arch/arm/nwfpe/entry.S

Source file repositories/reference/linux-study-clean/arch/arm/nwfpe/entry.S

File Facts

System
Linux kernel
Corpus path
arch/arm/nwfpe/entry.S
Extension
.S
Size
5971 bytes
Lines
191
Domain
Architecture Layer
Bucket
arch/arm
Inferred role
Architecture Layer: arch/arm
Status
atlas-only

Why This File Exists

CPU and platform-specific kernel glue: boot entry, traps, syscall entry, interrupts, page tables, context switch, and low-level barriers.

Dependency Surface

Detected Declarations

Annotated Snippet

NetWinder Floating Point Emulator
    (c) Rebel.COM, 1998
    (c) 1998, 1999 Philip Blundell

    Direct questions, comments to Scott Bambrough <scottb@netwinder.org>

*/
#include <linux/linkage.h>
#include <asm/assembler.h>
#include <asm/opcodes.h>

/* This is the kernel's entry point into the floating point emulator.
It is called from the kernel with code similar to this:

	sub	r4, r5, #4
	ldrt	r0, [r4]			@ r0  = instruction
	adrsvc	al, r9, ret_from_exception	@ r9  = normal FP return
	adrsvc	al, lr, fpundefinstr		@ lr  = undefined instr return

	get_current_task r10
	mov	r8, #1
	strb	r8, [r10, #TSK_USED_MATH]	@ set current->used_math
	add	r10, r10, #TSS_FPESAVE		@ r10 = workspace
	ldr	r4, .LC2
	ldr	pc, [r4]			@ Call FP emulator entry point

The kernel expects the emulator to return via one of two possible
points of return it passes to the emulator.  The emulator, if
successful in its emulation, jumps to ret_from_exception (passed in
r9) and the kernel takes care of returning control from the trap to
the user code.  If the emulator is unable to emulate the instruction,
it returns via _fpundefinstr (passed via lr) and the kernel halts the
user program with a core dump.

On entry to the emulator r10 points to an area of private FP workspace
reserved in the thread structure for this process.  This is where the
emulator saves its registers across calls.  The first word of this area
is used as a flag to detect the first time a process uses floating point,
so that the emulator startup cost can be avoided for tasks that don't
want it.

This routine does three things:

1) The kernel has created a struct pt_regs on the stack and saved the
user registers into it.  See /usr/include/asm/proc/ptrace.h for details.

2) It calls EmulateAll to emulate a floating point instruction.
EmulateAll returns 1 if the emulation was successful, or 0 if not.

3) If an instruction has been emulated successfully, it looks ahead at
the next instruction.  If it is a floating point instruction, it
executes the instruction, without returning to user space.  In this
way it repeatedly looks ahead and executes floating point instructions
until it encounters a non floating point instruction, at which time it
returns via _fpreturn.

This is done to reduce the effect of the trap overhead on each
floating point instructions.  GCC attempts to group floating point
instructions to allow the emulator to spread the cost of the trap over
several floating point instructions.  */

#include <asm/asm-offsets.h>

	.globl	nwfpe_enter
nwfpe_enter:
	mov	r4, lr			@ save the failure-return addresses
	mov	sl, sp			@ we access the registers via 'sl'

	ldr	r5, [sp, #S_PC]		@ get contents of PC;
	mov	r6, r0			@ save the opcode

Annotation

Implementation Notes