[PATCH] powerpc: strncmp optimization for POWER8

[Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>]

This patch adds an optimized POWER8 strncmp.  The implementation focus
on speeding up unaligned cases following the ideas of power8 strcmp.

The algorithm first check the initial 16 bytes, then align the first
function source and uses unaligned loads on second argument only.
Aditional checks for page boundaries are done for unaligned cases
(where sources alignment are different).

Checked on powerpc64 and powerpc64le.

--

	* sysdeps/powerpc/powerpc64/multiarch/strncmp-power8.S: New file.
	* sysdeps/powerpc/powerpc64/power8/strncmp.S: New file.
	* sysdeps/powerpc/powerpc64/multiarch/Makefile [sysdep_routines]: Add
	strncmp-power8 object.
	* sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
	(__libc_ifunc_impl_list): Add __strncmp_power8 implementation.
	* sysdeps/powerpc/powerpc64/multiarch/strncmp.c (strncmp): Likewise.
	* NEWS: Update.

--

diff --git a/sysdeps/powerpc/powerpc64/multiarch/Makefile b/sysdeps/powerpc/powerpc64/multiarch/Makefile
index ec4fca7..b7ea284 100644
--- a/sysdeps/powerpc/powerpc64/multiarch/Makefile
+++ b/sysdeps/powerpc/powerpc64/multiarch/Makefile
@@ -7,8 +7,9 @@ sysdep_routines += memcpy-power7 memcpy-a2 memcpy-power6 memcpy-cell \
 		   memrchr-power7 memrchr-ppc64 rawmemchr-power7 \
 		   rawmemchr-ppc64 strlen-power7 strlen-ppc64 strnlen-power7 \
 		   strnlen-ppc64 strcasecmp-power7 strcasecmp_l-power7 \
-		   strncase-power7 strncase_l-power7 strncmp-power7 \
-		   strncmp-power4 strncmp-ppc64 strchr-power7 strchr-ppc64 \
+		   strncase-power7 strncase_l-power7 \
+		   strncmp-power8 strncmp-power7 strncmp-power4 strncmp-ppc64 \
+		   strchr-power7 strchr-ppc64 \
 		   strchrnul-power7 strchrnul-ppc64 wcschr-power7 \
 		   wcschr-power6 wcschr-ppc64 wcsrchr-power7 wcsrchr-power6 \
 		   wcsrchr-ppc64 wcscpy-power7 wcscpy-power6 wcscpy-ppc64 \
diff --git a/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c b/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
index 2c03060..f5fdea5 100644
--- a/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
@@ -108,6 +108,8 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 
   /* Support sysdeps/powerpc/powerpc64/multiarch/strncmp.c.  */
   IFUNC_IMPL (i, name, strncmp,
+	      IFUNC_IMPL_ADD (array, i, strncmp, hwcap2 & PPC_FEATURE2_ARCH_2_07,
+			      __strncmp_power8)
 	      IFUNC_IMPL_ADD (array, i, strncmp, hwcap & PPC_FEATURE_HAS_VSX,
 			      __strncmp_power7)
 	      IFUNC_IMPL_ADD (array, i, strncmp, hwcap & PPC_FEATURE_POWER4,
diff --git a/sysdeps/powerpc/powerpc64/multiarch/strncmp-power8.S b/sysdeps/powerpc/powerpc64/multiarch/strncmp-power8.S
new file mode 100644
index 0000000..8d7223d
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/multiarch/strncmp-power8.S
@@ -0,0 +1,40 @@
+/* Copyright (C) 2015 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+
+#undef EALIGN
+#define EALIGN(name,alignt,words)				\
+  .section ".text";						\
+  ENTRY_2(__strncmp_power8)					\
+  .align ALIGNARG(alignt);					\
+  EALIGN_W_##words;						\
+  BODY_LABEL(__strncmp_power8):					\
+  cfi_startproc;						\
+  LOCALENTRY(__strncmp_power8)
+
+#undef END
+#define END(name)						\
+  cfi_endproc;							\
+  TRACEBACK(__strncmp_power8)					\
+  END_2(__strncmp_power8)
+
+
+#undef libc_hidden_builtin_def
+#define libc_hidden_builtin_def(name)
+
+#include <sysdeps/powerpc/powerpc64/power8/strncmp.S>
diff --git a/sysdeps/powerpc/powerpc64/multiarch/strncmp.c b/sysdeps/powerpc/powerpc64/multiarch/strncmp.c
index eb02aac..9b6a659 100644
--- a/sysdeps/powerpc/powerpc64/multiarch/strncmp.c
+++ b/sysdeps/powerpc/powerpc64/multiarch/strncmp.c
@@ -25,13 +25,16 @@
 extern __typeof (strncmp) __strncmp_ppc attribute_hidden;
 extern __typeof (strncmp) __strncmp_power4 attribute_hidden;
 extern __typeof (strncmp) __strncmp_power7 attribute_hidden;
+extern __typeof (strncmp) __strncmp_power8 attribute_hidden;
 
 /* Avoid DWARF definition DIE on ifunc symbol so that GDB can handle
    ifunc symbol properly.  */
 libc_ifunc (strncmp,
-            (hwcap & PPC_FEATURE_HAS_VSX)
-            ? __strncmp_power7 :
-	      (hwcap & PPC_FEATURE_POWER4)
+            (hwcap2 & PPC_FEATURE2_ARCH_2_07)
+            ? __strncmp_power8 :
+              (hwcap & PPC_FEATURE_HAS_VSX)
+              ? __strncmp_power7 :
+		(hwcap & PPC_FEATURE_POWER4)
 		? __strncmp_power4
             : __strncmp_ppc);
 #endif
diff --git a/sysdeps/powerpc/powerpc64/power8/strncmp.S b/sysdeps/powerpc/powerpc64/power8/strncmp.S
new file mode 100644
index 0000000..c13c29e
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/power8/strncmp.S
@@ -0,0 +1,323 @@
+/* Optimized strcpy/stpcpy implementation for PowerPC64/POWER8.
+   Copyright (C) 2015 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+
+/* Implements the function
+
+   int [r3] strncmp (const char *s1 [r3], const char *s2 [r4], size_t [r5] n)
+
+   The implementation uses unaligned doubleword access to avoid specialized
+   code paths depending of data alignment.  Although recent powerpc64 uses
+   64K as default, the page cross handling assumes minimum page size of
+   4k.  */
+
+	.machine  power7
+EALIGN (strncmp, 4, 0)
+	/* Check if size is 0.  */
+	mr.	r10,r5
+	beq	cr0,L(ret0)
+
+	/* Check if [s1]+16 or [s2]+16 will cross a 4K page boundary using
+	   the code:
+
+	    (((size_t) s1) % PAGE_SIZE > (PAGE_SIZE - ITER_SIZE))
+
+	   with PAGE_SIZE being 4096 and ITER_SIZE begin 16.  */
+	rldicl	r8,r3,0,52
+	cmpldi	cr7,r8,4096-16
+	bgt	cr7,L(pagecross)
+	rldicl	r9,r4,0,52
+	cmpldi	cr7,r9,4096-16
+	bgt	cr7,L(pagecross)
+
+	/* For short string up to 16 bytes, load both s1 and s2 using
+	   unaligned dwords and compare.  */
+	ld	r7,0(r3)
+	ld	r9,0(r4)
+	li	r8,0
+	cmpb	r8,r7,r8
+	cmpb	r6,r7,r9
+	orc.	r8,r8,r6
+	bne	cr0,L(different1)
+
+	/* If the string compared are equal, but size is less or equal
+	   to 8, return 0.  */
+	cmpldi	cr7,r10,8
+	li	r9,0
+	ble	cr7,L(ret1)
+	addi	r5,r10,-8
+
+	ld	r7,8(r3)
+	ld	r9,8(r4)
+	cmpb	r8,r7,r8
+	cmpb	r6,r7,r9
+	orc.	r8,r8,r6
+	bne	cr0,L(different0)
+
+	cmpldi	cr7,r5,8
+	mr	r9,r8
+	ble	cr7,L(ret1)
+
+	/* Update pointers and size.  */
+	addi	r10,r10,-16
+	addi	r3,r3,16
+	addi	r4,r4,16
+
+	/* Now it has checked for first 16 bytes, align source1 to doubleword
+	   and adjust source2 address.  */
+L(align_8b):
+	rldicl	r5,r3,0,61
+	rldicr	r3,r3,0,60
+	subf	r4,r5,r4
+	add	r10,r10,r5
+
+	/* At this point, source1 alignment is 0 and source2 alignment is
+	   between 0 and 7.  Check is source2 alignment is 0, meaning both
+	   sources have the same alignment.  */
+	andi.	r8,r4,0x7
+	beq	cr0,L(loop_eq_align_0)
+
+	li	r5,0
+	b	L(loop_ne_align_1)
+
+	/* If source2 is unaligned to doubleword, the code needs to check
+	   on each interation if the unaligned doubleword access will cross
+	   a 4k page boundary.  */
+	.align 4
+L(loop_ne_align_0):
+	ld	r7,0(r3)
+	ld	r9,0(r4)
+	cmpb	r8,r7,r5
+	cmpb	r6,r7,r9
+	orc.	r8,r8,r6
+	bne	cr0,L(different1)
+
+	cmpldi	cr7,r10,8
+	ble	cr7,L(ret0)
+	addi	r10,r10,-8
+	addi	r3,r3,8
+	addi	r4,r4,8
+L(loop_ne_align_1):
+	rldicl	r9,r4,0,52
+	cmpldi	r7,r9,4088
+	ble	cr7,L(loop_ne_align_0)
+	cmpdi	cr7,r10,0
+	beq	cr7,L(ret0)
+
+	lbz	r9,0(r3)
+	lbz	r8,0(r4)
+	cmplw	cr7,r9,r8
+	bne	cr7,L(byte_ne_4)
+	cmpdi	cr7,r9,0
+	beq	cr7,L(size_reached_0)
+
+	li	r9,r7
+	addi	r8,r3,1
+	mtctr	r9
+	addi	r4,r4,1
+	addi	r10,r10,-1
+	addi	r3,r3,8
+
+	/* The unaligned read of source2 will cross a 4K page boundary,
+	   and the different byte or NULL maybe be in the remaining page
+	   bytes.  Since it can not use the unaligned load the algorithm
+	   reads and compares 8 bytes to keep source1 doubleword aligned.  */
+	.align 4
+L(loop_ne_align_byte):
+	cmpdi	cr7,r10,0
+	addi	r10,r10,-1
+	beq	cr7,L(ret0)
+	lbz	r9,0(r8)
+	lbz	r7,0(r4)
+	addi	r8,r8,1
+	addi	r4,r4,1
+	cmplw	cr7,r9,r7
+	cmpdi	cr5,r9,0
+	bne	cr7,L(size_reached_2)
+	beq	cr5,L(size_reached_0)
+	bdnz	L(loop_ne_align_byte)
+
+	cmpdi	cr7,r10,0
+	bne+	cr7,L(loop_ne_align_0)
+
+	.align 4
+L(ret0):
+	li	r9,0
+L(ret1):
+	mr	r3,r9
+	blr
+
+	/* The code now check if r8 and r10 are different by issuing a
+	   cmpb and shift the result based on its output:
+
+	#ifdef __LITTLE_ENDIAN__
+	  leadzero = (__builtin_ffsl (z1) - 1);
+	  leadzero = leadzero > (n-1)*8 ? (n-1)*8 : leadzero;
+	  r1 = (r1 >> leadzero) & 0xFFUL;
+	  r2 = (r2 >> leadzero) & 0xFFUL;
+	#else
+	  leadzero = __builtin_clzl (z1);
+	  leadzero = leadzero > (n-1)*8 ? (n-1)*8 : leadzero;
+	  r1 = (r1 >> (56 - leadzero)) & 0xFFUL;
+	  r2 = (r2 >> (56 - leadzero)) & 0xFFUL;
+	#endif
+	  return r1 - r2;  */
+
+	.align 4
+L(different0):
+	mr	r10,r5
+#ifdef __LITTLE_ENDIAN__
+L(different1):
+        neg	r11,r8
+        sldi	r10,r10,3
+        and	r8,r11,r8
+        addi	r10,r10,-8
+        cntlzd	r8,r8
+        subfic	r8,r8,63
+        extsw 	r8,r8
+        cmpld	cr7,r8,r10
+        ble	cr7,L(different2)
+        mr	r8,r10
+L(different2):
+        extsw	r8,r8
+#else
+L(different1):
+	addi	r10,r10,-1
+	cntlzd	r8,r8
+	sldi	r10,r10,3
+	cmpld	cr7,r8,r10
+	blt	cr7,L(different2)
+	mr	r8,r10
+L(different2):
+	subfic	r8,r8,56
+#endif
+	srd	r7,r7,r8
+	srd	r9,r9,r8
+	rldicl	r3,r7,0,56
+	rldicl	r9,r9,0,56
+	subf	r9,r9,3
+	extsw	r9,r9
+	mr	r3,r9
+	blr
+
+	/* If unaligned 16 bytes reads across a 4K page boundary, it uses
+	   a simple byte a byte comparison until the page alignment for s1
+	   is reached.  */
+	.align 4
+L(pagecross):
+	lbz	r7,0(r3)
+	lbz	r9,0(r4)
+	subfic	r8,r8,4095
+	cmplw	cr7,r9,r7
+	bne	cr7,L(byte_ne_3)
+	cmpdi	cr7,r9,0
+	beq	cr7,L(byte_ne_0)
+	addi	r10,r10,-1
+	subf	r7,r8,r10
+	subf	r9,r7,r10
+	addi	r9,r9,1
+	mtctr	r9
+	b	L(pagecross_loop1)
+
+	.align 4
+L(pagecross_loop0):
+	beq	cr7,L(ret0)
+	lbz	r9,0(r3)
+	lbz	r8,0(r4)
+	addi	r10,r10,-1
+	cmplw	cr7,r9,r8
+	cmpdi	cr5,r9,0
+	bne	r7,L(byte_ne_2)
+	beq	r5,L(byte_ne_0)
+L(pagecross_loop1):
+	cmpdi	cr7,r10,0
+	addi	r3,r3,1
+	addi	r4,r4,1
+	bdnz	L(pagecross_loop0)
+	cmpdi	cr7,r7,0
+	li	r9,0
+	bne+	cr7,L(align_8b)
+	b	L(ret1)
+
+	/* If both source1 and source2 are doubleword aligned, there is no
+	   need for page boundary cross checks.  */
+	.align 4
+L(loop_eq_align_0):
+	ld	r7,0(r3)
+	ld	r9,0(r4)
+	cmpb	r8,r7,r8
+	cmpb	r6,r7,r9
+	orc.	r8,r8,r6
+	bne	cr0,L(different1)
+
+	cmpldi	cr7,r10,8
+	ble	cr7,L(ret0)
+	addi	r9,r10,-9
+
+	li	r5,0
+	srdi	r9,r9,3
+	addi	r9,r9,1
+	mtctr	r9
+	b	L(loop_eq_align_2)
+
+	.align 4
+L(loop_eq_align_1):
+	bdz	L(ret0)
+L(loop_eq_align_2):
+	ldu	r7,8(r3)
+	addi	r10,r10,-8
+	ldu	r9,8(r4)
+	cmpb	r8,r7,r5
+	cmpb	r6,r7,r9
+	orc.	r8,r8,r6
+	beq	cr0,L(loop_eq_align_1)
+	b	L(different1)
+
+	.align 4
+L(byte_ne_0):
+	li	r7,0
+L(byte_ne_1):
+	subf	r9,r9,r7
+	extsw	r9,r9
+	b	L(ret1)
+
+	.align 4
+L(byte_ne_2):
+	extsw	r7,r9
+	mr	r9,r8
+	b	L(byte_ne_1)
+L(size_reached_0):
+	li	r10,0
+L(size_reached_1):
+	subf	r9,r9,r10
+	extsw	r9,r9
+	b	L(ret1)
+L(size_reached_2):
+	extsw	r10,r9
+	mr	r9,r7
+	b	L(size_reached_1)
+L(byte_ne_3):
+	extsw	r7,r7
+	b	L(byte_ne_1)
+L(byte_ne_4):
+	extsw	r10,r9
+	mr	r9,r8
+	b	L(size_reached_1)
+END(strncmp)
+libc_hidden_builtin_def(strncmp)