  */
bca718
-	bne cr7,L(update2processBytes)
bca718
-
bca718
-	ld r7, 24(r3)		/* load s1 at offset=24  */
bca718
-	cmpb r6, r7, r10	/* verify for NULL  */
bca718
-	cmpdi cr7, r6, 0	/* is NULL found  */
bca718
-	bne cr7, L(update2processBytes)
bca718
-
bca718
-	cmpb r8, r7, r8		/* compare s1 and s2  */
bca718
-	cmpdi cr7, r8, -1	/* are s1 and s2 same ..?  */
bca718
-	bne cr7, L(update2processBytes)
bca718
-
bca718
-	addi r7, r9, 32		/* update s2 to next double word  */
bca718
-	addi r3, r3, 32		/* update s1 to next double word  */
bca718
-
bca718
-	ld r8, 32(r9)		/* load s2  */
bca718
-	mr r4, r7		/* save s2  */
bca718
-	cmpb r6, r8, r10	/* compare s2 with NULL  */
bca718
-	cmpdi cr7, r6, 0	/* end of s2 ..? */
bca718
-	bne cr7, L(process_unaligned_bytes)
bca718
-
bca718
-	ld r6, 0(r3)		/* load and compare s1 for NULL  */
bca718
-	cmpb r5, r6, r10
bca718
-	cmpdi cr7, r5, 0
bca718
-	bne cr7, L(process_unaligned_bytes)
bca718
-
bca718
-	cmpb r8, r6, r8		/* compare s1 and s2  */
bca718
-	cmpdi cr7, r8, -1
bca718
-	bne cr7, L(process_unaligned_bytes)
bca718
-
bca718
-	addi r5, r3, 8		/* increment s1 and d2 here  */
bca718
-	addi r11, r9, 40
bca718
-
bca718
-	ld r8, 40(r9)		/* process s2 now  */
bca718
-	cmpb r9, r8, r10
bca718
-	cmpdi cr7, r9, 0
bca718
-	bne cr7, L(processBytes)
bca718
-
bca718
-	mr r9, r7
bca718
-	ld r7, 8(r3)		/* process s1 now  */
bca718
-	cmpb r6, r7, r10
bca718
-	cmpdi cr7, r6, 0
bca718
-	beq cr7, L(unrollDword)	/* unroll to compare s1 and s2  */
bca718
-
bca718
-L(processBytes):
bca718
-	mr r4, r11		/* update input params  */
bca718
-	mr r3, r5
bca718
-
bca718
-	.p2align 4
bca718
+	ld	r8,0(r3)
bca718
+	ld	r10,0(r4)
bca718
+	cmpb	r7,r8,r5
bca718
+	cmpdi	cr7,r7,0
bca718
+	mr	r9,r7
bca718
+	bne 	cr7,L(null_found)
bca718
+	cmpld	cr7,r8,r10
bca718
+	bne	cr7,L(different)
bca718
+
bca718
+	ld	r8,8(r3)
bca718
+	ld	r10,8(r4)
bca718
+	cmpb	r7,r8,r5
bca718
+	cmpdi	cr7,r7,0
bca718
+	mr	r9,r7
bca718
+	bne 	cr7,L(null_found)
bca718
+	cmpld	cr7,r8,r10
bca718
+	bne	cr7,L(different)
bca718
+
bca718
+	ld	r8,16(r3)
bca718
+	ld	r10,16(r4)
bca718
+	cmpb	r7,r8,r5
bca718
+	cmpdi	cr7,r7,0
bca718
+	mr	r9,r7
bca718
+	bne 	cr7,L(null_found)
bca718
+	cmpld	cr7,r8,r10
bca718
+	bne	cr7,L(different)
bca718
+
bca718
+	ld	r8,24(r3)
bca718
+	ld	r10,24(r4)
bca718
+	cmpb	r7,r8,r5
bca718
+	cmpdi	cr7,r7,0
bca718
+	mr	r9,r7
bca718
+	bne 	cr7,L(null_found)
bca718
+	cmpld	cr7,r8,r10
bca718
+	bne	cr7,L(different)
bca718
+
bca718
+	addi r3, r3, 32
bca718
+	addi r4, r4, 32
bca718
+	beq cr7, L(unrollDword)
bca718
+
bca718
+	.align 4
bca718
+L(null_found):
bca718
+#ifdef __LITTLE_ENDIAN__
bca718
+	neg	r7,r9
bca718
+	and	r9,r9,r7
bca718
+	li	r7,-1
bca718
+	cntlzd	r9,r9
bca718
+	subfic	r9,r9,71
bca718
+	sld	r9,r7,r9
bca718
+#else
bca718
+	cntlzd	r9,r9
bca718
+	li	r7,-1
bca718
+	addi	r9,r9,8
bca718
+	srd	r9,r7,r9
bca718
+#endif
bca718
+	or	r8,r8,r9
bca718
+	or	r10,r10,r9
bca718
+
bca718
+L(different):
bca718
+	cmpb	r9,r8,r10
bca718
+#ifdef __LITTLE_ENDIAN__
bca718
+	addi	r7,r9,1
bca718
+	andc	r9,r7,r9
bca718
+	cntlzd	r9,r9
bca718
+	subfic	r9,r9,63
bca718
+#else
bca718
+	not	r9,r9
bca718
+	cntlzd	r9,r9
bca718
+	subfic	r9,r9,56
bca718
+#endif
bca718
+	srd	r3,r8,r9
bca718
+	srd	r10,r10,r9
bca718
+	rldicl	r10,r10,0,56
bca718
+	rldicl	r3,r3,0,56
bca718
+	subf	r3,r10,r3
bca718
+	blr
bca718
+
bca718
+	.align 4
bca718
 L(process_unaligned_bytes):
bca718
 	lbz r9, 0(r3)		/* load byte from s1  */
bca718
 	lbz r10, 0(r4)		/* load byte from s2  */
bca718
@@ -172,24 +146,19 @@ L(process_unaligned_bytes):
bca718
 	addi r4, r4, 4		/* increment s2 by unroll factor  */
bca718
 	beq cr6, L(process_unaligned_bytes)	/* unroll byte processing  */
bca718
 
bca718
-	.p2align 4
bca718
+	.align 4
bca718
 L(ComputeDiff):
bca718
 	extsw r9, r9
bca718
 	subf r10, r10, r9	/* compute s1 - s2  */
bca718
 	extsw r3, r10
bca718
 	blr			/* return  */
bca718
 
bca718
-	.p2align 4
bca718
+	.align 4
bca718
 L(diffOfNULL):
bca718
 	li r9, 0
bca718
 	subf r10, r10, r9	/* compute s1 - s2  */
bca718
 	extsw r3, r10		/* sign extend result  */
bca718
 	blr			/* return  */
bca718
 
bca718
-	.p2align 4
bca718
-L(update2processBytes):
bca718
-	mr r3, r5		/* update and proceed  */
bca718
-	b L(process_unaligned_bytes)
bca718
-
bca718
 END (strcmp)
bca718
 libc_hidden_builtin_def (strcmp)