[glibc] powerpc: Fix performance issues of strcmp power10

[glibc] powerpc: Fix performance issues of strcmp power10

[Rajalakshmi Srinivasaraghavan]

https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=90bcc8721ef82b7378d2b080141228660e862d56

commit 90bcc8721ef82b7378d2b080141228660e862d56
Author: Amrita H S <amritahs@linux.vnet.ibm.com>
Date:   Fri Dec 15 11:48:17 2023 -0500

    powerpc: Fix performance issues of strcmp power10
    
    Current implementation of strcmp for power10 has
    performance regression for multiple small sizes
    and alignment combination.
    
    Most of these performance issues are fixed by this
    patch. The compare loop is unrolled and page crosses
    of unrolled loop is handled.
    
    Thanks to Paul E. Murphy for helping in fixing the
    performance issues.
    
    Signed-off-by: Amrita H S <amritahs@linux.vnet.ibm.com>
    Co-Authored-By: Paul E. Murphy <murphyp@linux.ibm.com>
    Reviewed-by: Rajalakshmi Srinivasaraghavan <rajis@linux.ibm.com>

Diff:
---
 sysdeps/powerpc/powerpc64/le/power10/strcmp.S | 161 +++++++++++++++-----------
 1 file changed, 95 insertions(+), 66 deletions(-)

diff --git a/sysdeps/powerpc/powerpc64/le/power10/strcmp.S b/sysdeps/powerpc/powerpc64/le/power10/strcmp.S
index a3c1adad53..3406f4f26a 100644
--- a/sysdeps/powerpc/powerpc64/le/power10/strcmp.S
+++ b/sysdeps/powerpc/powerpc64/le/power10/strcmp.S
@@ -62,7 +62,7 @@
 	lxvl	  32+v5,reg2,r0;         \
 	add	  reg1,reg1,len_reg;     \
 	add	  reg2,reg2,len_reg;     \
-	vcmpnezb. v7,v4,v5;              \
+	vcmpnezb  v7,v4,v5;              \
 	vctzlsbb  r6,v7;                 \
 	cmpld	  cr7,r6,len_reg;        \
 	blt	  cr7,L(different);      \
@@ -72,70 +72,110 @@
 
 	.machine  power9
 ENTRY_TOCLESS (STRCMP, 4)
-	li	 r11,16
-	/* eq bit of cr1 used as swap status flag to indicate if
-	source pointers were swapped.  */
-	crclr	 4*cr1+eq
-	vspltisb v19,-1
-	andi.	 r7,r3,15
-	sub	 r7,r11,r7	/* r7(nalign1) = 16 - (str1 & 15).  */
-	andi.	 r9,r4,15
-	sub	 r5,r11,r9	/* r5(nalign2) = 16 - (str2 & 15).  */
-	cmpld	 cr7,r7,r5
-	beq	 cr7,L(same_aligned)
-	blt	 cr7,L(nalign1_min)
-	/* Swap r3 and r4, and r7 and r5 such that r3 and r7 hold the
-	pointer which is closer to the next 16B boundary so that only
-	one CHECK_N_BYTES is needed before entering the loop below.  */
-	mr	 r8,r4
-	mr	 r4,r3
-	mr	 r3,r8
-	mr	 r12,r7
-	mr	 r7,r5
-	mr	 r5,r12
-	crset	 4*cr1+eq	/* Set bit on swapping source pointers.  */
+	andi.	r7,r3,4095
+	andi.	r8,r4,4095
+	cmpldi	cr0,r7,4096-16
+	cmpldi	cr1,r8,4096-16
+	bgt	cr0,L(crosses)
+	bgt	cr1,L(crosses)
+	COMPARE_16(v4,v5,0)
 
-	.p2align 5
+L(crosses):
+	andi.	r7,r3,15
+	subfic	r7,r7,16	/* r7(nalign1) = 16 - (str1 & 15).  */
+	andi.	r9,r4,15
+	subfic	r5,r9,16	/* r5(nalign2) = 16 - (str2 & 15).  */
+	cmpld	cr7,r7,r5
+	beq	cr7,L(same_aligned)
+	blt	cr7,L(nalign1_min)
+
+	/* nalign2 is minimum and s2 pointer is aligned.  */
+	CHECK_N_BYTES(r3,r4,r5)
+	/* Are we on the 64B hunk which crosses a page?  */
+	andi.	r10,r3,63	/* Determine offset into 64B hunk.  */
+	andi.	r8,r3,15        /* The offset into the 16B hunk.  */
+	neg	r7,r3
+	andi.	r9,r7,15	/* Number of bytes after a 16B cross.  */
+	rlwinm.	r7,r7,26,0x3F	/* ((r3-4096))>>6&63.  */
+	beq	L(compare_64_pagecross)
+	mtctr	r7
+	b	L(compare_64B_unaligned)
+
+	/* nalign1 is minimum and s1 pointer is aligned.  */
 L(nalign1_min):
 	CHECK_N_BYTES(r3,r4,r7)
+	/* Are we on the 64B hunk which crosses a page?  */
+	andi.	r10,r4,63	/* Determine offset into 64B hunk.  */
+	andi.	r8,r4,15	/* The offset into the 16B hunk.  */
+	neg	r7,r4
+	andi.	r9,r7,15	/* Number of bytes after a 16B cross.  */
+	rlwinm. r7,r7,26,0x3F	/* ((r4-4096))>>6&63.  */
+	beq	L(compare_64_pagecross)
+	mtctr	r7
 
 	.p2align 5
-L(s1_aligned):
-	/* r9 and r5 is number of bytes to be read after and before
-	 page boundary correspondingly.  */
-	sub 	r5,r5,r7
-	subfic	r9,r5,16
-	/* Now let r7 hold the count of quadwords which can be
-	checked without crossing a page boundary. quadword offset is
-	(str2>>4)&0xFF.  */
-	rlwinm	r7,r4,28,0xFF
-	/* Below check is required only for first iteration. For second
-	iteration and beyond, the new loop counter is always 255.  */
-	cmpldi	r7,255
-	beq	L(L3)
-	/* Get the initial loop count by 255-((str2>>4)&0xFF).  */
-	subfic  r11,r7,255
+L(compare_64B_unaligned):
+	COMPARE_16(v4,v5,0)
+	COMPARE_16(v4,v5,16)
+	COMPARE_16(v4,v5,32)
+	COMPARE_16(v4,v5,48)
+	addi	r3,r3,64
+	addi	r4,r4,64
+	bdnz	L(compare_64B_unaligned)
 
-	.p2align 5
-L(L1):
+	/* Cross the page boundary of s2, carefully. Only for first
+	iteration we have to get the count of 64B blocks to be checked.
+	From second iteration and beyond, loop counter is always 63.  */
+L(compare_64_pagecross):
+	li	r11, 63
 	mtctr	r11
-
-	.p2align 5
-L(L2):
-	COMPARE_16(v4,v5,0)	/* Load 16B blocks using lxv.  */
+	cmpldi	r10,16
+	ble	L(cross_4)
+	cmpldi	r10,32
+	ble	L(cross_3)
+	cmpldi	r10,48
+	ble	L(cross_2)
+L(cross_1):
+	CHECK_N_BYTES(r3,r4,r9)
+	CHECK_N_BYTES(r3,r4,r8)
+	COMPARE_16(v4,v5,0)
+	COMPARE_16(v4,v5,16)
+	COMPARE_16(v4,v5,32)
+	addi	r3,r3,48
+	addi	r4,r4,48
+	b	L(compare_64B_unaligned)
+L(cross_2):
+	COMPARE_16(v4,v5,0)
 	addi	r3,r3,16
 	addi	r4,r4,16
-	bdnz	L(L2)
-	/* Cross the page boundary of s2, carefully.  */
-
-	.p2align 5
-L(L3):
-	CHECK_N_BYTES(r3,r4,r5)
 	CHECK_N_BYTES(r3,r4,r9)
-	li 	r11,255		/* Load the new loop counter.  */
-	b	L(L1)
+	CHECK_N_BYTES(r3,r4,r8)
+	COMPARE_16(v4,v5,0)
+	COMPARE_16(v4,v5,16)
+	addi	r3,r3,32
+	addi	r4,r4,32
+	b	L(compare_64B_unaligned)
+L(cross_3):
+	COMPARE_16(v4,v5,0)
+	COMPARE_16(v4,v5,16)
+	addi	r3,r3,32
+	addi	r4,r4,32
+	CHECK_N_BYTES(r3,r4,r9)
+	CHECK_N_BYTES(r3,r4,r8)
+	COMPARE_16(v4,v5,0)
+	addi	r3,r3,16
+	addi	r4,r4,16
+	b	L(compare_64B_unaligned)
+L(cross_4):
+	COMPARE_16(v4,v5,0)
+	COMPARE_16(v4,v5,16)
+	COMPARE_16(v4,v5,32)
+	addi	r3,r3,48
+	addi	r4,r4,48
+	CHECK_N_BYTES(r3,r4,r9)
+	CHECK_N_BYTES(r3,r4,r8)
+	b	L(compare_64B_unaligned)
 
-	.p2align 5
 L(same_aligned):
 	CHECK_N_BYTES(r3,r4,r7)
         /* Align s1 to 32B and adjust s2 address.
@@ -168,18 +208,7 @@ L(16B_aligned_loop):
 
 	/* Calculate and return the difference.  */
 L(different):
-	vctzlsbb r6,v7
-	vextubrx r5,r6,v4
-	vextubrx r4,r6,v5
-	bt  	 4*cr1+eq,L(swapped)
-	subf	 r3,r4,r5
-	blr
-
-	/* If src pointers were swapped, then swap the
-	indices and calculate the return value.  */
-L(swapped):
-	subf     r3,r5,r4
-	blr
+	TAIL(v4,v5)
 
 	.p2align 5
 L(32B_aligned_loop):