powerpc: POWER7 strcpy optimization for unaligned strings

This patch optimizes strcpy for ppc64/power7 for unaligned source or
destination address.  The source or destination address is aligned
to doubleword and data is shifted based on the alignment and
added with the previous loaded data to be written as a doubleword.
For each load, cmpb instruction is used for faster null check.

The word aligned optimization is also removed, since the new unaligned
code path shows better results handling word-aligned strings.

More combination of unaligned inputs is also added in benchtest
to measure the improvement.The new optimization shows 2 to 80% of
performance improvement for longer string though it does not show
big difference on string size less than 16 due to additional checks.

3 files changed, 268 insertions, 82 deletions

diff --git a/ChangeLog b/ChangeLog
index 3de34438ff..3e2d4b0d7e 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,10 @@
+2014-12-31  Rajalakshmi Srinivasaraghavan  <raji@linux.vnet.ibm.com>
+	    Adhemerval Zanella  <azanella@linux.vnet.ibm.com>
+
+	* sysdeps/powerpc/powerpc64/power7/strcpy.S (strcpy): Optimize unaligned
+	path.
+	* benchtests/bench-strcpy.c (test_main): Add more unaligned inputs.
+
 2014-12-31  Joseph Myers  <joseph@codesourcery.com>
 
 	* sysdeps/powerpc/bits/fenvinline.h (fegetround): Rename macro to
diff --git a/benchtests/bench-strcpy.c b/benchtests/bench-strcpy.c
index c3ab4cfcf7..e9445f290f 100644
--- a/benchtests/bench-strcpy.c
+++ b/benchtests/bench-strcpy.c
@@ -171,6 +171,22 @@ test_main (void)
       do_test (i, i, 8 << i, BIG_CHAR);
     }
 
+  for (i = 16; i <= 512; i+=4)
+    {
+      do_test (0, 4, i, SMALL_CHAR);
+      do_test (4, 0, i, BIG_CHAR);
+      do_test (4, 4, i, SMALL_CHAR);
+      do_test (2, 2, i, BIG_CHAR);
+      do_test (2, 6, i, SMALL_CHAR);
+      do_test (6, 2, i, BIG_CHAR);
+      do_test (1, 7, i, SMALL_CHAR);
+      do_test (7, 1, i, BIG_CHAR);
+      do_test (3, 4, i, SMALL_CHAR);
+      do_test (4, 3, i, BIG_CHAR);
+      do_test (5, 7, i, SMALL_CHAR);
+      do_test (7, 5, i, SMALL_CHAR);
+    }
+
   return ret;
 }
 
diff --git a/sysdeps/powerpc/powerpc64/power7/strcpy.S b/sysdeps/powerpc/powerpc64/power7/strcpy.S
index ce71982eaf..115f98a304 100644
--- a/sysdeps/powerpc/powerpc64/power7/strcpy.S
+++ b/sysdeps/powerpc/powerpc64/power7/strcpy.S
@@ -31,8 +31,6 @@
 
    if (((((uintptr_t)dst & 0x7UL) == 0) && ((uintptr_t)src & 0x7UL) == 0))
      goto aligned_doubleword_copy;
-   if (((((uintptr_t)dst & 0x3UL) == 0) && ((uintptr_t)src & 0x3UL) == 0))
-     goto aligned_word_copy;
    if (((uintptr_t)dst & 0x7UL) == ((uintptr_t)src & 0x7UL))
      goto same_alignment;
    goto unaligned;
@@ -70,9 +68,18 @@ EALIGN (FUNC_NAME, 4, 0)
 #endif
 	or	rTMP, rSRC, rRTN
 	clrldi.	rTMP, rTMP, 61
-	bne	L(check_word_alignment)
+	bne	L(check_alignment)
 	b	L(aligned_doubleword_copy)
 
+	.align 4
+L(check_alignment):
+	rldicl	rRTNAL, rRTN, 0, 61
+	rldicl	rSRCAL, rSRC, 0, 61
+	cmpld	cr7, rSRCAL, rRTNAL
+	beq	cr7, L(same_alignment)
+	b	L(unaligned)
+
+	.align 4
 L(same_alignment):
 /* Src and dst with same alignment: align both to doubleword.  */
 	mr	rALCNT, rRTN
@@ -180,93 +187,249 @@ L(g1):
 #endif
 	blr
 
-L(check_word_alignment):
-	clrldi. rTMP, rTMP, 62
-	beq	L(aligned_word_copy)
-	rldicl	rRTNAL, rRTN, 0, 61
-	rldicl	rSRCAL, rSRC, 0, 61
-	cmpld	cr7, rSRCAL, rRTNAL
-	beq	cr7, L(same_alignment)
-	b	L(unaligned)
-
-/* For word aligned memory, operate using word load and stores.  */
 	.align	4
-L(aligned_word_copy):
-	li	rMASK, 0
-	addi	rRTN, rRTN, -4
-	lwz	rWORD, 0(rSRC)
-	b	L(g5)
+L(unaligned):
+	cmpdi	rSRCAL, 0		/* Check src alignment */
+	beq	L(srcaligndstunalign)
+	/* src is unaligned */
+	rlwinm	r10, rSRC, 3,26,28	/* Calculate padding.  */
+	clrrdi	rSRC, rSRC, 3		/* Align the addr to dw boundary */
+	ld	rWORD, 0(rSRC)		/* Load doubleword from memory.  */
+	li	rTMP, 0
+	/* Discard bits not part of the string */
+#ifdef __LITTLE_ENDIAN__
+	srd	rALT, rWORD, r10
+#else
+	sld	rALT, rWORD, r10
+#endif
+	cmpb	rTMP, rALT, rTMP	/* Compare each byte against null */
+	/* Discard bits not part of the string */
+#ifdef __LITTLE_ENDIAN__
+	sld	rTMP, rTMP, r10
+#else
+	srd	rTMP, rTMP, r10
+#endif
+	cmpdi	rTMP, 0
+	bne	L(bytebybyte)		/* if it has null, copy byte by byte */
+	subfic	r8, r9, 8
+	rlwinm	r5, rRTN, 3,26,28	/* Calculate padding in bits.  */
+	rldicl	r9, rRTN, 0, 61		/* Calculate padding in bytes. */
+	addi	rRTN, rRTN, -1
 
-	.align	4
-L(g3):	lwzu	rALT, 4(rSRC)
-	stwu	rWORD, 4(rRTN)
-	cmpb	rTMP, rALT, rMASK
-	cmpwi	rTMP, 0
-	bne	L(g4)
-	lwzu	rWORD, 4(rSRC)
-	stwu	rALT, 4(rRTN)
-L(g5):	cmpb	rTMP, rWORD, rMASK
-	cmpwi	rTMP, 0		/* If rTMP is 0, no null in word.  */
-	beq	L(g3)
-
-	mr      rALT, rWORD
-/* We've hit the end of the string.  Do the rest byte-by-byte.  */
-L(g4):
+	cmpdi	r5, 0			/* check dest alignment */
+	beq	L(srcunaligndstalign)
+
+	/* both src and dst unaligned */
 #ifdef __LITTLE_ENDIAN__
-	rlwinm.	rTMP, rALT, 0, 24, 31
-	stbu	rALT, 4(rRTN)
-	beqlr-
-	rlwinm.	rTMP, rALT, 24, 24, 31
-	stbu	rTMP, 1(rRTN)
-	beqlr-
-	rlwinm.	rTMP, rALT, 16, 24, 31
-	stbu	rTMP, 1(rRTN)
-	beqlr-
-	rlwinm	rTMP, rALT, 8, 24, 31
-	stbu	rTMP, 1(rRTN)
+	sld	rWORD, rALT, r10
+	mr 	r11, r10
+	addi	r11, r11, -8		/* Adjust byte pointer on loaded dw */
 #else
-	rlwinm. rTMP, rALT, 8, 24, 31
-	stbu    rTMP, 4(rRTN)
-	beqlr
-	rlwinm. rTMP, rALT, 16, 24, 31
-	stbu    rTMP, 1(rRTN)
-	beqlr
-	rlwinm. rTMP, rALT, 24, 24, 31
-	stbu    rTMP, 1(rRTN)
-	beqlr
-	stbu    rALT, 1(rRTN)
+	srd	rWORD, rALT, r10
+	subfic	r11, r10, 64
 #endif
-	blr
+	/* dst alignment is greater then src alignment? */
+	cmpd	cr7, r5, r10
+	blt	cr7, L(dst_align_small)
+	/* src alignment is less than dst */
 
-/* Oh well.  In this case, we just do a byte-by-byte copy.  */
-	.align	4
-L(unaligned):
-	lbz	rWORD, 0(rSRC)
-	addi	rRTN, rRTN, -1
-	cmpdi	rWORD, 0
-	beq	L(u2)
-
-	.align 	5
-L(u0):	lbzu	rALT, 1(rSRC)
-	stbu	rWORD, 1(rRTN)
-	cmpdi	rALT, 0
-	beq	L(u1)
-	lbzu	rWORD, 1(rSRC)
+	/* Calculate the dst alignment differnce */
+	subfic	rALT, r9, 8
+	mtctr	rALT
+
+	/* Write till dst is aligned */
+	cmpdi	rTMP, rALT, 4
+	blt	L(storebyte1)		/* less than 4, store byte by byte */
+	beq	L(equal1)		/* if its 4, store word */
+	addi	rTMP, rALT, -4		/* greater than 4, so stb and stw */
+	mtctr	rTMP
+L(storebyte1):
+#ifdef __LITTLE_ENDIAN__
+	addi	r11, r11, 8		/* Adjust byte pointer on loaded dw */
+#else
+	addi	r11, r11, -8
+#endif
+	srd	rALT, rWORD, r11
+	stbu	rALT, 1(rRTN)
+	bdnz	L(storebyte1)
+
+	subfic	rALT, r9, 8		/* Check the remaining bytes */
+	cmpdi	rTMP, rALT, 4
+	blt	L(proceed)
+
+	.align 4
+L(equal1):
+#ifdef __LITTLE_ENDIAN__
+	addi	r11, r11, 8		/* Adjust byte pointer on loaded dw */
+	srd	rALT, rWORD, r11
+#else
+	subfic	r11, r11, 64
+	sld	rALT, rWORD, r11
+	srdi	rALT, rALT, 32
+#endif
+	stw	rALT, 1(rRTN)
+	addi	rRTN, rRTN, 4
+
+L(proceed):
+	mr	rALT, rWORD
+	/* calculate the Left over bytes to be written */
+	subfic	r11, r10, 64
+	subfic	r5, r5, 64
+	subf	r5, r5, r11		/* remaining bytes on second dw */
+        subfic	r10, r5, 64		/* remaining bytes on first dw */
+	subfic	r9, r9, 8
+	subf	r8, r9, r8		/* recalculate padding */
+L(srcunaligndstalign):
+	addi	rRTN, rRTN, 1
+	subfic	r5, r10, 64		/* remaining bytes on second dw */
+	addi	rSRC, rSRC, 8
+	li	rTMP,0
+	b	L(storedouble)
+
+	.align 4
+L(dst_align_small):
+	mtctr	r8
+	/* Write till src is aligned */
+L(storebyte2):
+#ifdef __LITTLE_ENDIAN__
+	addi	r11, r11, 8		/* Adjust byte pointer on dw */
+#else
+	addi	r11, r11, -8
+#endif
+	srd	rALT, rWORD, r11
 	stbu	rALT, 1(rRTN)
-	cmpdi	rWORD, 0
-	beq	L(u2)
-	lbzu	rALT, 1(rSRC)
-	stbu	rWORD, 1(rRTN)
-	cmpdi	rALT, 0
-	beq	L(u1)
-	lbzu	rWORD, 1(rSRC)
+	bdnz	L(storebyte2)
+
+	addi	rSRC, rSRC, 8		/* Increment src pointer */
+	addi	rRTN, rRTN, 1		/* Increment dst pointer */
+	rldicl	r8, rRTN, 0, 61		/* Recalculate padding */
+
+	/* src is aligned */
+L(srcaligndstunalign):
+	ld	rWORD, 0(rSRC)
+	mr	rALT, rWORD
+	li	rTMP, 0			/* Check null */
+	cmpb	rTMP, rWORD, rTMP
+	cmpdi	rTMP, 0
+	bne	L(bytebybyte)		/* Do byte by byte if there is NULL */
+	rlwinm	r5, rRTN, 3,26,28	/* Calculate padding */
+	addi	rRTN, rRTN, -1
+	subfic	r10, r8, 8
+	/* write byte by byte till aligned */
+#ifdef __LITTLE_ENDIAN__
+	li	r11, -8
+#else
+	li	r11, 64
+#endif
+	mtctr	r10
+	cmpdi	rTMP, r10, 4
+	blt	L(storebyte)
+	beq	L(equal)
+	addi	rTMP, r10, -4
+	mtctr	rTMP
+L(storebyte):
+#ifdef __LITTLE_ENDIAN__
+	addi	r11, r11, 8		/* Adjust byte pointer on  dw */
+#else
+	addi	r11, r11, -8
+#endif
+	srd	rALT, rWORD, r11
 	stbu	rALT, 1(rRTN)
-	cmpdi	rWORD, 0
-	bne	L(u0)
-L(u2):	stbu	rWORD, 1(rRTN)
-	blr
-L(u1):	stbu	rALT, 1(rRTN)
-	blr
+	bdnz	L(storebyte)
+
+	cmpdi	rTMP, r10, 4
+	blt	L(align)
+
+	.align 4
+L(equal):
+#ifdef __LITTLE_ENDIAN__
+	addi	r11, r11, 8
+	srd	rALT, rWORD, r11
+#else
+	subfic	r11, r11, 64
+	sld	rALT, rWORD, r11
+	srdi	rALT, rALT, 32
+#endif
+	stw	rALT, 1(rRTN)
+	addi	rRTN, rRTN, 4
+L(align):
+	addi	rRTN, rRTN, 1
+	addi	rSRC, rSRC, 8		/* Increment src pointer */
+	subfic	r10, r5, 64
+	li	rTMP, 0
+	/* dst addr aligned to 8 */
+L(storedouble):
+	ld	rALT, 0(rSRC)		/* load next dw */
+	cmpb	rTMP, rALT, rTMP
+	cmpdi	rTMP, 0			/* check for null on each new dw */
+	bne	L(null)
+#ifdef __LITTLE_ENDIAN__
+	srd	r9, rWORD, r10		/* bytes from first dw */
+	sld	r11, rALT, r5		/* bytes from second dw */
+#else
+	sld	r9, rWORD, r10
+	srd	r11, rALT, r5
+#endif
+	or	r11, r9, r11		/* make as a single dw */
+	std	r11, 0(rRTN)		/* store as std on aligned addr */
+	mr	rWORD, rALT		/* still few bytes left to be written */
+	addi	rRTN, rRTN, 8		/* increment dst addr */
+	addi	rSRC, rSRC, 8		/* increment src addr */
+	b	L(storedouble)		/* Loop till NULL */
+
+	.align 4
+
+/* We've hit the end of the string.  Do the rest byte-by-byte.  */
+L(null):
+	addi	rRTN, rRTN, -1
+	mr	r10, r5
+	mtctr	r8
+#ifdef __LITTLE_ENDIAN__
+	subfic	r10, r10, 64
+	addi	r10, r10, -8
+#endif
+	cmpdi	rTMP, r8, 4
+	blt	L(loop)
+
+	/* we can still use stw if leftover >= 4*/
+#ifdef __LITTLE_ENDIAN__
+	addi	r10, r10, 8
+	srd	r11, rWORD, r10
+#else
+	subfic	r10, r10, 64
+	sld	r11, rWORD, r10
+	srdi	r11, r11, 32
+#endif
+	stw	r11, 1(rRTN)
+	addi	rRTN, rRTN, 4
+
+	beq	L(bytebybyte1)
+	addi	r10, r10, 32
+#ifdef __LITTLE_ENDIAN__
+	addi	r10, r10, -8
+#else
+	subfic	r10, r10, 64
+#endif
+	addi	rTMP, r8, -4
+	mtctr	rTMP
+	/* remaining byte by byte part of first dw */
+L(loop):
+#ifdef __LITTLE_ENDIAN__
+	addi	r10, r10, 8
+#else
+	addi	r10, r10, -8
+#endif
+	srd	rTMP, rWORD, r10
+	stbu	rTMP, 1(rRTN)
+	bdnz	L(loop)
+
+L(bytebybyte1):
+	addi	rRTN, rRTN, 1
+	/* remaining byte by byte part of second dw */
+L(bytebybyte):
+	addi	rRTN, rRTN, -8
+	b	L(g1)
+
 END (FUNC_NAME)
 
 #ifndef USE_AS_STPCPY