[PATCH V2] powerpc: Fix performance issues of strcmp power10

Amrita H S amritahs@linux.vnet.ibm.com
Fri Dec 15 11:33:48 GMT 2023


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>
---
 sysdeps/powerpc/powerpc64/le/power10/strcmp.S | 160 +++++++++++-------
 1 file changed, 95 insertions(+), 65 deletions(-)

diff --git a/sysdeps/powerpc/powerpc64/le/power10/strcmp.S b/sysdeps/powerpc/powerpc64/le/power10/strcmp.S
index a3c1adad53..c80eced745 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 +72,111 @@
 
 	.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(nalign1_min):
-	CHECK_N_BYTES(r3,r4,r7)
+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)
 
-	.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
+	/* 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(L3)
-	/* Get the initial loop count by 255-((str2>>4)&0xFF).  */
-	subfic  r11,r7,255
+	mtctr	r7
+	b	L(L2)
 
-	.p2align 5
-L(L1):
-	mtctr	r11
+	/* 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(L3)
+	mtctr	r7
 
 	.p2align 5
 L(L2):
-	COMPARE_16(v4,v5,0)	/* Load 16B blocks using lxv.  */
-	addi	r3,r3,16
-	addi	r4,r4,16
+	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(L2)
-	/* Cross the page boundary of s2, carefully.  */
 
-	.p2align 5
+	/* 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(L3):
-	CHECK_N_BYTES(r3,r4,r5)
+	li	r11, 63
+	mtctr	r11
+	cmpldi	r10,16
+	ble	L(cross_4)
+	cmpldi	r10,32
+	ble	L(cross_3)
+	cmpldi	r10,48
+	ble	L(cross_2)
+
+	/* cross_1  */
 	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)
+	COMPARE_16(v4,v5,32)
+	addi	r3,r3,48
+	addi	r4,r4,48
+	b	L(L2)
+L(cross_2):
+	COMPARE_16(v4,v5,0)
+	addi	r3,r3,16
+	addi	r4,r4,16
+	CHECK_N_BYTES(r3,r4,r9)
+	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(L2)
+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(L2)
+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(L2)
 
-	.p2align 5
 L(same_aligned):
 	CHECK_N_BYTES(r3,r4,r7)
         /* Align s1 to 32B and adjust s2 address.
@@ -168,18 +209,7 @@ 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):
-- 
2.41.0



More information about the Libc-alpha mailing list