GNU C Library master sources branch roland/arm-memcpy created. glibc-2.17-823-gd5eebff

[roland at sourceware dot org]

This is an automated email from the git hooks/post-receive script. It was
generated because a ref change was pushed to the repository containing
the project "GNU C Library master sources".

The branch, roland/arm-memcpy has been created
        at  d5eebff9fdd4cd6609e16eccf40a3499f8cef870 (commit)

- Log -----------------------------------------------------------------
http://sourceware.org/git/gitweb.cgi?p=glibc.git;a=commitdiff;h=d5eebff9fdd4cd6609e16eccf40a3499f8cef870

commit d5eebff9fdd4cd6609e16eccf40a3499f8cef870
Author: Roland McGrath <roland@hack.frob.com>
Date:   Mon Jun 17 11:12:05 2013 -0700

    unfinished

diff --git a/ports/sysdeps/arm/armv7/multiarch/memcpy_impl.S b/ports/sysdeps/arm/armv7/multiarch/memcpy_impl.S
index f83276a..902b650 100644
--- a/ports/sysdeps/arm/armv7/multiarch/memcpy_impl.S
+++ b/ports/sysdeps/arm/armv7/multiarch/memcpy_impl.S
@@ -33,6 +33,7 @@
 #define NO_THUMB
 #endif
 #include <sysdep.h>
+#include <arm-features.h>
 
 	.syntax unified
 	/* This implementation requires ARM state.  */
@@ -73,6 +74,122 @@
 #define dst	ip
 #define tmp2	r10
 
+/* These two macros both work by repeated invocation of the macro
+   dispatch_step (not defined here).  That macro performs one "step",
+   doing one load instruction and one store instruction to copy one
+   "unit".  On entry, TMP1 contains the number of bytes to be copied,
+   a multiple of the unit size.  The macro clobbers TMP1 in the
+   process of doing a computed jump to the tail containing the
+   appropriate number of steps.
+
+   In dispatch_7_dword, dispatch_step is invoked seven times, with an
+   argument that is 7 for the first and 1 for the last.  Units are
+   double-words (8 bytes).  TMP1 is at most 56.
+
+   In dispatch_15_word, dispatch_step is invoked fifteen times,
+   with an argument that is 15 for the first and 1 for the last.
+   Units are words (4 bytes).  TMP1 is at most 60.  */
+
+#ifndef ARM_ALWAYS_BX
+# if ARM_BX_ALIGN_LOG2 != 2
+#  error case not handled
+# endif
+	.macro dispatch_7_dword
+	rsb	tmp1, tmp1, #((7 * 8) - PC_OFS + INSN_SIZE)
+	add	pc, pc, tmp1
+	dispatch_step 7
+	dispatch_step 6
+	dispatch_step 5
+	dispatch_step 4
+	dispatch_step 3
+	dispatch_step 2
+	dispatch_step 1
+	.purgem dispatch_step
+	.endm
+
+	.macro dispatch_15_word
+	rsb	tmp1, tmp1, #((15 * 4) - PC_OFS/2 + INSN_SIZE/2)
+	add	pc, pc, tmp1, lsl #1
+	dispatch_step 15
+	dispatch_step 14
+	dispatch_step 13
+	dispatch_step 12
+	dispatch_step 11
+	dispatch_step 10
+	dispatch_step 9
+	dispatch_step 8
+	dispatch_step 7
+	dispatch_step 6
+	dispatch_step 5
+	dispatch_step 4
+	dispatch_step 3
+	dispatch_step 2
+	dispatch_step 1
+	.purgem dispatch_step
+	.endm
+#else
+# if ARM_BX_ALIGN_LOG2 < 3
+#  error case not handled
+# endif
+# error unfinished
+	.macro dispatch_7_dword
+	rsb	tmp1, tmp1, #((7 * 8) - PC_OFS + INSN_SIZE)
+	add	pc, pc, tmp1
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 7
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 6
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 5
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 4
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 3
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 2
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 1
+	.purgem dispatch_step
+	.endm
+
+	.macro dispatch_15_word
+	rsb	tmp1, tmp1, #((15 * 4) - PC_OFS/2 + INSN_SIZE/2)
+	add	pc, pc, tmp1, lsl #1
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 15
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 14
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 13
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 12
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 11
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 10
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 9
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 8
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 7
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 6
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 5
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 4
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 3
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 2
+	.p2align ARM_BX_ALIGN_LOG2
+	dispatch_step 1
+	.purgem dispatch_step
+	.endm
+
+#endif
+
 #ifndef USE_NEON
 /* For bulk copies using GP registers.  */
 #define	A_l	r2		/* Call-clobbered.  */
@@ -140,23 +257,22 @@ ENTRY(memcpy)
 
 .Ltail63unaligned:
 #ifdef USE_NEON
-	and	tmp1, count, #0x38
-	rsb	tmp1, tmp1, #(56 - PC_OFS + INSN_SIZE)
-	add	pc, pc, tmp1
-	vld1.8	{d0}, [src]!	/* 14 words to go.  */
-	vst1.8	{d0}, [dst]!
-	vld1.8	{d0}, [src]!	/* 12 words to go.  */
-	vst1.8	{d0}, [dst]!
-	vld1.8	{d0}, [src]!	/* 10 words to go.  */
-	vst1.8	{d0}, [dst]!
-	vld1.8	{d0}, [src]!	/* 8 words to go.  */
-	vst1.8	{d0}, [dst]!
-	vld1.8	{d0}, [src]!	/* 6 words to go.  */
-	vst1.8	{d0}, [dst]!
-	vld1.8	{d0}, [src]!	/* 4 words to go.  */
-	vst1.8	{d0}, [dst]!
-	vld1.8	{d0}, [src]!	/* 2 words to go.  */
+	/* These need an extra layer of macro just to work around a
+	   bug in the assembler's parser when an operand starts with
+	   a {...}.  */
+	.macro neon_load_d0 src
+	vld1.8	{d0}, [src]!
+	.endm
+	.macro neon_store_d0 dst
 	vst1.8	{d0}, [dst]!
+	.endm
+
+	and	tmp1, count, #0x38
+	.macro dispatch_step i
+	sfi_breg src, neon_load_d0 src
+	sfi_breg dst, neon_store_d0 dst
+	.endm
+	dispatch_7_dword
 
 	tst	count, #4
 	ldrne	tmp1, [src], #4
@@ -167,47 +283,14 @@ ENTRY(memcpy)
 	and	tmp1, count, #0x3c
 	add	dst, dst, tmp1
 	add	src, src, tmp1
-	rsb	tmp1, tmp1, #(60 - PC_OFS/2 + INSN_SIZE/2)
 	/* Jump directly into the sequence below at the correct offset.  */
-	add	pc, pc, tmp1, lsl #1
-
-	ldr	tmp1, [src, #-60]	/* 15 words to go.  */
-	str	tmp1, [dst, #-60]
-
-	ldr	tmp1, [src, #-56]	/* 14 words to go.  */
-	str	tmp1, [dst, #-56]
-	ldr	tmp1, [src, #-52]
-	str	tmp1, [dst, #-52]
-
-	ldr	tmp1, [src, #-48]	/* 12 words to go.  */
-	str	tmp1, [dst, #-48]
-	ldr	tmp1, [src, #-44]
-	str	tmp1, [dst, #-44]
-
-	ldr	tmp1, [src, #-40]	/* 10 words to go.  */
-	str	tmp1, [dst, #-40]
-	ldr	tmp1, [src, #-36]
-	str	tmp1, [dst, #-36]
-
-	ldr	tmp1, [src, #-32]	/* 8 words to go.  */
-	str	tmp1, [dst, #-32]
-	ldr	tmp1, [src, #-28]
-	str	tmp1, [dst, #-28]
-
-	ldr	tmp1, [src, #-24]	/* 6 words to go.  */
-	str	tmp1, [dst, #-24]
-	ldr	tmp1, [src, #-20]
-	str	tmp1, [dst, #-20]
-
-	ldr	tmp1, [src, #-16]	/* 4 words to go.  */
-	str	tmp1, [dst, #-16]
-	ldr	tmp1, [src, #-12]
-	str	tmp1, [dst, #-12]
-
-	ldr	tmp1, [src, #-8]	/* 2 words to go.  */
-	str	tmp1, [dst, #-8]
-	ldr	tmp1, [src, #-4]
-	str	tmp1, [dst, #-4]
+	.macro dispatch_step i
+	sfi_breg src, \
+	ldr	tmp1, [src, #-(\i * 4)]
+	sfi_breg dst, \
+	str	tmp1, [dst, #-(\i * 4)]
+	.endm
+	dispatch_15_word
 #endif
 
 	lsls	count, count, #31
@@ -287,23 +370,13 @@ ENTRY(memcpy)
 	and	tmp1, tmp2, #0x38
 	add	dst, dst, tmp1
 	add	src, src, tmp1
-	rsb	tmp1, tmp1, #(56 - PC_OFS + INSN_SIZE)
-	add	pc, pc, tmp1
-
-	vldr	d0, [src, #-56]	/* 14 words to go.  */
-	vstr	d0, [dst, #-56]
-	vldr	d0, [src, #-48]	/* 12 words to go.  */
-	vstr	d0, [dst, #-48]
-	vldr	d0, [src, #-40]	/* 10 words to go.  */
-	vstr	d0, [dst, #-40]
-	vldr	d0, [src, #-32]	/* 8 words to go.  */
-	vstr	d0, [dst, #-32]
-	vldr	d0, [src, #-24]	/* 6 words to go.  */
-	vstr	d0, [dst, #-24]
-	vldr	d0, [src, #-16]	/* 4 words to go.  */
-	vstr	d0, [dst, #-16]
-	vldr	d0, [src, #-8]	/* 2 words to go.  */
-	vstr	d0, [dst, #-8]
+	.macro dispatch_step i
+	sfi_breg src, \
+	vldr	d0, [src, #-(\i * 8)]
+	sfi_breg dst, \
+	vstr	d0, [dst, #-(\i * 8)]
+	.endm
+	dispatch_7_dword
 #else
 	sub	src, src, #8
 	sub	dst, dst, #8
@@ -349,24 +422,15 @@ ENTRY(memcpy)
 	and	tmp1, tmp2, #0x38
 	add	dst, dst, tmp1
 	add	src, src, tmp1
-	rsb	tmp1, tmp1, #(56 - PC_OFS + INSN_SIZE)
-	add	pc, pc, tmp1
-	ldrd	A_l, A_h, [src, #-56]	/* 14 words to go.  */
-	strd	A_l, A_h, [dst, #-56]
-	ldrd	A_l, A_h, [src, #-48]	/* 12 words to go.  */
-	strd	A_l, A_h, [dst, #-48]
-	ldrd	A_l, A_h, [src, #-40]	/* 10 words to go.  */
-	strd	A_l, A_h, [dst, #-40]
-	ldrd	A_l, A_h, [src, #-32]	/* 8 words to go.  */
-	strd	A_l, A_h, [dst, #-32]
-	ldrd	A_l, A_h, [src, #-24]	/* 6 words to go.  */
-	strd	A_l, A_h, [dst, #-24]
-	ldrd	A_l, A_h, [src, #-16]	/* 4 words to go.  */
-	strd	A_l, A_h, [dst, #-16]
-	ldrd	A_l, A_h, [src, #-8]	/* 2 words to go.  */
-	strd	A_l, A_h, [dst, #-8]
-
+	.macro dispatch_step i
+	sfi_breg src, \
+	ldrd	A_l, A_h, [src, #-(\i * 8)]
+	sfi_breg dst, \
+	strd	A_l, A_h, [dst, #-(\i * 8)]
+	.endm
+	dispatch_7_dword
 #endif
+
 	tst	tmp2, #4
 	ldrne	tmp1, [src], #4
 	strne	tmp1, [dst], #4

-----------------------------------------------------------------------


hooks/post-receive
-- 
GNU C Library master sources