2005-09-26 14:04:21 +08:00
|
|
|
/*
|
|
|
|
* Single-step support.
|
|
|
|
*
|
|
|
|
* Copyright (C) 2004 Paul Mackerras <paulus@au.ibm.com>, IBM
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public License
|
|
|
|
* as published by the Free Software Foundation; either version
|
|
|
|
* 2 of the License, or (at your option) any later version.
|
|
|
|
*/
|
|
|
|
#include <linux/kernel.h>
|
2006-11-01 10:50:15 +08:00
|
|
|
#include <linux/kprobes.h>
|
2005-09-26 14:04:21 +08:00
|
|
|
#include <linux/ptrace.h>
|
2011-05-21 03:50:29 +08:00
|
|
|
#include <linux/prefetch.h>
|
2005-09-26 14:04:21 +08:00
|
|
|
#include <asm/sstep.h>
|
|
|
|
#include <asm/processor.h>
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
#include <asm/uaccess.h>
|
|
|
|
#include <asm/cputable.h>
|
2005-09-26 14:04:21 +08:00
|
|
|
|
|
|
|
extern char system_call_common[];
|
|
|
|
|
2005-10-28 20:48:08 +08:00
|
|
|
#ifdef CONFIG_PPC64
|
2005-09-26 14:04:21 +08:00
|
|
|
/* Bits in SRR1 that are copied from MSR */
|
2006-03-23 14:38:10 +08:00
|
|
|
#define MSR_MASK 0xffffffff87c0ffffUL
|
2005-10-28 20:48:08 +08:00
|
|
|
#else
|
|
|
|
#define MSR_MASK 0x87c0ffff
|
|
|
|
#endif
|
2005-09-26 14:04:21 +08:00
|
|
|
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
/* Bits in XER */
|
|
|
|
#define XER_SO 0x80000000U
|
|
|
|
#define XER_OV 0x40000000U
|
|
|
|
#define XER_CA 0x20000000U
|
|
|
|
|
2010-09-01 15:21:21 +08:00
|
|
|
#ifdef CONFIG_PPC_FPU
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
/*
|
|
|
|
* Functions in ldstfp.S
|
|
|
|
*/
|
|
|
|
extern int do_lfs(int rn, unsigned long ea);
|
|
|
|
extern int do_lfd(int rn, unsigned long ea);
|
|
|
|
extern int do_stfs(int rn, unsigned long ea);
|
|
|
|
extern int do_stfd(int rn, unsigned long ea);
|
|
|
|
extern int do_lvx(int rn, unsigned long ea);
|
|
|
|
extern int do_stvx(int rn, unsigned long ea);
|
|
|
|
extern int do_lxvd2x(int rn, unsigned long ea);
|
|
|
|
extern int do_stxvd2x(int rn, unsigned long ea);
|
2010-09-01 15:21:21 +08:00
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
2011-04-08 05:56:04 +08:00
|
|
|
/*
|
|
|
|
* Emulate the truncation of 64 bit values in 32-bit mode.
|
|
|
|
*/
|
|
|
|
static unsigned long truncate_if_32bit(unsigned long msr, unsigned long val)
|
|
|
|
{
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
if ((msr & MSR_64BIT) == 0)
|
|
|
|
val &= 0xffffffffUL;
|
|
|
|
#endif
|
|
|
|
return val;
|
|
|
|
}
|
|
|
|
|
2005-09-26 14:04:21 +08:00
|
|
|
/*
|
|
|
|
* Determine whether a conditional branch instruction would branch.
|
|
|
|
*/
|
2006-11-01 10:50:15 +08:00
|
|
|
static int __kprobes branch_taken(unsigned int instr, struct pt_regs *regs)
|
2005-09-26 14:04:21 +08:00
|
|
|
{
|
|
|
|
unsigned int bo = (instr >> 21) & 0x1f;
|
|
|
|
unsigned int bi;
|
|
|
|
|
|
|
|
if ((bo & 4) == 0) {
|
|
|
|
/* decrement counter */
|
|
|
|
--regs->ctr;
|
|
|
|
if (((bo >> 1) & 1) ^ (regs->ctr == 0))
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
if ((bo & 0x10) == 0) {
|
|
|
|
/* check bit from CR */
|
|
|
|
bi = (instr >> 16) & 0x1f;
|
|
|
|
if (((regs->ccr >> (31 - bi)) & 1) != ((bo >> 3) & 1))
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
static long __kprobes address_ok(struct pt_regs *regs, unsigned long ea, int nb)
|
|
|
|
{
|
|
|
|
if (!user_mode(regs))
|
|
|
|
return 1;
|
|
|
|
return __access_ok(ea, nb, USER_DS);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Calculate effective address for a D-form instruction
|
|
|
|
*/
|
|
|
|
static unsigned long __kprobes dform_ea(unsigned int instr, struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
int ra;
|
|
|
|
unsigned long ea;
|
|
|
|
|
|
|
|
ra = (instr >> 16) & 0x1f;
|
|
|
|
ea = (signed short) instr; /* sign-extend */
|
2014-09-02 12:35:07 +08:00
|
|
|
if (ra)
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
ea += regs->gpr[ra];
|
2011-04-08 05:56:04 +08:00
|
|
|
|
|
|
|
return truncate_if_32bit(regs->msr, ea);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
/*
|
|
|
|
* Calculate effective address for a DS-form instruction
|
|
|
|
*/
|
|
|
|
static unsigned long __kprobes dsform_ea(unsigned int instr, struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
int ra;
|
|
|
|
unsigned long ea;
|
|
|
|
|
|
|
|
ra = (instr >> 16) & 0x1f;
|
|
|
|
ea = (signed short) (instr & ~3); /* sign-extend */
|
2014-09-02 12:35:07 +08:00
|
|
|
if (ra)
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
ea += regs->gpr[ra];
|
2011-04-08 05:56:04 +08:00
|
|
|
|
|
|
|
return truncate_if_32bit(regs->msr, ea);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
}
|
|
|
|
#endif /* __powerpc64 */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Calculate effective address for an X-form instruction
|
|
|
|
*/
|
2014-09-02 12:35:07 +08:00
|
|
|
static unsigned long __kprobes xform_ea(unsigned int instr,
|
|
|
|
struct pt_regs *regs)
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
{
|
|
|
|
int ra, rb;
|
|
|
|
unsigned long ea;
|
|
|
|
|
|
|
|
ra = (instr >> 16) & 0x1f;
|
|
|
|
rb = (instr >> 11) & 0x1f;
|
|
|
|
ea = regs->gpr[rb];
|
2014-09-02 12:35:07 +08:00
|
|
|
if (ra)
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
ea += regs->gpr[ra];
|
2011-04-08 05:56:04 +08:00
|
|
|
|
|
|
|
return truncate_if_32bit(regs->msr, ea);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Return the largest power of 2, not greater than sizeof(unsigned long),
|
|
|
|
* such that x is a multiple of it.
|
|
|
|
*/
|
|
|
|
static inline unsigned long max_align(unsigned long x)
|
|
|
|
{
|
|
|
|
x |= sizeof(unsigned long);
|
|
|
|
return x & -x; /* isolates rightmost bit */
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static inline unsigned long byterev_2(unsigned long x)
|
|
|
|
{
|
|
|
|
return ((x >> 8) & 0xff) | ((x & 0xff) << 8);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline unsigned long byterev_4(unsigned long x)
|
|
|
|
{
|
|
|
|
return ((x >> 24) & 0xff) | ((x >> 8) & 0xff00) |
|
|
|
|
((x & 0xff00) << 8) | ((x & 0xff) << 24);
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
static inline unsigned long byterev_8(unsigned long x)
|
|
|
|
{
|
|
|
|
return (byterev_4(x) << 32) | byterev_4(x >> 32);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
static int __kprobes read_mem_aligned(unsigned long *dest, unsigned long ea,
|
|
|
|
int nb)
|
|
|
|
{
|
|
|
|
int err = 0;
|
|
|
|
unsigned long x = 0;
|
|
|
|
|
|
|
|
switch (nb) {
|
|
|
|
case 1:
|
|
|
|
err = __get_user(x, (unsigned char __user *) ea);
|
|
|
|
break;
|
|
|
|
case 2:
|
|
|
|
err = __get_user(x, (unsigned short __user *) ea);
|
|
|
|
break;
|
|
|
|
case 4:
|
|
|
|
err = __get_user(x, (unsigned int __user *) ea);
|
|
|
|
break;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 8:
|
|
|
|
err = __get_user(x, (unsigned long __user *) ea);
|
|
|
|
break;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
if (!err)
|
|
|
|
*dest = x;
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int __kprobes read_mem_unaligned(unsigned long *dest, unsigned long ea,
|
|
|
|
int nb, struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
unsigned long x, b, c;
|
2013-10-19 03:42:08 +08:00
|
|
|
#ifdef __LITTLE_ENDIAN__
|
|
|
|
int len = nb; /* save a copy of the length for byte reversal */
|
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
/* unaligned, do this in pieces */
|
|
|
|
x = 0;
|
|
|
|
for (; nb > 0; nb -= c) {
|
2013-10-19 03:42:08 +08:00
|
|
|
#ifdef __LITTLE_ENDIAN__
|
|
|
|
c = 1;
|
|
|
|
#endif
|
|
|
|
#ifdef __BIG_ENDIAN__
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
c = max_align(ea);
|
2013-10-19 03:42:08 +08:00
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
if (c > nb)
|
|
|
|
c = max_align(nb);
|
|
|
|
err = read_mem_aligned(&b, ea, c);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
x = (x << (8 * c)) + b;
|
|
|
|
ea += c;
|
|
|
|
}
|
2013-10-19 03:42:08 +08:00
|
|
|
#ifdef __LITTLE_ENDIAN__
|
|
|
|
switch (len) {
|
|
|
|
case 2:
|
|
|
|
*dest = byterev_2(x);
|
|
|
|
break;
|
|
|
|
case 4:
|
|
|
|
*dest = byterev_4(x);
|
|
|
|
break;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 8:
|
|
|
|
*dest = byterev_8(x);
|
|
|
|
break;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
#ifdef __BIG_ENDIAN__
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
*dest = x;
|
2013-10-19 03:42:08 +08:00
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Read memory at address ea for nb bytes, return 0 for success
|
|
|
|
* or -EFAULT if an error occurred.
|
|
|
|
*/
|
|
|
|
static int __kprobes read_mem(unsigned long *dest, unsigned long ea, int nb,
|
|
|
|
struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
if (!address_ok(regs, ea, nb))
|
|
|
|
return -EFAULT;
|
|
|
|
if ((ea & (nb - 1)) == 0)
|
|
|
|
return read_mem_aligned(dest, ea, nb);
|
|
|
|
return read_mem_unaligned(dest, ea, nb, regs);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int __kprobes write_mem_aligned(unsigned long val, unsigned long ea,
|
|
|
|
int nb)
|
|
|
|
{
|
|
|
|
int err = 0;
|
|
|
|
|
|
|
|
switch (nb) {
|
|
|
|
case 1:
|
|
|
|
err = __put_user(val, (unsigned char __user *) ea);
|
|
|
|
break;
|
|
|
|
case 2:
|
|
|
|
err = __put_user(val, (unsigned short __user *) ea);
|
|
|
|
break;
|
|
|
|
case 4:
|
|
|
|
err = __put_user(val, (unsigned int __user *) ea);
|
|
|
|
break;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 8:
|
|
|
|
err = __put_user(val, (unsigned long __user *) ea);
|
|
|
|
break;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int __kprobes write_mem_unaligned(unsigned long val, unsigned long ea,
|
|
|
|
int nb, struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
unsigned long c;
|
|
|
|
|
2013-10-19 03:42:08 +08:00
|
|
|
#ifdef __LITTLE_ENDIAN__
|
|
|
|
switch (nb) {
|
|
|
|
case 2:
|
|
|
|
val = byterev_2(val);
|
|
|
|
break;
|
|
|
|
case 4:
|
|
|
|
val = byterev_4(val);
|
|
|
|
break;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 8:
|
|
|
|
val = byterev_8(val);
|
|
|
|
break;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
/* unaligned or little-endian, do this in pieces */
|
|
|
|
for (; nb > 0; nb -= c) {
|
2013-10-19 03:42:08 +08:00
|
|
|
#ifdef __LITTLE_ENDIAN__
|
|
|
|
c = 1;
|
|
|
|
#endif
|
|
|
|
#ifdef __BIG_ENDIAN__
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
c = max_align(ea);
|
2013-10-19 03:42:08 +08:00
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
if (c > nb)
|
|
|
|
c = max_align(nb);
|
|
|
|
err = write_mem_aligned(val >> (nb - c) * 8, ea, c);
|
|
|
|
if (err)
|
|
|
|
return err;
|
2013-08-22 22:25:28 +08:00
|
|
|
ea += c;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Write memory at address ea for nb bytes, return 0 for success
|
|
|
|
* or -EFAULT if an error occurred.
|
|
|
|
*/
|
|
|
|
static int __kprobes write_mem(unsigned long val, unsigned long ea, int nb,
|
|
|
|
struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
if (!address_ok(regs, ea, nb))
|
|
|
|
return -EFAULT;
|
|
|
|
if ((ea & (nb - 1)) == 0)
|
|
|
|
return write_mem_aligned(val, ea, nb);
|
|
|
|
return write_mem_unaligned(val, ea, nb, regs);
|
|
|
|
}
|
|
|
|
|
2010-09-01 15:21:21 +08:00
|
|
|
#ifdef CONFIG_PPC_FPU
|
2005-09-26 14:04:21 +08:00
|
|
|
/*
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
* Check the address and alignment, and call func to do the actual
|
|
|
|
* load or store.
|
|
|
|
*/
|
|
|
|
static int __kprobes do_fp_load(int rn, int (*func)(int, unsigned long),
|
|
|
|
unsigned long ea, int nb,
|
|
|
|
struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
int err;
|
2013-10-19 03:44:17 +08:00
|
|
|
union {
|
|
|
|
double dbl;
|
|
|
|
unsigned long ul[2];
|
|
|
|
struct {
|
|
|
|
#ifdef __BIG_ENDIAN__
|
|
|
|
unsigned _pad_;
|
|
|
|
unsigned word;
|
|
|
|
#endif
|
|
|
|
#ifdef __LITTLE_ENDIAN__
|
|
|
|
unsigned word;
|
|
|
|
unsigned _pad_;
|
|
|
|
#endif
|
|
|
|
} single;
|
|
|
|
} data;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
unsigned long ptr;
|
|
|
|
|
|
|
|
if (!address_ok(regs, ea, nb))
|
|
|
|
return -EFAULT;
|
|
|
|
if ((ea & 3) == 0)
|
|
|
|
return (*func)(rn, ea);
|
2013-10-19 03:44:17 +08:00
|
|
|
ptr = (unsigned long) &data.ul;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
if (sizeof(unsigned long) == 8 || nb == 4) {
|
2013-10-19 03:44:17 +08:00
|
|
|
err = read_mem_unaligned(&data.ul[0], ea, nb, regs);
|
|
|
|
if (nb == 4)
|
|
|
|
ptr = (unsigned long)&(data.single.word);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
} else {
|
|
|
|
/* reading a double on 32-bit */
|
2013-10-19 03:44:17 +08:00
|
|
|
err = read_mem_unaligned(&data.ul[0], ea, 4, regs);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
if (!err)
|
2013-10-19 03:44:17 +08:00
|
|
|
err = read_mem_unaligned(&data.ul[1], ea + 4, 4, regs);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
}
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
return (*func)(rn, ptr);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int __kprobes do_fp_store(int rn, int (*func)(int, unsigned long),
|
|
|
|
unsigned long ea, int nb,
|
|
|
|
struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
int err;
|
2013-10-19 03:44:17 +08:00
|
|
|
union {
|
|
|
|
double dbl;
|
|
|
|
unsigned long ul[2];
|
|
|
|
struct {
|
|
|
|
#ifdef __BIG_ENDIAN__
|
|
|
|
unsigned _pad_;
|
|
|
|
unsigned word;
|
|
|
|
#endif
|
|
|
|
#ifdef __LITTLE_ENDIAN__
|
|
|
|
unsigned word;
|
|
|
|
unsigned _pad_;
|
|
|
|
#endif
|
|
|
|
} single;
|
|
|
|
} data;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
unsigned long ptr;
|
|
|
|
|
|
|
|
if (!address_ok(regs, ea, nb))
|
|
|
|
return -EFAULT;
|
|
|
|
if ((ea & 3) == 0)
|
|
|
|
return (*func)(rn, ea);
|
2013-10-19 03:44:17 +08:00
|
|
|
ptr = (unsigned long) &data.ul[0];
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
if (sizeof(unsigned long) == 8 || nb == 4) {
|
2013-10-19 03:44:17 +08:00
|
|
|
if (nb == 4)
|
|
|
|
ptr = (unsigned long)&(data.single.word);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
err = (*func)(rn, ptr);
|
|
|
|
if (err)
|
|
|
|
return err;
|
2013-10-19 03:44:17 +08:00
|
|
|
err = write_mem_unaligned(data.ul[0], ea, nb, regs);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
} else {
|
|
|
|
/* writing a double on 32-bit */
|
|
|
|
err = (*func)(rn, ptr);
|
|
|
|
if (err)
|
|
|
|
return err;
|
2013-10-19 03:44:17 +08:00
|
|
|
err = write_mem_unaligned(data.ul[0], ea, 4, regs);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
if (!err)
|
2013-10-19 03:44:17 +08:00
|
|
|
err = write_mem_unaligned(data.ul[1], ea + 4, 4, regs);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
}
|
|
|
|
return err;
|
|
|
|
}
|
2010-09-01 15:21:21 +08:00
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#ifdef CONFIG_ALTIVEC
|
|
|
|
/* For Altivec/VMX, no need to worry about alignment */
|
|
|
|
static int __kprobes do_vec_load(int rn, int (*func)(int, unsigned long),
|
|
|
|
unsigned long ea, struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
if (!address_ok(regs, ea & ~0xfUL, 16))
|
|
|
|
return -EFAULT;
|
|
|
|
return (*func)(rn, ea);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int __kprobes do_vec_store(int rn, int (*func)(int, unsigned long),
|
|
|
|
unsigned long ea, struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
if (!address_ok(regs, ea & ~0xfUL, 16))
|
|
|
|
return -EFAULT;
|
|
|
|
return (*func)(rn, ea);
|
|
|
|
}
|
|
|
|
#endif /* CONFIG_ALTIVEC */
|
|
|
|
|
|
|
|
#ifdef CONFIG_VSX
|
|
|
|
static int __kprobes do_vsx_load(int rn, int (*func)(int, unsigned long),
|
|
|
|
unsigned long ea, struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
unsigned long val[2];
|
|
|
|
|
|
|
|
if (!address_ok(regs, ea, 16))
|
|
|
|
return -EFAULT;
|
|
|
|
if ((ea & 3) == 0)
|
|
|
|
return (*func)(rn, ea);
|
|
|
|
err = read_mem_unaligned(&val[0], ea, 8, regs);
|
|
|
|
if (!err)
|
|
|
|
err = read_mem_unaligned(&val[1], ea + 8, 8, regs);
|
|
|
|
if (!err)
|
|
|
|
err = (*func)(rn, (unsigned long) &val[0]);
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int __kprobes do_vsx_store(int rn, int (*func)(int, unsigned long),
|
|
|
|
unsigned long ea, struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
unsigned long val[2];
|
|
|
|
|
|
|
|
if (!address_ok(regs, ea, 16))
|
|
|
|
return -EFAULT;
|
|
|
|
if ((ea & 3) == 0)
|
|
|
|
return (*func)(rn, ea);
|
|
|
|
err = (*func)(rn, (unsigned long) &val[0]);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
err = write_mem_unaligned(val[0], ea, 8, regs);
|
|
|
|
if (!err)
|
|
|
|
err = write_mem_unaligned(val[1], ea + 8, 8, regs);
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
#endif /* CONFIG_VSX */
|
|
|
|
|
|
|
|
#define __put_user_asmx(x, addr, err, op, cr) \
|
|
|
|
__asm__ __volatile__( \
|
|
|
|
"1: " op " %2,0,%3\n" \
|
|
|
|
" mfcr %1\n" \
|
|
|
|
"2:\n" \
|
|
|
|
".section .fixup,\"ax\"\n" \
|
|
|
|
"3: li %0,%4\n" \
|
|
|
|
" b 2b\n" \
|
|
|
|
".previous\n" \
|
|
|
|
".section __ex_table,\"a\"\n" \
|
|
|
|
PPC_LONG_ALIGN "\n" \
|
|
|
|
PPC_LONG "1b,3b\n" \
|
|
|
|
".previous" \
|
|
|
|
: "=r" (err), "=r" (cr) \
|
|
|
|
: "r" (x), "r" (addr), "i" (-EFAULT), "0" (err))
|
|
|
|
|
|
|
|
#define __get_user_asmx(x, addr, err, op) \
|
|
|
|
__asm__ __volatile__( \
|
|
|
|
"1: "op" %1,0,%2\n" \
|
|
|
|
"2:\n" \
|
|
|
|
".section .fixup,\"ax\"\n" \
|
|
|
|
"3: li %0,%3\n" \
|
|
|
|
" b 2b\n" \
|
|
|
|
".previous\n" \
|
|
|
|
".section __ex_table,\"a\"\n" \
|
|
|
|
PPC_LONG_ALIGN "\n" \
|
|
|
|
PPC_LONG "1b,3b\n" \
|
|
|
|
".previous" \
|
|
|
|
: "=r" (err), "=r" (x) \
|
|
|
|
: "r" (addr), "i" (-EFAULT), "0" (err))
|
|
|
|
|
|
|
|
#define __cacheop_user_asmx(addr, err, op) \
|
|
|
|
__asm__ __volatile__( \
|
|
|
|
"1: "op" 0,%1\n" \
|
|
|
|
"2:\n" \
|
|
|
|
".section .fixup,\"ax\"\n" \
|
|
|
|
"3: li %0,%3\n" \
|
|
|
|
" b 2b\n" \
|
|
|
|
".previous\n" \
|
|
|
|
".section __ex_table,\"a\"\n" \
|
|
|
|
PPC_LONG_ALIGN "\n" \
|
|
|
|
PPC_LONG "1b,3b\n" \
|
|
|
|
".previous" \
|
|
|
|
: "=r" (err) \
|
|
|
|
: "r" (addr), "i" (-EFAULT), "0" (err))
|
|
|
|
|
|
|
|
static void __kprobes set_cr0(struct pt_regs *regs, int rd)
|
|
|
|
{
|
|
|
|
long val = regs->gpr[rd];
|
|
|
|
|
|
|
|
regs->ccr = (regs->ccr & 0x0fffffff) | ((regs->xer >> 3) & 0x10000000);
|
|
|
|
#ifdef __powerpc64__
|
2011-04-08 05:56:04 +08:00
|
|
|
if (!(regs->msr & MSR_64BIT))
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
val = (int) val;
|
|
|
|
#endif
|
|
|
|
if (val < 0)
|
|
|
|
regs->ccr |= 0x80000000;
|
|
|
|
else if (val > 0)
|
|
|
|
regs->ccr |= 0x40000000;
|
|
|
|
else
|
|
|
|
regs->ccr |= 0x20000000;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __kprobes add_with_carry(struct pt_regs *regs, int rd,
|
|
|
|
unsigned long val1, unsigned long val2,
|
|
|
|
unsigned long carry_in)
|
|
|
|
{
|
|
|
|
unsigned long val = val1 + val2;
|
|
|
|
|
|
|
|
if (carry_in)
|
|
|
|
++val;
|
|
|
|
regs->gpr[rd] = val;
|
|
|
|
#ifdef __powerpc64__
|
2011-04-08 05:56:04 +08:00
|
|
|
if (!(regs->msr & MSR_64BIT)) {
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
val = (unsigned int) val;
|
|
|
|
val1 = (unsigned int) val1;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
if (val < val1 || (carry_in && val == val1))
|
|
|
|
regs->xer |= XER_CA;
|
|
|
|
else
|
|
|
|
regs->xer &= ~XER_CA;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __kprobes do_cmp_signed(struct pt_regs *regs, long v1, long v2,
|
|
|
|
int crfld)
|
|
|
|
{
|
|
|
|
unsigned int crval, shift;
|
|
|
|
|
|
|
|
crval = (regs->xer >> 31) & 1; /* get SO bit */
|
|
|
|
if (v1 < v2)
|
|
|
|
crval |= 8;
|
|
|
|
else if (v1 > v2)
|
|
|
|
crval |= 4;
|
|
|
|
else
|
|
|
|
crval |= 2;
|
|
|
|
shift = (7 - crfld) * 4;
|
|
|
|
regs->ccr = (regs->ccr & ~(0xf << shift)) | (crval << shift);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __kprobes do_cmp_unsigned(struct pt_regs *regs, unsigned long v1,
|
|
|
|
unsigned long v2, int crfld)
|
|
|
|
{
|
|
|
|
unsigned int crval, shift;
|
|
|
|
|
|
|
|
crval = (regs->xer >> 31) & 1; /* get SO bit */
|
|
|
|
if (v1 < v2)
|
|
|
|
crval |= 8;
|
|
|
|
else if (v1 > v2)
|
|
|
|
crval |= 4;
|
|
|
|
else
|
|
|
|
crval |= 2;
|
|
|
|
shift = (7 - crfld) * 4;
|
|
|
|
regs->ccr = (regs->ccr & ~(0xf << shift)) | (crval << shift);
|
|
|
|
}
|
|
|
|
|
2014-09-02 12:35:08 +08:00
|
|
|
static int __kprobes trap_compare(long v1, long v2)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
|
|
|
|
if (v1 < v2)
|
|
|
|
ret |= 0x10;
|
|
|
|
else if (v1 > v2)
|
|
|
|
ret |= 0x08;
|
|
|
|
else
|
|
|
|
ret |= 0x04;
|
|
|
|
if ((unsigned long)v1 < (unsigned long)v2)
|
|
|
|
ret |= 0x02;
|
|
|
|
else if ((unsigned long)v1 > (unsigned long)v2)
|
|
|
|
ret |= 0x01;
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
/*
|
|
|
|
* Elements of 32-bit rotate and mask instructions.
|
|
|
|
*/
|
|
|
|
#define MASK32(mb, me) ((0xffffffffUL >> (mb)) + \
|
|
|
|
((signed long)-0x80000000L >> (me)) + ((me) >= (mb)))
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
#define MASK64_L(mb) (~0UL >> (mb))
|
|
|
|
#define MASK64_R(me) ((signed long)-0x8000000000000000L >> (me))
|
|
|
|
#define MASK64(mb, me) (MASK64_L(mb) + MASK64_R(me) + ((me) >= (mb)))
|
|
|
|
#define DATA32(x) (((x) & 0xffffffffUL) | (((x) & 0xffffffffUL) << 32))
|
|
|
|
#else
|
|
|
|
#define DATA32(x) (x)
|
|
|
|
#endif
|
|
|
|
#define ROTATE(x, n) ((n) ? (((x) << (n)) | ((x) >> (8 * sizeof(long) - (n)))) : (x))
|
|
|
|
|
|
|
|
/*
|
2014-09-02 12:35:07 +08:00
|
|
|
* Decode an instruction, and execute it if that can be done just by
|
|
|
|
* modifying *regs (i.e. integer arithmetic and logical instructions,
|
|
|
|
* branches, and barrier instructions).
|
|
|
|
* Returns 1 if the instruction has been executed, or 0 if not.
|
|
|
|
* Sets *op to indicate what the instruction does.
|
2005-09-26 14:04:21 +08:00
|
|
|
*/
|
2014-09-02 12:35:07 +08:00
|
|
|
int __kprobes analyse_instr(struct instruction_op *op, struct pt_regs *regs,
|
|
|
|
unsigned int instr)
|
2005-09-26 14:04:21 +08:00
|
|
|
{
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
unsigned int opcode, ra, rb, rd, spr, u;
|
2005-09-26 14:04:21 +08:00
|
|
|
unsigned long int imm;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
unsigned long int val, val2;
|
2014-09-02 12:35:07 +08:00
|
|
|
unsigned int mb, me, sh;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
long ival;
|
2005-09-26 14:04:21 +08:00
|
|
|
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = COMPUTE;
|
|
|
|
|
2005-09-26 14:04:21 +08:00
|
|
|
opcode = instr >> 26;
|
|
|
|
switch (opcode) {
|
|
|
|
case 16: /* bc */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = BRANCH;
|
2005-09-26 14:04:21 +08:00
|
|
|
imm = (signed short)(instr & 0xfffc);
|
|
|
|
if ((instr & 2) == 0)
|
|
|
|
imm += regs->nip;
|
|
|
|
regs->nip += 4;
|
2011-04-08 05:56:04 +08:00
|
|
|
regs->nip = truncate_if_32bit(regs->msr, regs->nip);
|
2005-09-26 14:04:21 +08:00
|
|
|
if (instr & 1)
|
|
|
|
regs->link = regs->nip;
|
|
|
|
if (branch_taken(instr, regs))
|
2013-05-06 19:32:40 +08:00
|
|
|
regs->nip = truncate_if_32bit(regs->msr, imm);
|
2005-09-26 14:04:21 +08:00
|
|
|
return 1;
|
2005-10-28 20:48:08 +08:00
|
|
|
#ifdef CONFIG_PPC64
|
2005-09-26 14:04:21 +08:00
|
|
|
case 17: /* sc */
|
2014-09-02 12:35:07 +08:00
|
|
|
if ((instr & 0xfe2) == 2)
|
|
|
|
op->type = SYSCALL;
|
|
|
|
else
|
|
|
|
op->type = UNKNOWN;
|
|
|
|
return 0;
|
2005-10-28 20:48:08 +08:00
|
|
|
#endif
|
2005-09-26 14:04:21 +08:00
|
|
|
case 18: /* b */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = BRANCH;
|
2005-09-26 14:04:21 +08:00
|
|
|
imm = instr & 0x03fffffc;
|
|
|
|
if (imm & 0x02000000)
|
|
|
|
imm -= 0x04000000;
|
|
|
|
if ((instr & 2) == 0)
|
|
|
|
imm += regs->nip;
|
2011-04-08 05:56:04 +08:00
|
|
|
if (instr & 1)
|
|
|
|
regs->link = truncate_if_32bit(regs->msr, regs->nip + 4);
|
|
|
|
imm = truncate_if_32bit(regs->msr, imm);
|
2005-09-26 14:04:21 +08:00
|
|
|
regs->nip = imm;
|
|
|
|
return 1;
|
|
|
|
case 19:
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
switch ((instr >> 1) & 0x3ff) {
|
2014-09-02 12:35:08 +08:00
|
|
|
case 0: /* mcrf */
|
|
|
|
rd = (instr >> 21) & 0x1c;
|
|
|
|
ra = (instr >> 16) & 0x1c;
|
|
|
|
val = (regs->ccr >> ra) & 0xf;
|
|
|
|
regs->ccr = (regs->ccr & ~(0xfUL << rd)) | (val << rd);
|
|
|
|
goto instr_done;
|
|
|
|
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 16: /* bclr */
|
|
|
|
case 528: /* bcctr */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = BRANCH;
|
2005-09-26 14:04:21 +08:00
|
|
|
imm = (instr & 0x400)? regs->ctr: regs->link;
|
2011-04-08 05:56:04 +08:00
|
|
|
regs->nip = truncate_if_32bit(regs->msr, regs->nip + 4);
|
|
|
|
imm = truncate_if_32bit(regs->msr, imm);
|
2005-09-26 14:04:21 +08:00
|
|
|
if (instr & 1)
|
|
|
|
regs->link = regs->nip;
|
|
|
|
if (branch_taken(instr, regs))
|
|
|
|
regs->nip = imm;
|
|
|
|
return 1;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 18: /* rfid, scary */
|
2014-09-02 12:35:07 +08:00
|
|
|
if (regs->msr & MSR_PR)
|
|
|
|
goto priv;
|
|
|
|
op->type = RFI;
|
|
|
|
return 0;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 150: /* isync */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = BARRIER;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
isync();
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 33: /* crnor */
|
|
|
|
case 129: /* crandc */
|
|
|
|
case 193: /* crxor */
|
|
|
|
case 225: /* crnand */
|
|
|
|
case 257: /* crand */
|
|
|
|
case 289: /* creqv */
|
|
|
|
case 417: /* crorc */
|
|
|
|
case 449: /* cror */
|
|
|
|
ra = (instr >> 16) & 0x1f;
|
|
|
|
rb = (instr >> 11) & 0x1f;
|
|
|
|
rd = (instr >> 21) & 0x1f;
|
|
|
|
ra = (regs->ccr >> (31 - ra)) & 1;
|
|
|
|
rb = (regs->ccr >> (31 - rb)) & 1;
|
|
|
|
val = (instr >> (6 + ra * 2 + rb)) & 1;
|
|
|
|
regs->ccr = (regs->ccr & ~(1UL << (31 - rd))) |
|
|
|
|
(val << (31 - rd));
|
|
|
|
goto instr_done;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case 31:
|
|
|
|
switch ((instr >> 1) & 0x3ff) {
|
|
|
|
case 598: /* sync */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = BARRIER;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
#ifdef __powerpc64__
|
|
|
|
switch ((instr >> 21) & 3) {
|
|
|
|
case 1: /* lwsync */
|
|
|
|
asm volatile("lwsync" : : : "memory");
|
|
|
|
goto instr_done;
|
|
|
|
case 2: /* ptesync */
|
|
|
|
asm volatile("ptesync" : : : "memory");
|
|
|
|
goto instr_done;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
mb();
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 854: /* eieio */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = BARRIER;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
eieio();
|
|
|
|
goto instr_done;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Following cases refer to regs->gpr[], so we need all regs */
|
|
|
|
if (!FULL_REGS(regs))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
rd = (instr >> 21) & 0x1f;
|
|
|
|
ra = (instr >> 16) & 0x1f;
|
|
|
|
rb = (instr >> 11) & 0x1f;
|
|
|
|
|
|
|
|
switch (opcode) {
|
2014-09-02 12:35:08 +08:00
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 2: /* tdi */
|
|
|
|
if (rd & trap_compare(regs->gpr[ra], (short) instr))
|
|
|
|
goto trap;
|
|
|
|
goto instr_done;
|
|
|
|
#endif
|
|
|
|
case 3: /* twi */
|
|
|
|
if (rd & trap_compare((int)regs->gpr[ra], (short) instr))
|
|
|
|
goto trap;
|
|
|
|
goto instr_done;
|
|
|
|
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 7: /* mulli */
|
|
|
|
regs->gpr[rd] = regs->gpr[ra] * (short) instr;
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 8: /* subfic */
|
|
|
|
imm = (short) instr;
|
|
|
|
add_with_carry(regs, rd, ~regs->gpr[ra], imm, 1);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 10: /* cmpli */
|
|
|
|
imm = (unsigned short) instr;
|
|
|
|
val = regs->gpr[ra];
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
if ((rd & 1) == 0)
|
|
|
|
val = (unsigned int) val;
|
|
|
|
#endif
|
|
|
|
do_cmp_unsigned(regs, val, imm, rd >> 2);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 11: /* cmpi */
|
|
|
|
imm = (short) instr;
|
|
|
|
val = regs->gpr[ra];
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
if ((rd & 1) == 0)
|
|
|
|
val = (int) val;
|
|
|
|
#endif
|
|
|
|
do_cmp_signed(regs, val, imm, rd >> 2);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 12: /* addic */
|
|
|
|
imm = (short) instr;
|
|
|
|
add_with_carry(regs, rd, regs->gpr[ra], imm, 0);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 13: /* addic. */
|
|
|
|
imm = (short) instr;
|
|
|
|
add_with_carry(regs, rd, regs->gpr[ra], imm, 0);
|
|
|
|
set_cr0(regs, rd);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 14: /* addi */
|
|
|
|
imm = (short) instr;
|
|
|
|
if (ra)
|
|
|
|
imm += regs->gpr[ra];
|
|
|
|
regs->gpr[rd] = imm;
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 15: /* addis */
|
|
|
|
imm = ((short) instr) << 16;
|
|
|
|
if (ra)
|
|
|
|
imm += regs->gpr[ra];
|
|
|
|
regs->gpr[rd] = imm;
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 20: /* rlwimi */
|
|
|
|
mb = (instr >> 6) & 0x1f;
|
|
|
|
me = (instr >> 1) & 0x1f;
|
|
|
|
val = DATA32(regs->gpr[rd]);
|
|
|
|
imm = MASK32(mb, me);
|
|
|
|
regs->gpr[ra] = (regs->gpr[ra] & ~imm) | (ROTATE(val, rb) & imm);
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 21: /* rlwinm */
|
|
|
|
mb = (instr >> 6) & 0x1f;
|
|
|
|
me = (instr >> 1) & 0x1f;
|
|
|
|
val = DATA32(regs->gpr[rd]);
|
|
|
|
regs->gpr[ra] = ROTATE(val, rb) & MASK32(mb, me);
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 23: /* rlwnm */
|
|
|
|
mb = (instr >> 6) & 0x1f;
|
|
|
|
me = (instr >> 1) & 0x1f;
|
|
|
|
rb = regs->gpr[rb] & 0x1f;
|
|
|
|
val = DATA32(regs->gpr[rd]);
|
|
|
|
regs->gpr[ra] = ROTATE(val, rb) & MASK32(mb, me);
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 24: /* ori */
|
|
|
|
imm = (unsigned short) instr;
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] | imm;
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 25: /* oris */
|
|
|
|
imm = (unsigned short) instr;
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] | (imm << 16);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 26: /* xori */
|
|
|
|
imm = (unsigned short) instr;
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] ^ imm;
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 27: /* xoris */
|
|
|
|
imm = (unsigned short) instr;
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] ^ (imm << 16);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 28: /* andi. */
|
|
|
|
imm = (unsigned short) instr;
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] & imm;
|
|
|
|
set_cr0(regs, ra);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 29: /* andis. */
|
|
|
|
imm = (unsigned short) instr;
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] & (imm << 16);
|
|
|
|
set_cr0(regs, ra);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 30: /* rld* */
|
|
|
|
mb = ((instr >> 6) & 0x1f) | (instr & 0x20);
|
|
|
|
val = regs->gpr[rd];
|
|
|
|
if ((instr & 0x10) == 0) {
|
|
|
|
sh = rb | ((instr & 2) << 4);
|
|
|
|
val = ROTATE(val, sh);
|
|
|
|
switch ((instr >> 2) & 3) {
|
|
|
|
case 0: /* rldicl */
|
|
|
|
regs->gpr[ra] = val & MASK64_L(mb);
|
|
|
|
goto logical_done;
|
|
|
|
case 1: /* rldicr */
|
|
|
|
regs->gpr[ra] = val & MASK64_R(mb);
|
|
|
|
goto logical_done;
|
|
|
|
case 2: /* rldic */
|
|
|
|
regs->gpr[ra] = val & MASK64(mb, 63 - sh);
|
|
|
|
goto logical_done;
|
|
|
|
case 3: /* rldimi */
|
|
|
|
imm = MASK64(mb, 63 - sh);
|
|
|
|
regs->gpr[ra] = (regs->gpr[ra] & ~imm) |
|
|
|
|
(val & imm);
|
|
|
|
goto logical_done;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
sh = regs->gpr[rb] & 0x3f;
|
|
|
|
val = ROTATE(val, sh);
|
|
|
|
switch ((instr >> 1) & 7) {
|
|
|
|
case 0: /* rldcl */
|
|
|
|
regs->gpr[ra] = val & MASK64_L(mb);
|
|
|
|
goto logical_done;
|
|
|
|
case 1: /* rldcr */
|
|
|
|
regs->gpr[ra] = val & MASK64_R(mb);
|
|
|
|
goto logical_done;
|
|
|
|
}
|
2005-09-26 14:04:21 +08:00
|
|
|
}
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
#endif
|
|
|
|
|
2005-09-26 14:04:21 +08:00
|
|
|
case 31:
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
switch ((instr >> 1) & 0x3ff) {
|
2014-09-02 12:35:08 +08:00
|
|
|
case 4: /* tw */
|
|
|
|
if (rd == 0x1f ||
|
|
|
|
(rd & trap_compare((int)regs->gpr[ra],
|
|
|
|
(int)regs->gpr[rb])))
|
|
|
|
goto trap;
|
|
|
|
goto instr_done;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 68: /* td */
|
|
|
|
if (rd & trap_compare(regs->gpr[ra], regs->gpr[rb]))
|
|
|
|
goto trap;
|
|
|
|
goto instr_done;
|
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 83: /* mfmsr */
|
|
|
|
if (regs->msr & MSR_PR)
|
2014-09-02 12:35:07 +08:00
|
|
|
goto priv;
|
|
|
|
op->type = MFMSR;
|
|
|
|
op->reg = rd;
|
|
|
|
return 0;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 146: /* mtmsr */
|
|
|
|
if (regs->msr & MSR_PR)
|
2014-09-02 12:35:07 +08:00
|
|
|
goto priv;
|
|
|
|
op->type = MTMSR;
|
|
|
|
op->reg = rd;
|
|
|
|
op->val = 0xffffffff & ~(MSR_ME | MSR_LE);
|
|
|
|
return 0;
|
2005-10-28 20:48:08 +08:00
|
|
|
#ifdef CONFIG_PPC64
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 178: /* mtmsrd */
|
|
|
|
if (regs->msr & MSR_PR)
|
2014-09-02 12:35:07 +08:00
|
|
|
goto priv;
|
|
|
|
op->type = MTMSR;
|
|
|
|
op->reg = rd;
|
|
|
|
/* only MSR_EE and MSR_RI get changed if bit 15 set */
|
|
|
|
/* mtmsrd doesn't change MSR_HV, MSR_ME or MSR_LE */
|
|
|
|
imm = (instr & 0x10000)? 0x8002: 0xefffffffffffeffeUL;
|
|
|
|
op->val = imm;
|
|
|
|
return 0;
|
2005-10-28 20:48:08 +08:00
|
|
|
#endif
|
2014-09-02 12:35:07 +08:00
|
|
|
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 19: /* mfcr */
|
2007-04-18 13:56:38 +08:00
|
|
|
regs->gpr[rd] = regs->ccr;
|
|
|
|
regs->gpr[rd] &= 0xffffffffUL;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 144: /* mtcrf */
|
|
|
|
imm = 0xf0000000UL;
|
|
|
|
val = regs->gpr[rd];
|
|
|
|
for (sh = 0; sh < 8; ++sh) {
|
|
|
|
if (instr & (0x80000 >> sh))
|
|
|
|
regs->ccr = (regs->ccr & ~imm) |
|
|
|
|
(val & imm);
|
|
|
|
imm >>= 4;
|
|
|
|
}
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 339: /* mfspr */
|
2014-09-02 12:35:07 +08:00
|
|
|
spr = ((instr >> 16) & 0x1f) | ((instr >> 6) & 0x3e0);
|
2007-04-18 13:56:38 +08:00
|
|
|
switch (spr) {
|
2014-09-02 12:35:07 +08:00
|
|
|
case SPRN_XER: /* mfxer */
|
2007-04-18 13:56:38 +08:00
|
|
|
regs->gpr[rd] = regs->xer;
|
|
|
|
regs->gpr[rd] &= 0xffffffffUL;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
goto instr_done;
|
2014-09-02 12:35:07 +08:00
|
|
|
case SPRN_LR: /* mflr */
|
2007-04-18 13:56:38 +08:00
|
|
|
regs->gpr[rd] = regs->link;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
goto instr_done;
|
2014-09-02 12:35:07 +08:00
|
|
|
case SPRN_CTR: /* mfctr */
|
2007-04-18 13:56:38 +08:00
|
|
|
regs->gpr[rd] = regs->ctr;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
goto instr_done;
|
2014-09-02 12:35:07 +08:00
|
|
|
default:
|
|
|
|
op->type = MFSPR;
|
|
|
|
op->reg = rd;
|
|
|
|
op->spr = spr;
|
|
|
|
return 0;
|
2007-04-18 13:56:38 +08:00
|
|
|
}
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 467: /* mtspr */
|
2014-09-02 12:35:07 +08:00
|
|
|
spr = ((instr >> 16) & 0x1f) | ((instr >> 6) & 0x3e0);
|
2007-04-18 13:56:38 +08:00
|
|
|
switch (spr) {
|
2014-09-02 12:35:07 +08:00
|
|
|
case SPRN_XER: /* mtxer */
|
2007-04-18 13:56:38 +08:00
|
|
|
regs->xer = (regs->gpr[rd] & 0xffffffffUL);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
goto instr_done;
|
2014-09-02 12:35:07 +08:00
|
|
|
case SPRN_LR: /* mtlr */
|
2007-04-18 13:56:38 +08:00
|
|
|
regs->link = regs->gpr[rd];
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
goto instr_done;
|
2014-09-02 12:35:07 +08:00
|
|
|
case SPRN_CTR: /* mtctr */
|
2007-04-18 13:56:38 +08:00
|
|
|
regs->ctr = regs->gpr[rd];
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
goto instr_done;
|
2014-09-02 12:35:07 +08:00
|
|
|
default:
|
|
|
|
op->type = MTSPR;
|
|
|
|
op->val = regs->gpr[rd];
|
|
|
|
op->spr = spr;
|
|
|
|
return 0;
|
2007-04-18 13:56:38 +08:00
|
|
|
}
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
break;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Compare instructions
|
|
|
|
*/
|
|
|
|
case 0: /* cmp */
|
|
|
|
val = regs->gpr[ra];
|
|
|
|
val2 = regs->gpr[rb];
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
if ((rd & 1) == 0) {
|
|
|
|
/* word (32-bit) compare */
|
|
|
|
val = (int) val;
|
|
|
|
val2 = (int) val2;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
do_cmp_signed(regs, val, val2, rd >> 2);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case 32: /* cmpl */
|
|
|
|
val = regs->gpr[ra];
|
|
|
|
val2 = regs->gpr[rb];
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
if ((rd & 1) == 0) {
|
|
|
|
/* word (32-bit) compare */
|
|
|
|
val = (unsigned int) val;
|
|
|
|
val2 = (unsigned int) val2;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
do_cmp_unsigned(regs, val, val2, rd >> 2);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Arithmetic instructions
|
|
|
|
*/
|
|
|
|
case 8: /* subfc */
|
|
|
|
add_with_carry(regs, rd, ~regs->gpr[ra],
|
|
|
|
regs->gpr[rb], 1);
|
|
|
|
goto arith_done;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 9: /* mulhdu */
|
|
|
|
asm("mulhdu %0,%1,%2" : "=r" (regs->gpr[rd]) :
|
|
|
|
"r" (regs->gpr[ra]), "r" (regs->gpr[rb]));
|
|
|
|
goto arith_done;
|
|
|
|
#endif
|
|
|
|
case 10: /* addc */
|
|
|
|
add_with_carry(regs, rd, regs->gpr[ra],
|
|
|
|
regs->gpr[rb], 0);
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 11: /* mulhwu */
|
|
|
|
asm("mulhwu %0,%1,%2" : "=r" (regs->gpr[rd]) :
|
|
|
|
"r" (regs->gpr[ra]), "r" (regs->gpr[rb]));
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 40: /* subf */
|
|
|
|
regs->gpr[rd] = regs->gpr[rb] - regs->gpr[ra];
|
|
|
|
goto arith_done;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 73: /* mulhd */
|
|
|
|
asm("mulhd %0,%1,%2" : "=r" (regs->gpr[rd]) :
|
|
|
|
"r" (regs->gpr[ra]), "r" (regs->gpr[rb]));
|
|
|
|
goto arith_done;
|
|
|
|
#endif
|
|
|
|
case 75: /* mulhw */
|
|
|
|
asm("mulhw %0,%1,%2" : "=r" (regs->gpr[rd]) :
|
|
|
|
"r" (regs->gpr[ra]), "r" (regs->gpr[rb]));
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 104: /* neg */
|
|
|
|
regs->gpr[rd] = -regs->gpr[ra];
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 136: /* subfe */
|
|
|
|
add_with_carry(regs, rd, ~regs->gpr[ra], regs->gpr[rb],
|
|
|
|
regs->xer & XER_CA);
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 138: /* adde */
|
|
|
|
add_with_carry(regs, rd, regs->gpr[ra], regs->gpr[rb],
|
|
|
|
regs->xer & XER_CA);
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 200: /* subfze */
|
|
|
|
add_with_carry(regs, rd, ~regs->gpr[ra], 0L,
|
|
|
|
regs->xer & XER_CA);
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 202: /* addze */
|
|
|
|
add_with_carry(regs, rd, regs->gpr[ra], 0L,
|
|
|
|
regs->xer & XER_CA);
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 232: /* subfme */
|
|
|
|
add_with_carry(regs, rd, ~regs->gpr[ra], -1L,
|
|
|
|
regs->xer & XER_CA);
|
|
|
|
goto arith_done;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 233: /* mulld */
|
|
|
|
regs->gpr[rd] = regs->gpr[ra] * regs->gpr[rb];
|
|
|
|
goto arith_done;
|
|
|
|
#endif
|
|
|
|
case 234: /* addme */
|
|
|
|
add_with_carry(regs, rd, regs->gpr[ra], -1L,
|
|
|
|
regs->xer & XER_CA);
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 235: /* mullw */
|
|
|
|
regs->gpr[rd] = (unsigned int) regs->gpr[ra] *
|
|
|
|
(unsigned int) regs->gpr[rb];
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
case 266: /* add */
|
|
|
|
regs->gpr[rd] = regs->gpr[ra] + regs->gpr[rb];
|
|
|
|
goto arith_done;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 457: /* divdu */
|
|
|
|
regs->gpr[rd] = regs->gpr[ra] / regs->gpr[rb];
|
|
|
|
goto arith_done;
|
|
|
|
#endif
|
|
|
|
case 459: /* divwu */
|
|
|
|
regs->gpr[rd] = (unsigned int) regs->gpr[ra] /
|
|
|
|
(unsigned int) regs->gpr[rb];
|
|
|
|
goto arith_done;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 489: /* divd */
|
|
|
|
regs->gpr[rd] = (long int) regs->gpr[ra] /
|
|
|
|
(long int) regs->gpr[rb];
|
|
|
|
goto arith_done;
|
|
|
|
#endif
|
|
|
|
case 491: /* divw */
|
|
|
|
regs->gpr[rd] = (int) regs->gpr[ra] /
|
|
|
|
(int) regs->gpr[rb];
|
|
|
|
goto arith_done;
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Logical instructions
|
|
|
|
*/
|
|
|
|
case 26: /* cntlzw */
|
|
|
|
asm("cntlzw %0,%1" : "=r" (regs->gpr[ra]) :
|
|
|
|
"r" (regs->gpr[rd]));
|
|
|
|
goto logical_done;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 58: /* cntlzd */
|
|
|
|
asm("cntlzd %0,%1" : "=r" (regs->gpr[ra]) :
|
|
|
|
"r" (regs->gpr[rd]));
|
|
|
|
goto logical_done;
|
|
|
|
#endif
|
|
|
|
case 28: /* and */
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] & regs->gpr[rb];
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 60: /* andc */
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] & ~regs->gpr[rb];
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 124: /* nor */
|
|
|
|
regs->gpr[ra] = ~(regs->gpr[rd] | regs->gpr[rb]);
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 284: /* xor */
|
|
|
|
regs->gpr[ra] = ~(regs->gpr[rd] ^ regs->gpr[rb]);
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 316: /* xor */
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] ^ regs->gpr[rb];
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 412: /* orc */
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] | ~regs->gpr[rb];
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 444: /* or */
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] | regs->gpr[rb];
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 476: /* nand */
|
|
|
|
regs->gpr[ra] = ~(regs->gpr[rd] & regs->gpr[rb]);
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 922: /* extsh */
|
|
|
|
regs->gpr[ra] = (signed short) regs->gpr[rd];
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 954: /* extsb */
|
|
|
|
regs->gpr[ra] = (signed char) regs->gpr[rd];
|
|
|
|
goto logical_done;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 986: /* extsw */
|
|
|
|
regs->gpr[ra] = (signed int) regs->gpr[rd];
|
|
|
|
goto logical_done;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Shift instructions
|
|
|
|
*/
|
|
|
|
case 24: /* slw */
|
|
|
|
sh = regs->gpr[rb] & 0x3f;
|
|
|
|
if (sh < 32)
|
|
|
|
regs->gpr[ra] = (regs->gpr[rd] << sh) & 0xffffffffUL;
|
|
|
|
else
|
|
|
|
regs->gpr[ra] = 0;
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 536: /* srw */
|
|
|
|
sh = regs->gpr[rb] & 0x3f;
|
|
|
|
if (sh < 32)
|
|
|
|
regs->gpr[ra] = (regs->gpr[rd] & 0xffffffffUL) >> sh;
|
|
|
|
else
|
|
|
|
regs->gpr[ra] = 0;
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 792: /* sraw */
|
|
|
|
sh = regs->gpr[rb] & 0x3f;
|
|
|
|
ival = (signed int) regs->gpr[rd];
|
|
|
|
regs->gpr[ra] = ival >> (sh < 32 ? sh : 31);
|
2014-07-19 15:47:57 +08:00
|
|
|
if (ival < 0 && (sh >= 32 || (ival & ((1ul << sh) - 1)) != 0))
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
regs->xer |= XER_CA;
|
|
|
|
else
|
|
|
|
regs->xer &= ~XER_CA;
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 824: /* srawi */
|
|
|
|
sh = rb;
|
|
|
|
ival = (signed int) regs->gpr[rd];
|
|
|
|
regs->gpr[ra] = ival >> sh;
|
2014-07-19 15:47:57 +08:00
|
|
|
if (ival < 0 && (ival & ((1ul << sh) - 1)) != 0)
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
regs->xer |= XER_CA;
|
|
|
|
else
|
|
|
|
regs->xer &= ~XER_CA;
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 27: /* sld */
|
2014-07-19 15:47:57 +08:00
|
|
|
sh = regs->gpr[rb] & 0x7f;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
if (sh < 64)
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] << sh;
|
|
|
|
else
|
|
|
|
regs->gpr[ra] = 0;
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 539: /* srd */
|
|
|
|
sh = regs->gpr[rb] & 0x7f;
|
|
|
|
if (sh < 64)
|
|
|
|
regs->gpr[ra] = regs->gpr[rd] >> sh;
|
|
|
|
else
|
|
|
|
regs->gpr[ra] = 0;
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 794: /* srad */
|
|
|
|
sh = regs->gpr[rb] & 0x7f;
|
|
|
|
ival = (signed long int) regs->gpr[rd];
|
|
|
|
regs->gpr[ra] = ival >> (sh < 64 ? sh : 63);
|
2014-07-19 15:47:57 +08:00
|
|
|
if (ival < 0 && (sh >= 64 || (ival & ((1ul << sh) - 1)) != 0))
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
regs->xer |= XER_CA;
|
|
|
|
else
|
|
|
|
regs->xer &= ~XER_CA;
|
|
|
|
goto logical_done;
|
|
|
|
|
|
|
|
case 826: /* sradi with sh_5 = 0 */
|
|
|
|
case 827: /* sradi with sh_5 = 1 */
|
|
|
|
sh = rb | ((instr & 2) << 4);
|
|
|
|
ival = (signed long int) regs->gpr[rd];
|
|
|
|
regs->gpr[ra] = ival >> sh;
|
2014-07-19 15:47:57 +08:00
|
|
|
if (ival < 0 && (ival & ((1ul << sh) - 1)) != 0)
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
regs->xer |= XER_CA;
|
|
|
|
else
|
|
|
|
regs->xer &= ~XER_CA;
|
|
|
|
goto logical_done;
|
|
|
|
#endif /* __powerpc64__ */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Cache instructions
|
|
|
|
*/
|
|
|
|
case 54: /* dcbst */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(CACHEOP, DCBST, 0);
|
|
|
|
op->ea = xform_ea(instr, regs);
|
|
|
|
return 0;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 86: /* dcbf */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(CACHEOP, DCBF, 0);
|
|
|
|
op->ea = xform_ea(instr, regs);
|
|
|
|
return 0;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 246: /* dcbtst */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(CACHEOP, DCBTST, 0);
|
|
|
|
op->ea = xform_ea(instr, regs);
|
|
|
|
op->reg = rd;
|
|
|
|
return 0;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 278: /* dcbt */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(CACHEOP, DCBTST, 0);
|
|
|
|
op->ea = xform_ea(instr, regs);
|
|
|
|
op->reg = rd;
|
|
|
|
return 0;
|
2014-09-02 12:35:08 +08:00
|
|
|
|
|
|
|
case 982: /* icbi */
|
|
|
|
op->type = MKOP(CACHEOP, ICBI, 0);
|
|
|
|
op->ea = xform_ea(instr, regs);
|
|
|
|
return 0;
|
2005-09-26 14:04:21 +08:00
|
|
|
}
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
break;
|
2005-09-26 14:04:21 +08:00
|
|
|
}
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
/*
|
2014-09-02 12:35:07 +08:00
|
|
|
* Loads and stores.
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
*/
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = UNKNOWN;
|
|
|
|
op->update_reg = ra;
|
|
|
|
op->reg = rd;
|
|
|
|
op->val = regs->gpr[rd];
|
|
|
|
u = (instr >> 20) & UPDATE;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
switch (opcode) {
|
|
|
|
case 31:
|
2014-09-02 12:35:07 +08:00
|
|
|
u = instr & UPDATE;
|
|
|
|
op->ea = xform_ea(instr, regs);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
switch ((instr >> 1) & 0x3ff) {
|
|
|
|
case 20: /* lwarx */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LARX, 0, 4);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 150: /* stwcx. */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STCX, 0, 4);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 84: /* ldarx */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LARX, 0, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 214: /* stdcx. */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STCX, 0, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 21: /* ldx */
|
|
|
|
case 53: /* ldux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, u, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
#endif
|
|
|
|
|
|
|
|
case 23: /* lwzx */
|
|
|
|
case 55: /* lwzux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, u, 4);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 87: /* lbzx */
|
|
|
|
case 119: /* lbzux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, u, 1);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#ifdef CONFIG_ALTIVEC
|
|
|
|
case 103: /* lvx */
|
|
|
|
case 359: /* lvxl */
|
|
|
|
if (!(regs->msr & MSR_VEC))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto vecunavail;
|
|
|
|
op->type = MKOP(LOAD_VMX, 0, 16);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 231: /* stvx */
|
|
|
|
case 487: /* stvxl */
|
|
|
|
if (!(regs->msr & MSR_VEC))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto vecunavail;
|
|
|
|
op->type = MKOP(STORE_VMX, 0, 16);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
#endif /* CONFIG_ALTIVEC */
|
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 149: /* stdx */
|
|
|
|
case 181: /* stdux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, u, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
#endif
|
|
|
|
|
|
|
|
case 151: /* stwx */
|
|
|
|
case 183: /* stwux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, u, 4);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 215: /* stbx */
|
|
|
|
case 247: /* stbux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, u, 1);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 279: /* lhzx */
|
|
|
|
case 311: /* lhzux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, u, 2);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 341: /* lwax */
|
|
|
|
case 373: /* lwaux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, SIGNEXT | u, 4);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
#endif
|
|
|
|
|
|
|
|
case 343: /* lhax */
|
|
|
|
case 375: /* lhaux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, SIGNEXT | u, 2);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 407: /* sthx */
|
|
|
|
case 439: /* sthux */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, u, 2);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 532: /* ldbrx */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, BYTEREV, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#endif
|
2014-09-02 12:35:09 +08:00
|
|
|
case 533: /* lswx */
|
|
|
|
op->type = MKOP(LOAD_MULTI, 0, regs->xer & 0x7f);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 534: /* lwbrx */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, BYTEREV, 4);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
2014-09-02 12:35:09 +08:00
|
|
|
case 597: /* lswi */
|
|
|
|
if (rb == 0)
|
|
|
|
rb = 32; /* # bytes to load */
|
|
|
|
op->type = MKOP(LOAD_MULTI, 0, rb);
|
|
|
|
op->ea = 0;
|
|
|
|
if (ra)
|
|
|
|
op->ea = truncate_if_32bit(regs->msr,
|
|
|
|
regs->gpr[ra]);
|
|
|
|
break;
|
|
|
|
|
2014-05-21 03:59:42 +08:00
|
|
|
#ifdef CONFIG_PPC_FPU
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 535: /* lfsx */
|
|
|
|
case 567: /* lfsux */
|
|
|
|
if (!(regs->msr & MSR_FP))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto fpunavail;
|
|
|
|
op->type = MKOP(LOAD_FP, u, 4);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 599: /* lfdx */
|
|
|
|
case 631: /* lfdux */
|
|
|
|
if (!(regs->msr & MSR_FP))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto fpunavail;
|
|
|
|
op->type = MKOP(LOAD_FP, u, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 663: /* stfsx */
|
|
|
|
case 695: /* stfsux */
|
|
|
|
if (!(regs->msr & MSR_FP))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto fpunavail;
|
|
|
|
op->type = MKOP(STORE_FP, u, 4);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 727: /* stfdx */
|
|
|
|
case 759: /* stfdux */
|
|
|
|
if (!(regs->msr & MSR_FP))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto fpunavail;
|
|
|
|
op->type = MKOP(STORE_FP, u, 8);
|
|
|
|
break;
|
2010-09-01 15:21:21 +08:00
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 660: /* stdbrx */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, BYTEREV, 8);
|
|
|
|
op->val = byterev_8(regs->gpr[rd]);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#endif
|
2014-09-02 12:35:09 +08:00
|
|
|
case 661: /* stswx */
|
|
|
|
op->type = MKOP(STORE_MULTI, 0, regs->xer & 0x7f);
|
|
|
|
break;
|
|
|
|
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 662: /* stwbrx */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, BYTEREV, 4);
|
|
|
|
op->val = byterev_4(regs->gpr[rd]);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
2014-09-02 12:35:09 +08:00
|
|
|
case 725:
|
|
|
|
if (rb == 0)
|
|
|
|
rb = 32; /* # bytes to store */
|
|
|
|
op->type = MKOP(STORE_MULTI, 0, rb);
|
|
|
|
op->ea = 0;
|
|
|
|
if (ra)
|
|
|
|
op->ea = truncate_if_32bit(regs->msr,
|
|
|
|
regs->gpr[ra]);
|
|
|
|
break;
|
|
|
|
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 790: /* lhbrx */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, BYTEREV, 2);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 918: /* sthbrx */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, BYTEREV, 2);
|
|
|
|
op->val = byterev_2(regs->gpr[rd]);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#ifdef CONFIG_VSX
|
|
|
|
case 844: /* lxvd2x */
|
|
|
|
case 876: /* lxvd2ux */
|
|
|
|
if (!(regs->msr & MSR_VSX))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto vsxunavail;
|
|
|
|
op->reg = rd | ((instr & 1) << 5);
|
|
|
|
op->type = MKOP(LOAD_VSX, u, 16);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 972: /* stxvd2x */
|
|
|
|
case 1004: /* stxvd2ux */
|
|
|
|
if (!(regs->msr & MSR_VSX))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto vsxunavail;
|
|
|
|
op->reg = rd | ((instr & 1) << 5);
|
|
|
|
op->type = MKOP(STORE_VSX, u, 16);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#endif /* CONFIG_VSX */
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case 32: /* lwz */
|
|
|
|
case 33: /* lwzu */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, u, 4);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 34: /* lbz */
|
|
|
|
case 35: /* lbzu */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, u, 1);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 36: /* stw */
|
2012-09-17 07:54:31 +08:00
|
|
|
case 37: /* stwu */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, u, 4);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
2012-09-17 07:54:31 +08:00
|
|
|
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 38: /* stb */
|
|
|
|
case 39: /* stbu */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, u, 1);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 40: /* lhz */
|
|
|
|
case 41: /* lhzu */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, u, 2);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 42: /* lha */
|
|
|
|
case 43: /* lhau */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, SIGNEXT | u, 2);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 44: /* sth */
|
|
|
|
case 45: /* sthu */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, u, 2);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 46: /* lmw */
|
|
|
|
if (ra >= rd)
|
|
|
|
break; /* invalid form, ra in range to load */
|
2014-09-02 12:35:09 +08:00
|
|
|
op->type = MKOP(LOAD_MULTI, 0, 4 * (32 - rd));
|
2014-09-02 12:35:07 +08:00
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 47: /* stmw */
|
2014-09-02 12:35:09 +08:00
|
|
|
op->type = MKOP(STORE_MULTI, 0, 4 * (32 - rd));
|
2014-09-02 12:35:07 +08:00
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
2010-09-01 15:21:21 +08:00
|
|
|
#ifdef CONFIG_PPC_FPU
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 48: /* lfs */
|
|
|
|
case 49: /* lfsu */
|
|
|
|
if (!(regs->msr & MSR_FP))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto fpunavail;
|
|
|
|
op->type = MKOP(LOAD_FP, u, 4);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 50: /* lfd */
|
|
|
|
case 51: /* lfdu */
|
|
|
|
if (!(regs->msr & MSR_FP))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto fpunavail;
|
|
|
|
op->type = MKOP(LOAD_FP, u, 8);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 52: /* stfs */
|
|
|
|
case 53: /* stfsu */
|
|
|
|
if (!(regs->msr & MSR_FP))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto fpunavail;
|
|
|
|
op->type = MKOP(STORE_FP, u, 4);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
case 54: /* stfd */
|
|
|
|
case 55: /* stfdu */
|
|
|
|
if (!(regs->msr & MSR_FP))
|
2014-09-02 12:35:07 +08:00
|
|
|
goto fpunavail;
|
|
|
|
op->type = MKOP(STORE_FP, u, 8);
|
|
|
|
op->ea = dform_ea(instr, regs);
|
|
|
|
break;
|
2010-09-01 15:21:21 +08:00
|
|
|
#endif
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 58: /* ld[u], lwa */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->ea = dsform_ea(instr, regs);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
switch (instr & 3) {
|
|
|
|
case 0: /* ld */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, 0, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 1: /* ldu */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, UPDATE, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 2: /* lwa */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(LOAD, SIGNEXT, 4);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case 62: /* std[u] */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->ea = dsform_ea(instr, regs);
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
switch (instr & 3) {
|
|
|
|
case 0: /* std */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, 0, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
case 1: /* stdu */
|
2014-09-02 12:35:07 +08:00
|
|
|
op->type = MKOP(STORE, UPDATE, 8);
|
|
|
|
break;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
#endif /* __powerpc64__ */
|
|
|
|
|
|
|
|
}
|
2014-09-02 12:35:07 +08:00
|
|
|
return 0;
|
powerpc: Emulate most Book I instructions in emulate_step()
This extends the emulate_step() function to handle a large proportion
of the Book I instructions implemented on current 64-bit server
processors. The aim is to handle all the load and store instructions
used in the kernel, plus all of the instructions that appear between
l[wd]arx and st[wd]cx., so this handles the Altivec/VMX lvx and stvx
and the VSX lxv2dx and stxv2dx instructions (implemented in POWER7).
The new code can emulate user mode instructions, and checks the
effective address for a load or store if the saved state is for
user mode. It doesn't handle little-endian mode at present.
For floating-point, Altivec/VMX and VSX instructions, it checks
that the saved MSR has the enable bit for the relevant facility
set, and if so, assumes that the FP/VMX/VSX registers contain
valid state, and does loads or stores directly to/from the
FP/VMX/VSX registers, using assembly helpers in ldstfp.S.
Instructions supported now include:
* Loads and stores, including some but not all VMX and VSX instructions,
and lmw/stmw
* Atomic loads and stores (l[dw]arx, st[dw]cx.)
* Arithmetic instructions (add, subtract, multiply, divide, etc.)
* Compare instructions
* Rotate and mask instructions
* Shift instructions
* Logical instructions (and, or, xor, etc.)
* Condition register logical instructions
* mtcrf, cntlz[wd], exts[bhw]
* isync, sync, lwsync, ptesync, eieio
* Cache operations (dcbf, dcbst, dcbt, dcbtst)
The overflow-checking arithmetic instructions are not included, but
they appear not to be ever used in C code.
This uses decimal values for the minor opcodes in the switch statements
because that is what appears in the Power ISA specification, thus it is
easier to check that they are correct if they are in decimal.
If this is used to single-step an instruction where a data breakpoint
interrupt occurred, then there is the possibility that the instruction
is a lwarx or ldarx. In that case we have to be careful not to lose the
reservation until we get to the matching st[wd]cx., or we'll never make
forward progress. One alternative is to try to arrange that we can
return from interrupts and handle data breakpoint interrupts without
losing the reservation, which means not using any spinlocks, mutexes,
or atomic ops (including bitops). That seems rather fragile. The
other alternative is to emulate the larx/stcx and all the instructions
in between. This is why this commit adds support for a wide range
of integer instructions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
2010-06-15 12:48:58 +08:00
|
|
|
|
|
|
|
logical_done:
|
|
|
|
if (instr & 1)
|
|
|
|
set_cr0(regs, ra);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
arith_done:
|
|
|
|
if (instr & 1)
|
|
|
|
set_cr0(regs, rd);
|
2014-09-02 12:35:07 +08:00
|
|
|
|
|
|
|
instr_done:
|
|
|
|
regs->nip = truncate_if_32bit(regs->msr, regs->nip + 4);
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
priv:
|
|
|
|
op->type = INTERRUPT | 0x700;
|
|
|
|
op->val = SRR1_PROGPRIV;
|
|
|
|
return 0;
|
|
|
|
|
2014-09-02 12:35:08 +08:00
|
|
|
trap:
|
|
|
|
op->type = INTERRUPT | 0x700;
|
|
|
|
op->val = SRR1_PROGTRAP;
|
|
|
|
return 0;
|
|
|
|
|
2014-09-02 12:35:07 +08:00
|
|
|
#ifdef CONFIG_PPC_FPU
|
|
|
|
fpunavail:
|
|
|
|
op->type = INTERRUPT | 0x800;
|
|
|
|
return 0;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#ifdef CONFIG_ALTIVEC
|
|
|
|
vecunavail:
|
|
|
|
op->type = INTERRUPT | 0xf20;
|
|
|
|
return 0;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#ifdef CONFIG_VSX
|
|
|
|
vsxunavail:
|
|
|
|
op->type = INTERRUPT | 0xf40;
|
|
|
|
return 0;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(analyse_instr);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* For PPC32 we always use stwu with r1 to change the stack pointer.
|
|
|
|
* So this emulated store may corrupt the exception frame, now we
|
|
|
|
* have to provide the exception frame trampoline, which is pushed
|
|
|
|
* below the kprobed function stack. So we only update gpr[1] but
|
|
|
|
* don't emulate the real store operation. We will do real store
|
|
|
|
* operation safely in exception return code by checking this flag.
|
|
|
|
*/
|
|
|
|
static __kprobes int handle_stack_update(unsigned long ea, struct pt_regs *regs)
|
|
|
|
{
|
|
|
|
#ifdef CONFIG_PPC32
|
|
|
|
/*
|
|
|
|
* Check if we will touch kernel stack overflow
|
|
|
|
*/
|
|
|
|
if (ea - STACK_INT_FRAME_SIZE <= current->thread.ksp_limit) {
|
|
|
|
printk(KERN_CRIT "Can't kprobe this since kernel stack would overflow.\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
#endif /* CONFIG_PPC32 */
|
|
|
|
/*
|
|
|
|
* Check if we already set since that means we'll
|
|
|
|
* lose the previous value.
|
|
|
|
*/
|
|
|
|
WARN_ON(test_thread_flag(TIF_EMULATE_STACK_STORE));
|
|
|
|
set_thread_flag(TIF_EMULATE_STACK_STORE);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static __kprobes void do_signext(unsigned long *valp, int size)
|
|
|
|
{
|
|
|
|
switch (size) {
|
|
|
|
case 2:
|
|
|
|
*valp = (signed short) *valp;
|
|
|
|
break;
|
|
|
|
case 4:
|
|
|
|
*valp = (signed int) *valp;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static __kprobes void do_byterev(unsigned long *valp, int size)
|
|
|
|
{
|
|
|
|
switch (size) {
|
|
|
|
case 2:
|
|
|
|
*valp = byterev_2(*valp);
|
|
|
|
break;
|
|
|
|
case 4:
|
|
|
|
*valp = byterev_4(*valp);
|
|
|
|
break;
|
|
|
|
#ifdef __powerpc64__
|
|
|
|
case 8:
|
|
|
|
*valp = byterev_8(*valp);
|
|
|
|
break;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Emulate instructions that cause a transfer of control,
|
|
|
|
* loads and stores, and a few other instructions.
|
|
|
|
* Returns 1 if the step was emulated, 0 if not,
|
|
|
|
* or -1 if the instruction is one that should not be stepped,
|
|
|
|
* such as an rfid, or a mtmsrd that would clear MSR_RI.
|
|
|
|
*/
|
|
|
|
int __kprobes emulate_step(struct pt_regs *regs, unsigned int instr)
|
|
|
|
{
|
|
|
|
struct instruction_op op;
|
|
|
|
int r, err, size;
|
|
|
|
unsigned long val;
|
|
|
|
unsigned int cr;
|
2014-09-02 12:35:09 +08:00
|
|
|
int i, rd, nb;
|
2014-09-02 12:35:07 +08:00
|
|
|
|
|
|
|
r = analyse_instr(&op, regs, instr);
|
|
|
|
if (r != 0)
|
|
|
|
return r;
|
|
|
|
|
|
|
|
err = 0;
|
|
|
|
size = GETSIZE(op.type);
|
|
|
|
switch (op.type & INSTR_TYPE_MASK) {
|
|
|
|
case CACHEOP:
|
|
|
|
if (!address_ok(regs, op.ea, 8))
|
|
|
|
return 0;
|
|
|
|
switch (op.type & CACHEOP_MASK) {
|
|
|
|
case DCBST:
|
|
|
|
__cacheop_user_asmx(op.ea, err, "dcbst");
|
|
|
|
break;
|
|
|
|
case DCBF:
|
|
|
|
__cacheop_user_asmx(op.ea, err, "dcbf");
|
|
|
|
break;
|
|
|
|
case DCBTST:
|
|
|
|
if (op.reg == 0)
|
|
|
|
prefetchw((void *) op.ea);
|
|
|
|
break;
|
|
|
|
case DCBT:
|
|
|
|
if (op.reg == 0)
|
|
|
|
prefetch((void *) op.ea);
|
|
|
|
break;
|
2014-09-02 12:35:08 +08:00
|
|
|
case ICBI:
|
|
|
|
__cacheop_user_asmx(op.ea, err, "icbi");
|
|
|
|
break;
|
2014-09-02 12:35:07 +08:00
|
|
|
}
|
|
|
|
if (err)
|
|
|
|
return 0;
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case LARX:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
if (op.ea & (size - 1))
|
|
|
|
break; /* can't handle misaligned */
|
|
|
|
err = -EFAULT;
|
|
|
|
if (!address_ok(regs, op.ea, size))
|
|
|
|
goto ldst_done;
|
|
|
|
err = 0;
|
|
|
|
switch (size) {
|
|
|
|
case 4:
|
|
|
|
__get_user_asmx(val, op.ea, err, "lwarx");
|
|
|
|
break;
|
|
|
|
case 8:
|
|
|
|
__get_user_asmx(val, op.ea, err, "ldarx");
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
if (!err)
|
|
|
|
regs->gpr[op.reg] = val;
|
|
|
|
goto ldst_done;
|
|
|
|
|
|
|
|
case STCX:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
if (op.ea & (size - 1))
|
|
|
|
break; /* can't handle misaligned */
|
|
|
|
err = -EFAULT;
|
|
|
|
if (!address_ok(regs, op.ea, size))
|
|
|
|
goto ldst_done;
|
|
|
|
err = 0;
|
|
|
|
switch (size) {
|
|
|
|
case 4:
|
|
|
|
__put_user_asmx(op.val, op.ea, err, "stwcx.", cr);
|
|
|
|
break;
|
|
|
|
case 8:
|
|
|
|
__put_user_asmx(op.val, op.ea, err, "stdcx.", cr);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
if (!err)
|
|
|
|
regs->ccr = (regs->ccr & 0x0fffffff) |
|
|
|
|
(cr & 0xe0000000) |
|
|
|
|
((regs->xer >> 3) & 0x10000000);
|
|
|
|
goto ldst_done;
|
|
|
|
|
|
|
|
case LOAD:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
err = read_mem(®s->gpr[op.reg], op.ea, size, regs);
|
|
|
|
if (!err) {
|
|
|
|
if (op.type & SIGNEXT)
|
|
|
|
do_signext(®s->gpr[op.reg], size);
|
|
|
|
if (op.type & BYTEREV)
|
|
|
|
do_byterev(®s->gpr[op.reg], size);
|
|
|
|
}
|
|
|
|
goto ldst_done;
|
|
|
|
|
|
|
|
case LOAD_FP:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
if (size == 4)
|
|
|
|
err = do_fp_load(op.reg, do_lfs, op.ea, size, regs);
|
|
|
|
else
|
|
|
|
err = do_fp_load(op.reg, do_lfd, op.ea, size, regs);
|
|
|
|
goto ldst_done;
|
|
|
|
|
|
|
|
#ifdef CONFIG_ALTIVEC
|
|
|
|
case LOAD_VMX:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
err = do_vec_load(op.reg, do_lvx, op.ea & ~0xfUL, regs);
|
|
|
|
goto ldst_done;
|
|
|
|
#endif
|
|
|
|
#ifdef CONFIG_VSX
|
|
|
|
case LOAD_VSX:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
err = do_vsx_load(op.reg, do_lxvd2x, op.ea, regs);
|
|
|
|
goto ldst_done;
|
|
|
|
#endif
|
|
|
|
case LOAD_MULTI:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
rd = op.reg;
|
2014-09-02 12:35:09 +08:00
|
|
|
for (i = 0; i < size; i += 4) {
|
|
|
|
nb = size - i;
|
|
|
|
if (nb > 4)
|
|
|
|
nb = 4;
|
|
|
|
err = read_mem(®s->gpr[rd], op.ea, nb, regs);
|
2014-09-02 12:35:07 +08:00
|
|
|
if (err)
|
|
|
|
return 0;
|
2014-09-02 12:35:09 +08:00
|
|
|
if (nb < 4) /* left-justify last bytes */
|
|
|
|
regs->gpr[rd] <<= 32 - 8 * nb;
|
2014-09-02 12:35:07 +08:00
|
|
|
op.ea += 4;
|
2014-09-02 12:35:09 +08:00
|
|
|
++rd;
|
|
|
|
}
|
2014-09-02 12:35:07 +08:00
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case STORE:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
if ((op.type & UPDATE) && size == sizeof(long) &&
|
|
|
|
op.reg == 1 && op.update_reg == 1 &&
|
|
|
|
!(regs->msr & MSR_PR) &&
|
|
|
|
op.ea >= regs->gpr[1] - STACK_INT_FRAME_SIZE) {
|
|
|
|
err = handle_stack_update(op.ea, regs);
|
|
|
|
goto ldst_done;
|
|
|
|
}
|
|
|
|
err = write_mem(op.val, op.ea, size, regs);
|
|
|
|
goto ldst_done;
|
|
|
|
|
|
|
|
case STORE_FP:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
if (size == 4)
|
|
|
|
err = do_fp_store(op.reg, do_stfs, op.ea, size, regs);
|
|
|
|
else
|
|
|
|
err = do_fp_store(op.reg, do_stfd, op.ea, size, regs);
|
|
|
|
goto ldst_done;
|
|
|
|
|
|
|
|
#ifdef CONFIG_ALTIVEC
|
|
|
|
case STORE_VMX:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
err = do_vec_store(op.reg, do_stvx, op.ea & ~0xfUL, regs);
|
|
|
|
goto ldst_done;
|
|
|
|
#endif
|
|
|
|
#ifdef CONFIG_VSX
|
|
|
|
case STORE_VSX:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
err = do_vsx_store(op.reg, do_stxvd2x, op.ea, regs);
|
|
|
|
goto ldst_done;
|
|
|
|
#endif
|
|
|
|
case STORE_MULTI:
|
|
|
|
if (regs->msr & MSR_LE)
|
|
|
|
return 0;
|
|
|
|
rd = op.reg;
|
2014-09-02 12:35:09 +08:00
|
|
|
for (i = 0; i < size; i += 4) {
|
|
|
|
val = regs->gpr[rd];
|
|
|
|
nb = size - i;
|
|
|
|
if (nb > 4)
|
|
|
|
nb = 4;
|
|
|
|
else
|
|
|
|
val >>= 32 - 8 * nb;
|
|
|
|
err = write_mem(val, op.ea, nb, regs);
|
2014-09-02 12:35:07 +08:00
|
|
|
if (err)
|
|
|
|
return 0;
|
|
|
|
op.ea += 4;
|
2014-09-02 12:35:09 +08:00
|
|
|
++rd;
|
|
|
|
}
|
2014-09-02 12:35:07 +08:00
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case MFMSR:
|
|
|
|
regs->gpr[op.reg] = regs->msr & MSR_MASK;
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
case MTMSR:
|
|
|
|
val = regs->gpr[op.reg];
|
|
|
|
if ((val & MSR_RI) == 0)
|
|
|
|
/* can't step mtmsr[d] that would clear MSR_RI */
|
|
|
|
return -1;
|
|
|
|
/* here op.val is the mask of bits to change */
|
|
|
|
regs->msr = (regs->msr & ~op.val) | (val & op.val);
|
|
|
|
goto instr_done;
|
|
|
|
|
|
|
|
#ifdef CONFIG_PPC64
|
|
|
|
case SYSCALL: /* sc */
|
|
|
|
/*
|
|
|
|
* N.B. this uses knowledge about how the syscall
|
|
|
|
* entry code works. If that is changed, this will
|
|
|
|
* need to be changed also.
|
|
|
|
*/
|
|
|
|
if (regs->gpr[0] == 0x1ebe &&
|
|
|
|
cpu_has_feature(CPU_FTR_REAL_LE)) {
|
|
|
|
regs->msr ^= MSR_LE;
|
|
|
|
goto instr_done;
|
|
|
|
}
|
|
|
|
regs->gpr[9] = regs->gpr[13];
|
|
|
|
regs->gpr[10] = MSR_KERNEL;
|
|
|
|
regs->gpr[11] = regs->nip + 4;
|
|
|
|
regs->gpr[12] = regs->msr & MSR_MASK;
|
|
|
|
regs->gpr[13] = (unsigned long) get_paca();
|
|
|
|
regs->nip = (unsigned long) &system_call_common;
|
|
|
|
regs->msr = MSR_KERNEL;
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
case RFI:
|
|
|
|
return -1;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
ldst_done:
|
|
|
|
if (err)
|
|
|
|
return 0;
|
|
|
|
if (op.type & UPDATE)
|
|
|
|
regs->gpr[op.update_reg] = op.ea;
|
|
|
|
|
|
|
|
instr_done:
|
|
|
|
regs->nip = truncate_if_32bit(regs->msr, regs->nip + 4);
|
|
|
|
return 1;
|
2005-09-26 14:04:21 +08:00
|
|
|
}
|