[PATCH] x86_64: Undo the earlier changes to remove unrolled copy/memset functions

They cause quite bad performance regressions on Netburst
This is temporary until we can get new optimized functions
for these CPUs.

This undoes changes that were done in 2.6.15 and in 2.6.16-rc1,
essentially bringing the code back to 2.6.14 level. Only change
is I renamed the X86_FEATURE_K8_C flag to X86_FEATURE_REP_GOOD
and fixed the check for the flag and also fixed some comments.

Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

arch/x86_64/kernel/setup.c

@@ -877,6 +877,7 @@ static void __init amd_detect_cmp(struct cpuinfo_x86 *c)
 static int __init init_amd(struct cpuinfo_x86 *c)
 {
 	int r;
+	unsigned level;
 
 #ifdef CONFIG_SMP
 	unsigned long value;
@@ -899,6 +900,11 @@ static int __init init_amd(struct cpuinfo_x86 *c)
 	   3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway */
 	clear_bit(0*32+31, &c->x86_capability);
 	
+	/* On C+ stepping K8 rep microcode works well for copy/memset */
+	level = cpuid_eax(1);
+	if (c->x86 == 15 && ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58))
+		set_bit(X86_FEATURE_REP_GOOD, &c->x86_capability);
+
 	r = get_model_name(c);
 	if (!r) { 
 		switch (c->x86) { 

arch/x86_64/lib/clear_page.S

@@ -5,8 +5,46 @@
 	.globl clear_page
 	.p2align 4
 clear_page:
+	xorl   %eax,%eax
+	movl   $4096/64,%ecx
+	.p2align 4
+.Lloop:
+	decl	%ecx
+#define PUT(x) movq %rax,x*8(%rdi)
+	movq %rax,(%rdi)
+	PUT(1)
+	PUT(2)
+	PUT(3)
+	PUT(4)
+	PUT(5)
+	PUT(6)
+	PUT(7)
+	leaq	64(%rdi),%rdi
+	jnz	.Lloop
+	nop
+	ret
+clear_page_end:
+
+	/* Some CPUs run faster using the string instructions.
+	   It is also a lot simpler. Use this when possible */
+
+#include <asm/cpufeature.h>
+
+	.section .altinstructions,"a"
+	.align 8
+	.quad  clear_page
+	.quad  clear_page_c
+	.byte  X86_FEATURE_REP_GOOD
+	.byte  clear_page_end-clear_page
+	.byte  clear_page_c_end-clear_page_c
+	.previous
+
+	.section .altinstr_replacement,"ax"
+clear_page_c:
 	movl $4096/8,%ecx
 	xorl %eax,%eax
 	rep 
 	stosq
 	ret
+clear_page_c_end:
+	.previous

arch/x86_64/lib/copy_page.S

@@ -8,7 +8,94 @@
 	.globl copy_page
 	.p2align 4
 copy_page:
+	subq	$3*8,%rsp
+	movq	%rbx,(%rsp)
+	movq	%r12,1*8(%rsp)
+	movq	%r13,2*8(%rsp)
+
+	movl	$(4096/64)-5,%ecx
+	.p2align 4
+.Loop64:
+  	dec     %rcx
+
+	movq        (%rsi), %rax
+	movq      8 (%rsi), %rbx
+	movq     16 (%rsi), %rdx
+	movq     24 (%rsi), %r8
+	movq     32 (%rsi), %r9
+	movq     40 (%rsi), %r10
+	movq     48 (%rsi), %r11
+	movq     56 (%rsi), %r12
+
+	prefetcht0 5*64(%rsi)
+
+	movq     %rax,    (%rdi)
+	movq     %rbx,  8 (%rdi)
+	movq     %rdx, 16 (%rdi)
+	movq     %r8,  24 (%rdi)
+	movq     %r9,  32 (%rdi)
+	movq     %r10, 40 (%rdi)
+	movq     %r11, 48 (%rdi)
+	movq     %r12, 56 (%rdi)
+
+	leaq    64 (%rsi), %rsi
+	leaq    64 (%rdi), %rdi
+
+	jnz     .Loop64
+
+	movl	$5,%ecx
+	.p2align 4
+.Loop2:
+	decl   %ecx
+
+	movq        (%rsi), %rax
+	movq      8 (%rsi), %rbx
+	movq     16 (%rsi), %rdx
+	movq     24 (%rsi), %r8
+	movq     32 (%rsi), %r9
+	movq     40 (%rsi), %r10
+	movq     48 (%rsi), %r11
+	movq     56 (%rsi), %r12
+
+	movq     %rax,    (%rdi)
+	movq     %rbx,  8 (%rdi)
+	movq     %rdx, 16 (%rdi)
+	movq     %r8,  24 (%rdi)
+	movq     %r9,  32 (%rdi)
+	movq     %r10, 40 (%rdi)
+	movq     %r11, 48 (%rdi)
+	movq     %r12, 56 (%rdi)
+
+	leaq	64(%rdi),%rdi
+	leaq	64(%rsi),%rsi
+
+	jnz	.Loop2
+
+	movq	(%rsp),%rbx
+	movq	1*8(%rsp),%r12
+	movq	2*8(%rsp),%r13
+	addq	$3*8,%rsp
+	ret
+
+	/* Some CPUs run faster using the string copy instructions.
+	   It is also a lot simpler. Use this when possible */
+
+#include <asm/cpufeature.h>
+
+	.section .altinstructions,"a"
+	.align 8
+	.quad  copy_page
+	.quad  copy_page_c
+	.byte  X86_FEATURE_REP_GOOD
+	.byte  copy_page_c_end-copy_page_c
+	.byte  copy_page_c_end-copy_page_c
+	.previous
+
+	.section .altinstr_replacement,"ax"
+copy_page_c:
 	movl $4096/8,%ecx
 	rep 
 	movsq 
 	ret
+copy_page_c_end:
+	.previous

arch/x86_64/lib/copy_user.S

@@ -4,9 +4,12 @@
  * Functions to copy from and to user space.		
  */		 
 
+#define FIX_ALIGNMENT 1
+
 	#include <asm/current.h>
 	#include <asm/asm-offsets.h>
 	#include <asm/thread_info.h>
+	#include <asm/cpufeature.h>
 
 /* Standard copy_to_user with segment limit checking */		
 	.globl copy_to_user
@@ -18,7 +21,23 @@ copy_to_user:
 	jc  bad_to_user
 	cmpq threadinfo_addr_limit(%rax),%rcx
 	jae bad_to_user
-	jmp copy_user_generic
+2:
+	.byte 0xe9	/* 32bit jump */
+	.long .Lcug-1f
+1:
+
+	.section .altinstr_replacement,"ax"
+3:	.byte 0xe9			/* replacement jmp with 8 bit immediate */
+	.long copy_user_generic_c-1b	/* offset */
+	.previous
+	.section .altinstructions,"a"
+	.align 8
+	.quad  2b
+	.quad  3b
+	.byte  X86_FEATURE_REP_GOOD
+	.byte  5
+	.byte  5
+	.previous
 
 /* Standard copy_from_user with segment limit checking */	
 	.globl copy_from_user
@@ -53,44 +72,230 @@ bad_to_user:
  * rsi source
  * rdx count
  *
- * Only 4GB of copy is supported. This shouldn't be a problem
- * because the kernel normally only writes from/to page sized chunks
- * even if user space passed a longer buffer.
- * And more would be dangerous because both Intel and AMD have
- * errata with rep movsq > 4GB. If someone feels the need to fix
- * this please consider this.
- *
  * Output:		
  * eax uncopied bytes or 0 if successful.
  */
-
 	.globl copy_user_generic
+	.p2align 4
 copy_user_generic:
+	.byte 0x66,0x66,0x90	/* 5 byte nop for replacement jump */
+	.byte 0x66,0x90
+1:
+	.section .altinstr_replacement,"ax"
+2:	.byte 0xe9	             /* near jump with 32bit immediate */
+	.long copy_user_generic_c-1b /* offset */
+	.previous
+	.section .altinstructions,"a"
+	.align 8
+	.quad  copy_user_generic
+	.quad  2b
+	.byte  X86_FEATURE_REP_GOOD
+	.byte  5
+	.byte  5
+	.previous
+.Lcug:
+	pushq %rbx
+	xorl %eax,%eax		/*zero for the exception handler */
+
+#ifdef FIX_ALIGNMENT
+	/* check for bad alignment of destination */
+	movl %edi,%ecx
+	andl $7,%ecx
+	jnz  .Lbad_alignment
+.Lafter_bad_alignment:
+#endif
+
+	movq %rdx,%rcx
+
+	movl $64,%ebx
+	shrq $6,%rdx
+	decq %rdx
+	js   .Lhandle_tail
+
+	.p2align 4
+.Lloop:
+.Ls1:	movq (%rsi),%r11
+.Ls2:	movq 1*8(%rsi),%r8
+.Ls3:	movq 2*8(%rsi),%r9
+.Ls4:	movq 3*8(%rsi),%r10
+.Ld1:	movq %r11,(%rdi)
+.Ld2:	movq %r8,1*8(%rdi)
+.Ld3:	movq %r9,2*8(%rdi)
+.Ld4:	movq %r10,3*8(%rdi)
+
+.Ls5:	movq 4*8(%rsi),%r11
+.Ls6:	movq 5*8(%rsi),%r8
+.Ls7:	movq 6*8(%rsi),%r9
+.Ls8:	movq 7*8(%rsi),%r10
+.Ld5:	movq %r11,4*8(%rdi)
+.Ld6:	movq %r8,5*8(%rdi)
+.Ld7:	movq %r9,6*8(%rdi)
+.Ld8:	movq %r10,7*8(%rdi)
+
+	decq %rdx
+
+	leaq 64(%rsi),%rsi
+	leaq 64(%rdi),%rdi
+
+	jns  .Lloop
+
+	.p2align 4
+.Lhandle_tail:
+	movl %ecx,%edx
+	andl $63,%ecx
+	shrl $3,%ecx
+	jz   .Lhandle_7
+	movl $8,%ebx
+	.p2align 4
+.Lloop_8:
+.Ls9:	movq (%rsi),%r8
+.Ld9:	movq %r8,(%rdi)
+	decl %ecx
+	leaq 8(%rdi),%rdi
+	leaq 8(%rsi),%rsi
+	jnz .Lloop_8
+
+.Lhandle_7:
+	movl %edx,%ecx
+	andl $7,%ecx
+	jz   .Lende
+	.p2align 4
+.Lloop_1:
+.Ls10:	movb (%rsi),%bl
+.Ld10:	movb %bl,(%rdi)
+	incq %rdi
+	incq %rsi
+	decl %ecx
+	jnz .Lloop_1
+
+.Lende:
+	popq %rbx
+	ret
+
+#ifdef FIX_ALIGNMENT
+	/* align destination */
+	.p2align 4
+.Lbad_alignment:
+	movl $8,%r9d
+	subl %ecx,%r9d
+	movl %r9d,%ecx
+	cmpq %r9,%rdx
+	jz   .Lhandle_7
+	js   .Lhandle_7
+.Lalign_1:
+.Ls11:	movb (%rsi),%bl
+.Ld11:	movb %bl,(%rdi)
+	incq %rsi
+	incq %rdi
+	decl %ecx
+	jnz .Lalign_1
+	subq %r9,%rdx
+	jmp .Lafter_bad_alignment
+#endif
+
+	/* table sorted by exception address */
+	.section __ex_table,"a"
+	.align 8
+	.quad .Ls1,.Ls1e
+	.quad .Ls2,.Ls2e
+	.quad .Ls3,.Ls3e
+	.quad .Ls4,.Ls4e
+	.quad .Ld1,.Ls1e
+	.quad .Ld2,.Ls2e
+	.quad .Ld3,.Ls3e
+	.quad .Ld4,.Ls4e
+	.quad .Ls5,.Ls5e
+	.quad .Ls6,.Ls6e
+	.quad .Ls7,.Ls7e
+	.quad .Ls8,.Ls8e
+	.quad .Ld5,.Ls5e
+	.quad .Ld6,.Ls6e
+	.quad .Ld7,.Ls7e
+	.quad .Ld8,.Ls8e
+	.quad .Ls9,.Le_quad
+	.quad .Ld9,.Le_quad
+	.quad .Ls10,.Le_byte
+	.quad .Ld10,.Le_byte
+#ifdef FIX_ALIGNMENT
+	.quad .Ls11,.Lzero_rest
+	.quad .Ld11,.Lzero_rest
+#endif
+	.quad .Le5,.Le_zero
+	.previous
+
+	/* compute 64-offset for main loop. 8 bytes accuracy with error on the
+	   pessimistic side. this is gross. it would be better to fix the
+	   interface. */
+	/* eax: zero, ebx: 64 */
+.Ls1e: 	addl $8,%eax
+.Ls2e: 	addl $8,%eax
+.Ls3e: 	addl $8,%eax
+.Ls4e: 	addl $8,%eax
+.Ls5e: 	addl $8,%eax
+.Ls6e: 	addl $8,%eax
+.Ls7e: 	addl $8,%eax
+.Ls8e: 	addl $8,%eax
+	addq %rbx,%rdi	/* +64 */
+	subq %rax,%rdi  /* correct destination with computed offset */
+
+	shlq $6,%rdx	/* loop counter * 64 (stride length) */
+	addq %rax,%rdx	/* add offset to loopcnt */
+	andl $63,%ecx	/* remaining bytes */
+	addq %rcx,%rdx	/* add them */
+	jmp .Lzero_rest
+
+	/* exception on quad word loop in tail handling */
+	/* ecx:	loopcnt/8, %edx: length, rdi: correct */
+.Le_quad:
+	shll $3,%ecx
+	andl $7,%edx
+	addl %ecx,%edx
+	/* edx: bytes to zero, rdi: dest, eax:zero */
+.Lzero_rest:
+	movq %rdx,%rcx
+.Le_byte:
+	xorl %eax,%eax
+.Le5:	rep
+	stosb
+	/* when there is another exception while zeroing the rest just return */
+.Le_zero:
+	movq %rdx,%rax
+	jmp .Lende
+
+	/* Some CPUs run faster using the string copy instructions.
+	   This is also a lot simpler. Use them when possible.
+	   Patch in jmps to this code instead of copying it fully
+	   to avoid unwanted aliasing in the exception tables. */
+
+ /* rdi	destination
+  * rsi source
+  * rdx count
+  *
+  * Output:
+  * eax uncopied bytes or 0 if successfull.
+  *
+  * Only 4GB of copy is supported. This shouldn't be a problem
+  * because the kernel normally only writes from/to page sized chunks
+  * even if user space passed a longer buffer.
+  * And more would be dangerous because both Intel and AMD have
+  * errata with rep movsq > 4GB. If someone feels the need to fix
+  * this please consider this.
+   */
+copy_user_generic_c:
 	movl %edx,%ecx
 	shrl $3,%ecx
 	andl $7,%edx	
-	jz   5f
 1:	rep 
 	movsq 
 	movl %edx,%ecx
-	xor %eax,%eax
 2:	rep
 	movsb
+4:	movl %ecx,%eax
 	ret
-	/* align here? */
-5: 	xorl %eax,%eax
-6:	rep movsq
-	ret
-
-	.section .fixup,"ax"
 3:	lea (%rdx,%rcx,8),%rax
 	ret
-4:	movl %ecx,%eax
-	ret
-	.previous
 
 	.section __ex_table,"a"
 	.quad 1b,3b
 	.quad 2b,4b
-	.quad 6b,4b
 	.previous

arch/x86_64/lib/memcpy.S

@@ -11,8 +11,6 @@
  * 
  * Output:
  * rax original destination
- * 
- * TODO: check best memcpy for PSC
  */	
 
  	.globl __memcpy
@@ -20,6 +18,95 @@
 	.p2align 4
 __memcpy:
 memcpy:		
+	pushq %rbx
+	movq %rdi,%rax
+
+	movl %edx,%ecx
+	shrl $6,%ecx
+	jz .Lhandle_tail
+
+	.p2align 4
+.Lloop_64:
+	decl %ecx
+
+	movq (%rsi),%r11
+	movq 8(%rsi),%r8
+
+	movq %r11,(%rdi)
+	movq %r8,1*8(%rdi)
+
+	movq 2*8(%rsi),%r9
+	movq 3*8(%rsi),%r10
+
+	movq %r9,2*8(%rdi)
+	movq %r10,3*8(%rdi)
+
+	movq 4*8(%rsi),%r11
+	movq 5*8(%rsi),%r8
+
+	movq %r11,4*8(%rdi)
+	movq %r8,5*8(%rdi)
+
+	movq 6*8(%rsi),%r9
+	movq 7*8(%rsi),%r10
+
+	movq %r9,6*8(%rdi)
+	movq %r10,7*8(%rdi)
+
+	leaq 64(%rsi),%rsi
+	leaq 64(%rdi),%rdi
+	jnz  .Lloop_64
+
+.Lhandle_tail:
+	movl %edx,%ecx
+	andl $63,%ecx
+	shrl $3,%ecx
+	jz   .Lhandle_7
+	.p2align 4
+.Lloop_8:
+	decl %ecx
+	movq (%rsi),%r8
+	movq %r8,(%rdi)
+	leaq 8(%rdi),%rdi
+	leaq 8(%rsi),%rsi
+	jnz  .Lloop_8
+
+.Lhandle_7:
+	movl %edx,%ecx
+	andl $7,%ecx
+	jz .Lende
+	.p2align 4
+.Lloop_1:
+	movb (%rsi),%r8b
+	movb %r8b,(%rdi)
+	incq %rdi
+	incq %rsi
+	decl %ecx
+	jnz .Lloop_1
+
+.Lende:
+	popq %rbx
+	ret
+.Lfinal:
+
+	/* Some CPUs run faster using the string copy instructions.
+	   It is also a lot simpler. Use this when possible */
+
+	.section .altinstructions,"a"
+	.align 8
+	.quad  memcpy
+	.quad  memcpy_c
+	.byte  X86_FEATURE_REP_GOOD
+	.byte  .Lfinal-memcpy
+	.byte  memcpy_c_end-memcpy_c
+	.previous
+
+	.section .altinstr_replacement,"ax"
+ /* rdi	destination
+  * rsi source
+  * rdx count
+  */
+memcpy_c:
 	movq %rdi,%rax
 	movl %edx,%ecx
 	shrl $3,%ecx
@@ -30,3 +117,5 @@ memcpy:
 	rep
 	movsb
 	ret
+memcpy_c_end:
+	.previous

arch/x86_64/lib/memset.S

@@ -13,6 +13,98 @@
 	.p2align 4
 memset:	
 __memset:
+	movq %rdi,%r10
+	movq %rdx,%r11
+
+	/* expand byte value  */
+	movzbl %sil,%ecx
+	movabs $0x0101010101010101,%rax
+	mul    %rcx		/* with rax, clobbers rdx */
+
+	/* align dst */
+	movl  %edi,%r9d
+	andl  $7,%r9d
+	jnz  .Lbad_alignment
+.Lafter_bad_alignment:
+
+	movl %r11d,%ecx
+	shrl $6,%ecx
+	jz	 .Lhandle_tail
+
+	.p2align 4
+.Lloop_64:
+	decl   %ecx
+	movq  %rax,(%rdi)
+	movq  %rax,8(%rdi)
+	movq  %rax,16(%rdi)
+	movq  %rax,24(%rdi)
+	movq  %rax,32(%rdi)
+	movq  %rax,40(%rdi)
+	movq  %rax,48(%rdi)
+	movq  %rax,56(%rdi)
+	leaq  64(%rdi),%rdi
+	jnz    .Lloop_64
+
+	/* Handle tail in loops. The loops should be faster than hard
+	   to predict jump tables. */
+	.p2align 4
+.Lhandle_tail:
+	movl	%r11d,%ecx
+	andl    $63&(~7),%ecx
+	jz 		.Lhandle_7
+	shrl	$3,%ecx
+	.p2align 4
+.Lloop_8:
+	decl   %ecx
+	movq  %rax,(%rdi)
+	leaq  8(%rdi),%rdi
+	jnz    .Lloop_8
+
+.Lhandle_7:
+	movl	%r11d,%ecx
+	andl	$7,%ecx
+	jz      .Lende
+	.p2align 4
+.Lloop_1:
+	decl    %ecx
+	movb 	%al,(%rdi)
+	leaq	1(%rdi),%rdi
+	jnz     .Lloop_1
+
+.Lende:
+	movq	%r10,%rax
+	ret
+
+.Lbad_alignment:
+	cmpq $7,%r11
+	jbe	.Lhandle_7
+	movq %rax,(%rdi)	/* unaligned store */
+	movq $8,%r8
+	subq %r9,%r8
+	addq %r8,%rdi
+	subq %r8,%r11
+	jmp .Lafter_bad_alignment
+
+	/* Some CPUs run faster using the string instructions.
+	   It is also a lot simpler. Use this when possible */
+
+#include <asm/cpufeature.h>
+
+	.section .altinstructions,"a"
+	.align 8
+	.quad  memset
+	.quad  memset_c
+	.byte  X86_FEATURE_REP_GOOD
+	.byte  memset_c_end-memset_c
+	.byte  memset_c_end-memset_c
+	.previous
+
+	.section .altinstr_replacement,"ax"
+ /* rdi	destination
+  * rsi value
+  * rdx count
+  */
+memset_c:
 	movq %rdi,%r9
 	movl %edx,%r8d
 	andl $7,%r8d		
@@ -29,3 +121,5 @@ __memset:
 	stosb
 	movq %r9,%rax
 	ret
+memset_c_end:
+	.previous

include/asm-x86_64/cpufeature.h

@@ -61,7 +61,7 @@
 #define X86_FEATURE_K6_MTRR	(3*32+ 1) /* AMD K6 nonstandard MTRRs */
 #define X86_FEATURE_CYRIX_ARR	(3*32+ 2) /* Cyrix ARRs (= MTRRs) */
 #define X86_FEATURE_CENTAUR_MCR	(3*32+ 3) /* Centaur MCRs (= MTRRs) */
-/* 4 free */
+#define X86_FEATURE_REP_GOOD	(3*32+ 4) /* rep microcode works well on this CPU */
 #define X86_FEATURE_CONSTANT_TSC (3*32+5) /* TSC runs at constant rate */
 #define X86_FEATURE_SYNC_RDTSC  (3*32+6)  /* RDTSC syncs CPU core */