about summary refs log tree commit diff
path: root/sysdeps/powerpc/powerpc32/power7/memchr.S
diff options
context:
space:
mode:
authorAlan Modra <amodra@gmail.com>2013-08-17 18:48:36 +0930
committerAlan Modra <amodra@gmail.com>2013-10-04 10:41:46 +0930
commit466b03933234017473c12dd1d92bda5e7fe49df7 (patch)
treeb70778371ad8c8e5a1ab77dee110d8711618abfd /sysdeps/powerpc/powerpc32/power7/memchr.S
parent3be87c77d24c4456ccca4034363b6d1814cd0c84 (diff)
downloadglibc-466b03933234017473c12dd1d92bda5e7fe49df7.tar.gz
glibc-466b03933234017473c12dd1d92bda5e7fe49df7.tar.xz
glibc-466b03933234017473c12dd1d92bda5e7fe49df7.zip
PowerPC LE memchr and memrchr
http://sourceware.org/ml/libc-alpha/2013-08/msg00105.html

Like strnlen, memchr and memrchr had a number of defects fixed by this
patch as well as adding little-endian support.  The first one I
noticed was that the entry to the main loop needlessly checked for
"are we done yet?" when we know the size is large enough that we can't
be done.  The second defect I noticed was that the main loop count was
wrong, which in turn meant that the small loop needed to handle an
extra word.  Thirdly, there is nothing to say that the string can't
wrap around zero, except of course that we'd normally hit a segfault
on trying to read from address zero.  Fixing that simplified a number
of places:

-	/* Are we done already?  */
-	addi    r9,r8,8
-	cmpld	r9,r7
-	bge	L(null)

becomes

+	cmpld	r8,r7
+	beqlr

However, the exit gets an extra test because I test for being on the
last word then if so whether the byte offset is less than the end.
Overall, the change is a win.

Lastly, memrchr used the wrong cache hint.

	* sysdeps/powerpc/powerpc64/power7/memchr.S: Replace rlwimi with
	insrdi.  Make better use of reg selection to speed exit slightly.
	Schedule entry path a little better.  Remove useless "are we done"
	checks on entry to main loop.  Handle wrapping around zero address.
	Correct main loop count.  Handle single left-over word from main
	loop inline rather than by using loop_small.  Remove extra word
	case in loop_small caused by wrong loop count.  Add little-endian
	support.
	* sysdeps/powerpc/powerpc32/power7/memchr.S: Likewise.
	* sysdeps/powerpc/powerpc64/power7/memrchr.S: Likewise.  Use proper
	cache hint.
	* sysdeps/powerpc/powerpc32/power7/memrchr.S: Likewise.
	* sysdeps/powerpc/powerpc64/power7/rawmemchr.S: Add little-endian
	support.  Avoid rlwimi.
	* sysdeps/powerpc/powerpc32/power7/rawmemchr.S: Likewise.
Diffstat (limited to 'sysdeps/powerpc/powerpc32/power7/memchr.S')
-rw-r--r--sysdeps/powerpc/powerpc32/power7/memchr.S185
1 files changed, 82 insertions, 103 deletions
diff --git a/sysdeps/powerpc/powerpc32/power7/memchr.S b/sysdeps/powerpc/powerpc32/power7/memchr.S
index 369e5e0483..85754f3f17 100644
--- a/sysdeps/powerpc/powerpc32/power7/memchr.S
+++ b/sysdeps/powerpc/powerpc32/power7/memchr.S
@@ -25,107 +25,111 @@ ENTRY (__memchr)
 	CALL_MCOUNT
 	dcbt	0,r3
 	clrrwi  r8,r3,2
-	rlwimi	r4,r4,8,16,23
-	rlwimi	r4,r4,16,0,15
+	insrdi	r4,r4,8,48
 	add	r7,r3,r5      /* Calculate the last acceptable address.  */
+	insrdi	r4,r4,16,32
 	cmplwi	r5,16
+	li	r9, -1
+	rlwinm	r6,r3,3,27,28 /* Calculate padding.  */
+	addi	r7,r7,-1
+#ifdef __LITTLE_ENDIAN__
+	slw	r9,r9,r6
+#else
+	srw	r9,r9,r6
+#endif
 	ble	L(small_range)
 
-	cmplw	cr7,r3,r7     /* Compare the starting address (r3) with the
-				 ending address (r7).  If (r3 >= r7), the size
-				 passed in is zero or negative.  */
-	ble	cr7,L(proceed)
-
-	li	r7,-1	      /* Artificially set our ending address (r7)
-				 such that we will exit early. */
-L(proceed):
-	rlwinm	r6,r3,3,27,28 /* Calculate padding.  */
-	cmpli	cr6,r6,0      /* cr6 == Do we have padding?  */
 	lwz	r12,0(r8)     /* Load word from memory.  */
-	cmpb	r10,r12,r4    /* Check for BYTEs in WORD1.  */
-	beq	cr6,L(proceed_no_padding)
-	slw	r10,r10,r6
-	srw	r10,r10,r6
-L(proceed_no_padding):
-	cmplwi	cr7,r10,0     /* If r10 == 0, no BYTEs have been found.  */
+	cmpb	r3,r12,r4     /* Check for BYTEs in WORD1.  */
+	and	r3,r3,r9
+	clrlwi	r5,r7,30      /* Byte count - 1 in last word.  */
+	clrrwi	r7,r7,2       /* Address of last word.  */
+	cmplwi	cr7,r3,0      /* If r3 == 0, no BYTEs have been found.  */
 	bne	cr7,L(done)
 
-	/* Are we done already?  */
-	addi	r9,r8,4
-	cmplw	cr6,r9,r7
-	bge	cr6,L(null)
-
 	mtcrf   0x01,r8
 	/* Are we now aligned to a doubleword boundary?  If so, skip to
 	   the main loop.  Otherwise, go through the alignment code.  */
-
 	bt	29,L(loop_setup)
 
 	/* Handle WORD2 of pair.  */
 	lwzu	r12,4(r8)
-	cmpb	r10,r12,r4
-	cmplwi	cr7,r10,0
+	cmpb	r3,r12,r4
+	cmplwi	cr7,r3,0
 	bne	cr7,L(done)
 
-	/* Are we done already?  */
-	addi	r9,r8,4
-	cmplw	cr6,r9,r7
-	bge	cr6,L(null)
-
 L(loop_setup):
-	sub	r5,r7,r9
-	srwi	r6,r5,3	      /* Number of loop iterations.  */
+	/* The last word we want to read in the loop below is the one
+	   containing the last byte of the string, ie. the word at
+	   (s + size - 1) & ~3, or r7.  The first word read is at
+	   r8 + 4, we read 2 * cnt words, so the last word read will
+	   be at r8 + 4 + 8 * cnt - 4.  Solving for cnt gives
+	   cnt = (r7 - r8) / 8  */
+	sub	r6,r7,r8
+	srwi	r6,r6,3	      /* Number of loop iterations.  */
 	mtctr	r6            /* Setup the counter.  */
-	b	L(loop)
-	/* Main loop to look for BYTE backwards in the string.  Since
-	   it's a small loop (< 8 instructions), align it to 32-bytes.  */
-	.p2align  5
+
+	/* Main loop to look for BYTE in the string.  Since
+	   it's a small loop (8 instructions), align it to 32-bytes.  */
+	.align	5
 L(loop):
 	/* Load two words, compare and merge in a
 	   single register for speed.  This is an attempt
 	   to speed up the byte-checking process for bigger strings.  */
 	lwz	r12,4(r8)
 	lwzu	r11,8(r8)
-	cmpb	r10,r12,r4
+	cmpb	r3,r12,r4
 	cmpb	r9,r11,r4
-	or	r5,r9,r10     /* Merge everything in one word.  */
-	cmplwi	cr7,r5,0
+	or	r6,r9,r3      /* Merge everything in one word.  */
+	cmplwi	cr7,r6,0
 	bne	cr7,L(found)
 	bdnz	L(loop)
 
-	/* We're here because the counter reached 0, and that means we
-	   didn't have any matches for BYTE in the whole range.  */
-	subi	r11,r7,4
-	cmplw	cr6,r8,r11
-	blt	cr6,L(loop_small)
-	b	L(null)
+	/* We may have one more dword to read.  */
+	cmplw	r8,r7
+	beqlr
+
+	lwzu	r12,4(r8)
+	cmpb	r3,r12,r4
+	cmplwi	cr6,r3,0
+	bne	cr6,L(done)
+	blr
 
+	.align	4
+L(found):
 	/* OK, one (or both) of the words contains BYTE.  Check
 	   the first word and decrement the address in case the first
 	   word really contains BYTE.  */
-	.align	4
-L(found):
-	cmplwi	cr6,r10,0
+	cmplwi	cr6,r3,0
 	addi	r8,r8,-4
 	bne	cr6,L(done)
 
 	/* BYTE must be in the second word.  Adjust the address
-	   again and move the result of cmpb to r10 so we can calculate the
+	   again and move the result of cmpb to r3 so we can calculate the
 	   pointer.  */
 
-	mr	r10,r9
+	mr	r3,r9
 	addi	r8,r8,4
 
-	/* r10 has the output of the cmpb instruction, that is, it contains
+	/* r3 has the output of the cmpb instruction, that is, it contains
 	   0xff in the same position as BYTE in the original
 	   word from the string.  Use that to calculate the pointer.
 	   We need to make sure BYTE is *before* the end of the range.  */
 L(done):
-	cntlzw	r0,r10	      /* Count leading zeroes before the match.  */
-	srwi	r0,r0,3	      /* Convert leading zeroes to bytes.  */
+#ifdef __LITTLE_ENDIAN__
+	addi    r0,r3,-1
+	andc    r0,r0,r3
+	popcntw	r0,r0	      /* Count trailing zeros.  */
+#else
+	cntlzw	r0,r3	      /* Count leading zeros before the match.  */
+#endif
+	cmplw	r8,r7         /* Are we on the last word?  */
+	srwi	r0,r0,3	      /* Convert leading/trailing zeros to bytes.  */
 	add	r3,r8,r0
-	cmplw	r3,r7
-	bge	L(null)
+	cmplw	cr7,r0,r5     /* If on the last dword, check byte offset.  */
+	bnelr
+	blelr	cr7
+	li	r3,0
 	blr
 
 	.align	4
@@ -137,67 +141,42 @@ L(null):
 	.align	4
 L(small_range):
 	cmplwi	r5,0
-	rlwinm	r6,r3,3,27,28 /* Calculate padding.  */
-	beq	L(null)       /* This branch is for the cmplwi r5,0 above */
+	beq	L(null)
 	lwz	r12,0(r8)     /* Load word from memory.  */
-	cmplwi	cr6,r6,0      /* cr6 == Do we have padding?  */
-	cmpb	r10,r12,r4    /* Check for BYTE in DWORD1.  */
-	beq	cr6,L(small_no_padding)
-	slw	r10,r10,r6
-	srw	r10,r10,r6
-L(small_no_padding):
-	cmplwi	cr7,r10,0
+	cmpb	r3,r12,r4     /* Check for BYTE in DWORD1.  */
+	and	r3,r3,r9
+	cmplwi	cr7,r3,0
+	clrlwi	r5,r7,30      /* Byte count - 1 in last word.  */
+	clrrwi	r7,r7,2       /* Address of last word.  */
+	cmplw	r8,r7         /* Are we done already?  */
 	bne	cr7,L(done)
+	beqlr
 
-	/* Are we done already?  */
-	addi    r9,r8,4
-	cmplw	r9,r7
-	bge	L(null)
-
-L(loop_small):                /* loop_small has been unrolled.  */
 	lwzu	r12,4(r8)
-	cmpb	r10,r12,r4
-	addi	r9,r8,4
-	cmplwi	cr6,r10,0
-	cmplw	r9,r7
+	cmpb	r3,r12,r4
+	cmplwi	cr6,r3,0
+	cmplw	r8,r7
 	bne	cr6,L(done)
-	bge	L(null)
+	beqlr
 
 	lwzu	r12,4(r8)
-	cmpb	r10,r12,r4
-	addi	r9,r8,4
-	cmplwi	cr6,r10,0
-	cmplw	r9,r7
+	cmpb	r3,r12,r4
+	cmplwi	cr6,r3,0
+	cmplw	r8,r7
 	bne	cr6,L(done)
-	bge	L(null)
+	beqlr
 
 	lwzu	r12,4(r8)
-	cmpb	r10,r12,r4
-	addi	r9,r8,4
-	cmplwi	cr6,r10,0
-	cmplw	r9,r7
+	cmpb	r3,r12,r4
+	cmplwi	cr6,r3,0
+	cmplw	r8,r7
 	bne	cr6,L(done)
-	bge	L(null)
+	beqlr
 
 	lwzu	r12,4(r8)
-	cmpb	r10,r12,r4
-	addi	r9,r8,4
-	cmplwi	cr6,r10,0
-	cmplw	r9,r7
+	cmpb	r3,r12,r4
+	cmplwi	cr6,r3,0
 	bne	cr6,L(done)
-	bge	L(null)
-
-	/* For most cases we will never get here.  Under some combinations of
-	   padding + length there is a leftover word that still needs to be
-	   checked.  */
-	lwzu	r12,4(r8)
-	cmpb	r10,r12,r4
-	addi	r9,r8,4
-	cmplwi	cr6,r10,0
-	bne	cr6,L(done)
-
-	/* save a branch and exit directly */
-	li	r3,0
 	blr
 
 END (__memchr)