[Gmp-commit] /home/hgfiles/gmp: Tune.
mercurial at gmplib.org
mercurial at gmplib.org
Wed Jun 30 23:06:00 CEST 2010
details: /home/hgfiles/gmp/rev/5d980cdf49eb
changeset: 13660:5d980cdf49eb
user: Torbjorn Granlund <tege at gmplib.org>
date: Wed Jun 30 23:05:53 2010 +0200
description:
Tune.
diffstat:
ChangeLog | 4 +
mpn/x86_64/divrem_2.asm | 222 ++++++++++++++---------------------------------
2 files changed, 73 insertions(+), 153 deletions(-)
diffs (truncated from 311 to 300 lines):
diff -r 40b15fdbf204 -r 5d980cdf49eb ChangeLog
--- a/ChangeLog Sat Jun 19 16:16:52 2010 +0200
+++ b/ChangeLog Wed Jun 30 23:05:53 2010 +0200
@@ -1,3 +1,7 @@
+2010-06-30 Torbjorn Granlund <tege at gmplib.org>
+
+ * mpn/x86_64/divrem_2.asm: Tune.
+
2010-06-19 Niels Möller <nisse at lysator.liu.se>
* tune/speed.h (SPEED_ROUTINE_MPN_MOD_1_1): Pass normalized
diff -r 40b15fdbf204 -r 5d980cdf49eb mpn/x86_64/divrem_2.asm
--- a/mpn/x86_64/divrem_2.asm Sat Jun 19 16:16:52 2010 +0200
+++ b/mpn/x86_64/divrem_2.asm Wed Jun 30 23:05:53 2010 +0200
@@ -1,6 +1,6 @@
dnl x86-64 mpn_divrem_2 -- Divide an mpn number by a normalized 2-limb number.
-dnl Copyright 2007, 2008 Free Software Foundation, Inc.
+dnl Copyright 2007, 2008, 2010 Free Software Foundation, Inc.
dnl This file is part of the GNU MP Library.
@@ -20,22 +20,15 @@
include(`../config.m4')
-C NEW
-C norm frac norm
-C AMD K8,K9 21 21 19
-C AMD K10 21 21 19
-C Intel P4 73 73 70
-C Intel core2 37.5 37.5 36
-C Intel corei 33 33 32
-C Intel atom ? ? ?
-C VIA nano 37 ? 37
+C c/l
+C AMD K8,K9 18
+C AMD K10 18
+C Intel P4 68
+C Intel core2 34
+C Intel corei 30.5
+C Intel atom ?
+C VIA nano 33
-C TODO
-C * The loop has not been carefully tuned, nor analysed for critical path
-C length. It seems that 19 c/l is a bit long, compared to the 13 c/l for
-C mpn_divrem_1.
-C * Clean up. This code is really crude.
-C * Add testing to tests/devel/try.c before enabling NEW code.
C INPUT PARAMETERS
define(`qp', `%rdi')
@@ -44,43 +37,40 @@
define(`un_param', `%rcx')
define(`dp', `%r8')
-define(`dinv', `%r9')
-
-
-C rax rbx rcx rdx rsi rdi rbp r8 r9 r10 r11 r12 r13 r14 r15
-C cnt qp d dinv
ASM_START()
TEXT
ALIGN(16)
PROLOGUE(mpn_divrem_2)
-
push %r15
- lea (%rdx,%rcx,8), %rax
push %r14
push %r13
+ push %r12
+ lea -24(%rdx,%rcx,8), %r12 C r12 = &up[un-1]
mov %rsi, %r13
- push %r12
- lea -24(%rax), %r12
push %rbp
mov %rdi, %rbp
push %rbx
- mov 8(%r8), %r11
- mov -8(%rax), %r9
- mov (%r8), %r8
- mov -16(%rax), %r10
+ mov 8(%r8), %r11 C d1
+ mov 16(%r12), %rbx
+ mov (%r8), %r8 C d0
+ mov 8(%r12), %r10
+
xor R32(%r15), R32(%r15)
- cmp %r9, %r11
+ cmp %rbx, %r11
ja L(2)
setb %dl
cmp %r10, %r8
setbe %al
orb %al, %dl C "orb" form to placate Sun tools
- jne L(23)
+ je L(2)
+ inc R32(%r15)
+ sub %r8, %r10
+ sbb %r11, %rbx
L(2):
- lea -3(%rcx,%r13), %rbx C un + fn - 3
- test %rbx, %rbx
- js L(6)
+ lea -3(%rcx,%r13), %r14 C un + fn - 3
+ test %r14, %r14
+ js L(end)
push %r8
push %r10
@@ -94,124 +84,65 @@
mov %r11, %rdx
mov %rax, %rdi
imul %rax, %rdx
- mov %rdx, %r14
+ mov %rdx, %r9
mul %r8
- mov %rdx, %rcx
- mov $-1, %rdx
- add %r8, %r14
- adc $0, %rdx
- add %rcx, %r14
- adc $0, %rdx
- js L(8)
-L(18):
- dec %rdi
- sub %r11, %r14
- sbb $0, %rdx
- jns L(18)
-L(8):
+ xor R32(%rcx), R32(%rcx)
+ add %r8, %r9
+ adc $-1, %rcx
+ add %rdx, %r9
+ adc $0, %rcx
+ js 2f
+1: dec %rdi
+ sub %r11, %r9
+ sbb $0, %rcx
+ jns 1b
+2:
-C rax rbx rcx rdx rsi rdi rbp r8 r9 r10 r11 r12 r13 r14 r15
-C n2 un n1 dinv qp d0 d1 up fn msl
-C n2 un -d1 n1 dinv XX XX
-
-ifdef(`NEW',`
- lea (%rbp,%rbx,8), %rbp
- mov %rbx, %rcx C un
- mov %r9, %rbx
- mov %rdi, %r9 C di
- mov %r10, %r14
+ lea (%rbp,%r14,8), %rbp
mov %r11, %rsi
neg %rsi C -d1
+
+C rax rbx rcx rdx rsi rdi rbp r8 r9 r10 r11 r12 r13 r14 r15
+C n2 un -d1 dinv qp d0 q0 d1 up fn msl
+
ALIGN(16)
-L(loop):
- mov %r9, %rax C di ncp
+L(top): mov %rdi, %rax C di ncp
mul %rbx C 0, 17
- add %r14, %rax C 4
- mov %rax, %r10 C q0 5
+ mov %r10, %rcx C
+ add %rax, %rcx C 4
adc %rbx, %rdx C 5
- mov %rdx, %rdi C q 6
+ mov %rdx, %r9 C q 6
imul %rsi, %rdx C 6
mov %r8, %rax C ncp
- lea (%rdx, %r14), %rbx C n1 -= ... 7
- xor R32(%r14), R32(%r14) C
- mul %rdi C 7
- cmp %rcx, %r13 C
+ lea (%rdx, %r10), %rbx C n1 -= ... 10
+ xor R32(%r10), R32(%r10) C
+ mul %r9 C 7
+ cmp %r14, %r13 C
jg L(19) C
- mov (%r12), %r14 C
+ mov (%r12), %r10 C
sub $8, %r12 C
-L(19): sub %r8, %r14 C ncp
- sbb %r11, %rbx C 9
- sub %rax, %r14 C 11
+L(19): sub %r8, %r10 C ncp
+ sbb %r11, %rbx C 11
+ sub %rax, %r10 C 11
sbb %rdx, %rbx C 12
- inc %rdi C 7
+ xor R32(%rax), R32(%rax) C
xor R32(%rdx), R32(%rdx) C
- cmp %r10, %rbx C 13
- mov %r8, %rax C d0 ncp
- adc $-1, %rdx C mask 14
- add %rdx, %rdi C q-- 15
- and %rdx, %rax C d0 or 0 15
- and %r11, %rdx C d1 or 0 15
- add %rax, %r14 C 16
+ cmp %rcx, %rbx C 13
+ cmovnc %r8, %rax C 14
+ cmovnc %r11, %rdx C 14
+ adc $0, %r9 C adjust q 14
+ nop
+ add %rax, %r10 C 15
adc %rdx, %rbx C 16
- cmp %r11, %rbx C 17
+ cmp %r11, %rbx C
jae L(fix) C
-L(bck): mov %rdi, (%rbp) C
+L(bck): mov %r9, (%rbp) C
sub $8, %rbp C
- dec %rcx
- jns L(loop)
+ dec %r14
+ jns L(top)
- mov %r14, %r10
- mov %rbx, %r9
-',`
- lea (%rbp,%rbx,8), %rbp
- mov %rbx, %rcx
- mov %r9, %rax
- mov %r10, %rsi
- ALIGN(16)
-L(loop):
- mov %rax, %r14 C 0, 19
- mul %rdi C 0
- mov %r11, %r9 C 1
- add %rsi, %rax C 4
- mov %rax, %rbx C q0 5
- adc %r14, %rdx C q 5
- lea 1(%rdx), %r10 C 6
- mov %rdx, %rax C 6
- imul %rdx, %r9 C 6
- sub %r9, %rsi C 10
- xor R32(%r9), R32(%r9) C
- mul %r8 C 7
- cmp %rcx, %r13 C
- jg L(13) C
- mov (%r12), %r9 C
- sub $8, %r12 C
-L(13): sub %r8, %r9 C ncp
- sbb %r11, %rsi C 11
- sub %rax, %r9 C 11
- sbb %rdx, %rsi C 12
- cmp %rbx, %rsi C 13
- sbb %rax, %rax C 14
- not %rax C 15
- add %rax, %r10 C 16
- mov %r8, %rbx C ncp
- and %rax, %rbx C 16
- and %r11, %rax C 16
- add %rbx, %r9 C 17
- adc %rsi, %rax C 18
- cmp %rax, %r11 C 19
- jbe L(fix) C
-L(bck): mov %r10, (%rbp) C
- sub $8, %rbp C
- mov %r9, %rsi C 18
- dec %rcx
- jns L(loop)
-
- mov %rsi, %r10
- mov %rax, %r9
-')
-L(6):
- mov %r10, 8(%r12)
- mov %r9, 16(%r12)
+L(end): mov %r10, 8(%r12)
+ mov %rbx, 16(%r12)
pop %rbx
pop %rbp
pop %r12
@@ -221,28 +152,13 @@
pop %r15
ret
-L(23): inc R32(%r15)
- sub %r8, %r10
- sbb %r11, %r9
- jmp L(2)
-
-ifdef(`NEW',`
L(fix): seta %dl
- cmp %r8, %r14
+ cmp %r8, %r10
setae %al
orb %dl, %al C "orb" form to placate Sun tools
je L(bck)
- inc %rdi
- sub %r8, %r14
+ inc %r9
+ sub %r8, %r10
sbb %r11, %rbx
More information about the gmp-commit
mailing list