summaryrefslogtreecommitdiffstats
path: root/comm/third_party/libgcrypt/mpi/pentium4/sse2
diff options
context:
space:
mode:
Diffstat (limited to 'comm/third_party/libgcrypt/mpi/pentium4/sse2')
-rw-r--r--comm/third_party/libgcrypt/mpi/pentium4/sse2/distfiles5
-rw-r--r--comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-add1.S91
-rw-r--r--comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul1.S96
-rw-r--r--comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul2.S136
-rw-r--r--comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul3.S127
-rw-r--r--comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-sub1.S112
6 files changed, 567 insertions, 0 deletions
diff --git a/comm/third_party/libgcrypt/mpi/pentium4/sse2/distfiles b/comm/third_party/libgcrypt/mpi/pentium4/sse2/distfiles
new file mode 100644
index 0000000000..7252cd7e3f
--- /dev/null
+++ b/comm/third_party/libgcrypt/mpi/pentium4/sse2/distfiles
@@ -0,0 +1,5 @@
+mpih-add1.S
+mpih-mul1.S
+mpih-mul2.S
+mpih-mul3.S
+mpih-sub1.S
diff --git a/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-add1.S b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-add1.S
new file mode 100644
index 0000000000..55ed663032
--- /dev/null
+++ b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-add1.S
@@ -0,0 +1,91 @@
+/* Intel Pentium-4 mpn_add_n -- mpn addition.
+ *
+ * Copyright 2001, 2002 Free Software Foundation, Inc.
+ *
+ * This file is part of Libgcrypt.
+ *
+ * Libgcrypt is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as
+ * published by the Free Software Foundation; either version 2.1 of
+ * the License, or (at your option) any later version.
+ *
+ * Libgcrypt is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ * Actually it's the same code with only minor changes in the
+ * way the data is stored; this is to support the abstraction
+ * of an optional secure memory allocation which may be used
+ * to avoid revealing of sensitive data due to paging etc.
+ */
+
+
+#include "sysdep.h"
+#include "asm-syntax.h"
+
+
+ /*******************
+ * mpi_limb_t
+ * _gcry_mpih_add_n( mpi_ptr_t res_ptr, (sp + 4)
+ * mpi_ptr_t s1_ptr, (sp + 8)
+ * mpi_ptr_t s2_ptr, (sp + 12)
+ * mpi_size_t size) (sp + 16)
+ *
+ * P4 Willamette, Northwood: 4.0 cycles/limb if dst!=src1 and dst!=src2
+ * 6.0 cycles/limb if dst==src1 or dst==src2
+ * P4 Prescott: >= 5 cycles/limb
+ *
+ * The 4 c/l achieved here isn't particularly good, but is better than 9 c/l
+ * for a basic adc loop.
+ */
+
+ TEXT
+ ALIGN (3)
+ GLOBL C_SYMBOL_NAME(_gcry_mpih_add_n)
+C_SYMBOL_NAME(_gcry_mpih_add_n:)
+
+ pxor %mm0, %mm0
+
+ movl 8(%esp), %eax /* s1_ptr */
+ movl %ebx, 8(%esp) /* re-use parameter space */
+ movl 12(%esp), %ebx /* res_ptr */
+ movl 4(%esp), %edx /* s2_ptr */
+ movl 16(%esp), %ecx /* size */
+
+ leal (%eax,%ecx,4), %eax /* src1 end */
+ leal (%ebx,%ecx,4), %ebx /* src2 end */
+ leal (%edx,%ecx,4), %edx /* dst end */
+ negl %ecx /* -size */
+
+Ltop:
+/*
+ C eax src1 end
+ C ebx src2 end
+ C ecx counter, limbs, negative
+ C edx dst end
+ C mm0 carry bit
+*/
+
+ movd (%eax,%ecx,4), %mm1
+ movd (%ebx,%ecx,4), %mm2
+ paddq %mm2, %mm1
+
+ paddq %mm1, %mm0
+ movd %mm0, (%edx,%ecx,4)
+
+ psrlq $32, %mm0
+
+ addl $1, %ecx
+ jnz Ltop
+
+
+ movd %mm0, %eax
+ movl 8(%esp), %ebx /* restore saved EBX */
+ emms
+ ret
diff --git a/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul1.S b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul1.S
new file mode 100644
index 0000000000..a0c98fb4dd
--- /dev/null
+++ b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul1.S
@@ -0,0 +1,96 @@
+/* Intel Pentium-4 mpn_mul_1 -- Multiply a limb vector with a limb and store
+ * the result in a second limb vector.
+ *
+ * Copyright 2001, 2002, 2003, 2005 Free Software Foundation, Inc.
+ *
+ * This file is part of Libgcrypt.
+ *
+ * Libgcrypt is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as
+ * published by the Free Software Foundation; either version 2.1 of
+ * the License, or (at your option) any later version.
+ *
+ * Libgcrypt is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ * Actually it's the same code with only minor changes in the
+ * way the data is stored; this is to support the abstraction
+ * of an optional secure memory allocation which may be used
+ * to avoid revealing of sensitive data due to paging etc.
+ */
+
+
+#include "sysdep.h"
+#include "asm-syntax.h"
+
+
+/*******************
+ * mpi_limb_t
+ * _gcry_mpih_mul_1( mpi_ptr_t res_ptr, (sp + 4)
+ * mpi_ptr_t s1_ptr, (sp + 8)
+ * mpi_size_t s1_size, (sp + 12)
+ * mpi_limb_t s2_limb) (sp + 16)
+ *
+ * src != dst src == dst
+ * P6 model 9 (Banias) ?.?
+ * P6 model 13 (Dothan) 4.75 4.75
+ * P4 model 0 (Willamette) 4.0 6.0
+ * P4 model 1 (?) 4.0 6.0
+ * P4 model 2 (Northwood) 4.0 6.0
+ * P4 model 3 (Prescott) ?.? ?.?
+ * P4 model 4 (Nocona) ?.? ?.?
+ * Unfortunately when src==dst the write-combining described in
+ * pentium4/README takes us up to 6 c/l.
+ *
+ */
+
+ TEXT
+ ALIGN (3)
+ GLOBL C_SYMBOL_NAME(_gcry_mpih_mul_1)
+C_SYMBOL_NAME(_gcry_mpih_mul_1:);
+
+ pxor %mm0, %mm0
+
+.Lstart_1c:
+ movl 8(%esp), %eax
+ movd 16(%esp), %mm7
+ movl 4(%esp), %edx
+ movl 12(%esp), %ecx
+
+.Ltop:
+
+/*
+ C eax src, incrementing
+ C ebx
+ C ecx counter, size iterations
+ C edx dst, incrementing
+ C
+ C mm0 carry limb
+ C mm7 multiplier
+*/
+
+ movd (%eax), %mm1
+ addl $4, %eax
+ pmuludq %mm7, %mm1
+
+ paddq %mm1, %mm0
+ movd %mm0, (%edx)
+ addl $4, %edx
+
+ psrlq $32, %mm0
+
+ subl $1, %ecx
+ jnz .Ltop
+
+
+ movd %mm0, %eax
+ emms
+ ret
+
diff --git a/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul2.S b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul2.S
new file mode 100644
index 0000000000..f975adfca5
--- /dev/null
+++ b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul2.S
@@ -0,0 +1,136 @@
+/* Intel Pentium-4 mpn_addmul_1 -- Multiply a limb vector with a limb and add
+ * the result to a second limb vector.
+ *
+ * Copyright 2001, 2002, 2004, 2005 Free Software Foundation, Inc.
+ *
+ * This file is part of Libgcrypt.
+ *
+ * Libgcrypt is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as
+ * published by the Free Software Foundation; either version 2.1 of
+ * the License, or (at your option) any later version.
+ *
+ * Libgcrypt is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ * Actually it's the same code with only minor changes in the
+ * way the data is stored; this is to support the abstraction
+ * of an optional secure memory allocation which may be used
+ * to avoid revealing of sensitive data due to paging etc.
+ */
+
+
+#include "sysdep.h"
+#include "asm-syntax.h"
+
+
+/*******************
+ * mpi_limb_t
+ * _gcry_mpih_addmul_1( mpi_ptr_t res_ptr, (sp + 4)
+ * mpi_ptr_t s1_ptr, (sp + 8)
+ * mpi_size_t s1_size, (sp + 12)
+ * mpi_limb_t s2_limb) (sp + 16)
+ *
+ * P3 model 9 (Banias) ?.?
+ * P3 model 13 (Dothan) 5.8
+ * P4 model 0 (Willamette) 5.5
+ * P4 model 1 (?) 5.5
+ * P4 model 2 (Northwood) 5.5
+ * P4 model 3 (Prescott) 6.0
+ * P4 model 4 (Nocona)
+ *
+ * Only the carry limb propagation is on the dependent chain, but some other
+ * Pentium4 pipeline magic brings down performance to 6 cycles/l from the
+ * ideal 4 cycles/l.
+ */
+
+
+ TEXT
+ ALIGN (4)
+ GLOBL C_SYMBOL_NAME(_gcry_mpih_addmul_1)
+C_SYMBOL_NAME(_gcry_mpih_addmul_1:)
+
+ pxor %mm4, %mm4
+.Lstart_1c:
+ movl 8(%esp), %eax
+ movl 12(%esp), %ecx
+ movl 4(%esp), %edx
+ movd 16(%esp), %mm7
+
+/*
+ C eax src, incrementing ; 5B
+ C ecx loop counter, decrementing
+ C edx dst, incrementing
+ C
+ C mm4 carry, low 32-bits
+ C mm7 multiplier
+*/
+
+ movd (%eax), %mm2
+ pmuludq %mm7, %mm2
+
+ shrl $1, %ecx
+ jnc .Leven
+
+ leal 4(%eax), %eax
+ movd (%edx), %mm1
+ paddq %mm2, %mm1
+ paddq %mm1, %mm4
+ movd %mm4, (%edx)
+ psrlq $32, %mm4
+
+ testl %ecx, %ecx
+ jz .Lrtn
+ leal 4(%edx), %edx
+
+ movd (%eax), %mm2
+ pmuludq %mm7, %mm2
+.Leven:
+ movd 4(%eax), %mm0
+ movd (%edx), %mm1
+ pmuludq %mm7, %mm0
+
+ subl $1, %ecx
+ jz .Lend
+.Lloop:
+ paddq %mm2, %mm1
+ movd 8(%eax), %mm2
+ paddq %mm1, %mm4
+ movd 4(%edx), %mm3
+ pmuludq %mm7, %mm2
+ movd %mm4, (%edx)
+ psrlq $32, %mm4
+
+ paddq %mm0, %mm3
+ movd 12(%eax), %mm0
+ paddq %mm3, %mm4
+ movd 8(%edx), %mm1
+ pmuludq %mm7, %mm0
+ movd %mm4, 4(%edx)
+ psrlq $32, %mm4
+
+ leal 8(%eax), %eax
+ leal 8(%edx), %edx
+ subl $1, %ecx
+ jnz .Lloop
+.Lend:
+ paddq %mm2, %mm1
+ paddq %mm1, %mm4
+ movd 4(%edx), %mm3
+ movd %mm4, (%edx)
+ psrlq $32, %mm4
+ paddq %mm0, %mm3
+ paddq %mm3, %mm4
+ movd %mm4, 4(%edx)
+ psrlq $32, %mm4
+.Lrtn:
+ movd %mm4, %eax
+ emms
+ ret
diff --git a/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul3.S b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul3.S
new file mode 100644
index 0000000000..ebcd2a68ea
--- /dev/null
+++ b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-mul3.S
@@ -0,0 +1,127 @@
+/* Intel Pentium-4 mpn_submul_1 -- Multiply a limb vector with a limb and
+ * subtract the result from a second limb vector.
+ *
+ * Copyright 2001, 2002 Free Software Foundation, Inc.
+ *
+ * This file is part of Libgcrypt.
+ *
+ * Libgcrypt is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as
+ * published by the Free Software Foundation; either version 2.1 of
+ * the License, or (at your option) any later version.
+ *
+ * Libgcrypt is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ * Actually it's the same code with only minor changes in the
+ * way the data is stored; this is to support the abstraction
+ * of an optional secure memory allocation which may be used
+ * to avoid revealing of sensitive data due to paging etc.
+ */
+
+
+#include "sysdep.h"
+#include "asm-syntax.h"
+
+
+/*******************
+ * mpi_limb_t
+ * _gcry_mpih_submul_1( mpi_ptr_t res_ptr, (sp + 4)
+ * mpi_ptr_t s1_ptr, (sp + 8)
+ * mpi_size_t s1_size, (sp + 12)
+ * mpi_limb_t s2_limb) (sp + 16)
+ *
+ * P4: 7 cycles/limb, unstable timing, at least on early Pentium4 silicon
+ * (stepping 10).
+ *
+ * This code is not particularly good at 7 c/l. The dependent chain is only
+ * 4 c/l and there's only 4 MMX unit instructions, so it's not clear why that
+ * speed isn't achieved.
+ *
+ * The arrangements made here to get a two instruction dependent chain are
+ * slightly subtle. In the loop the carry (or borrow rather) is a negative
+ * so that a paddq can be used to give a low limb ready to store, and a high
+ * limb ready to become the new carry after a psrlq.
+ *
+ * If the carry was a simple twos complement negative then the psrlq shift
+ * would need to bring in 0 bits or 1 bits according to whether the high was
+ * zero or non-zero, since a non-zero value would represent a negative
+ * needing sign extension. That wouldn't be particularly easy to arrange and
+ * certainly would add an instruction to the dependent chain, so instead an
+ * offset is applied so that the high limb will be 0xFFFFFFFF+c. With c in
+ * the range -0xFFFFFFFF to 0, the value 0xFFFFFFFF+c is in the range 0 to
+ * 0xFFFFFFFF and is therefore always positive and can always have 0 bits
+ * shifted in, which is what psrlq does.
+ *
+ * The extra 0xFFFFFFFF must be subtracted before c is used, but that can be
+ * done off the dependent chain. The total adjustment then is to add
+ * 0xFFFFFFFF00000000 to offset the new carry, and subtract
+ * 0x00000000FFFFFFFF to remove the offset from the current carry, for a net
+ * add of 0xFFFFFFFE00000001. In the code this is applied to the destination
+ * limb when fetched.
+ *
+ * It's also possible to view the 0xFFFFFFFF adjustment as a ones-complement
+ * negative, which is how it's undone for the return value, but that doesn't
+ * seem as clear.
+*/
+
+ TEXT
+ ALIGN (4)
+ GLOBL C_SYMBOL_NAME(_gcry_mpih_submul_1)
+C_SYMBOL_NAME(_gcry_mpih_submul_1:)
+
+ pxor %mm1, %mm1
+
+.Lstart_1c:
+ movl 8(%esp), %eax
+ pcmpeqd %mm0, %mm0
+
+ movd 16(%esp), %mm7
+ pcmpeqd %mm6, %mm6
+
+ movl 4(%esp), %edx
+ psrlq $32, %mm0
+
+ movl 12(%esp), %ecx
+ psllq $32, %mm6
+
+ psubq %mm0, %mm6
+
+ psubq %mm1, %mm0
+
+/*
+ C eax src, incrementing
+ C ebx
+ C ecx loop counter, decrementing
+ C edx dst, incrementing
+ C
+ C mm0 0xFFFFFFFF - borrow
+ C mm6 0xFFFFFFFE00000001
+ C mm7 multiplier
+*/
+
+.Lloop:
+ movd (%eax), %mm1
+ leal 4(%eax), %eax
+ movd (%edx), %mm2
+ paddq %mm6, %mm2
+ pmuludq %mm7, %mm1
+ psubq %mm1, %mm2
+ paddq %mm2, %mm0
+ subl $1, %ecx
+ movd %mm0, (%edx)
+ psrlq $32, %mm0
+ leal 4(%edx), %edx
+ jnz .Lloop
+
+ movd %mm0, %eax
+ notl %eax
+ emms
+ ret
diff --git a/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-sub1.S b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-sub1.S
new file mode 100644
index 0000000000..33900c742e
--- /dev/null
+++ b/comm/third_party/libgcrypt/mpi/pentium4/sse2/mpih-sub1.S
@@ -0,0 +1,112 @@
+/* Intel Pentium-4 mpn_sub_n -- mpn subtraction.
+ *
+ * Copyright 2001, 2002 Free Software Foundation, Inc.
+ *
+ * This file is part of Libgcrypt.
+ *
+ * Libgcrypt is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as
+ * published by the Free Software Foundation; either version 2.1 of
+ * the License, or (at your option) any later version.
+ *
+ * Libgcrypt is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ * Actually it's the same code with only minor changes in the
+ * way the data is stored; this is to support the abstraction
+ * of an optional secure memory allocation which may be used
+ * to avoid revealing of sensitive data due to paging etc.
+ */
+
+
+#include "sysdep.h"
+#include "asm-syntax.h"
+
+
+/*******************
+ * mpi_limb_t
+ * _gcry_mpih_sub_n( mpi_ptr_t res_ptr, (sp + 4)
+ * mpi_ptr_t s1_ptr, (sp + 8)
+ * mpi_ptr_t s2_ptr, (sp + 12)
+ * mpi_size_t size) (sp + 16)
+ *
+ * P4 Willamette, Northwood: 4.0 cycles/limb if dst!=src1 and dst!=src2
+ * 6.0 cycles/limb if dst==src1 or dst==src2
+ * P4 Prescott: >= 5 cycles/limb
+ *
+ * The main loop code is 2x unrolled so that the carry bit can alternate
+ * between mm0 and mm1.
+ */
+
+
+.text
+ ALIGN (3)
+ .globl C_SYMBOL_NAME(_gcry_mpih_sub_n)
+C_SYMBOL_NAME(_gcry_mpih_sub_n:)
+
+ pxor %mm0, %mm0
+.Lstart_nc:
+ movl 8(%esp), %eax
+ movl %ebx, 8(%esp)
+ movl 12(%esp), %ebx
+ movl 4(%esp), %edx
+ movl 16(%esp), %ecx
+
+ leal (%eax,%ecx,4), %eax
+ leal (%ebx,%ecx,4), %ebx
+ leal (%edx,%ecx,4), %edx
+ negl %ecx
+
+.Ltop:
+/*
+ C eax src1 end
+ C ebx src2 end
+ C ecx counter, limbs, negative
+ C edx dst end
+ C mm0 carry bit
+*/
+
+ movd (%eax,%ecx,4), %mm1
+ movd (%ebx,%ecx,4), %mm2
+ psubq %mm2, %mm1
+
+ psubq %mm0, %mm1
+ movd %mm1, (%edx,%ecx,4)
+
+ psrlq $63, %mm1
+
+ addl $1, %ecx
+ jz .Ldone_mm1
+
+ movd (%eax,%ecx,4), %mm0
+ movd (%ebx,%ecx,4), %mm2
+ psubq %mm2, %mm0
+
+ psubq %mm1, %mm0
+ movd %mm0, (%edx,%ecx,4)
+
+ psrlq $63, %mm0
+
+ addl $1, %ecx
+ jnz .Ltop
+
+
+ movd %mm0, %eax
+ movl 8(%esp), %ebx
+ emms
+ ret
+
+
+
+.Ldone_mm1:
+ movd %mm1, %eax
+ movl 8(%esp), %ebx
+ emms
+ ret