linux_dsm_epyc7002/arch/nios2/kernel/insnemu.S

593 lines
15 KiB
ArmAsm
Raw Normal View History

/*
* Copyright (C) 2003-2013 Altera Corporation
* All rights reserved.
*
* 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.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include <linux/linkage.h>
#include <asm/entry.h>
.set noat
.set nobreak
/*
* Explicitly allow the use of r1 (the assembler temporary register)
* within this code. This register is normally reserved for the use of
* the compiler.
*/
ENTRY(instruction_trap)
ldw r1, PT_R1(sp) // Restore registers
ldw r2, PT_R2(sp)
ldw r3, PT_R3(sp)
ldw r4, PT_R4(sp)
ldw r5, PT_R5(sp)
ldw r6, PT_R6(sp)
ldw r7, PT_R7(sp)
ldw r8, PT_R8(sp)
ldw r9, PT_R9(sp)
ldw r10, PT_R10(sp)
ldw r11, PT_R11(sp)
ldw r12, PT_R12(sp)
ldw r13, PT_R13(sp)
ldw r14, PT_R14(sp)
ldw r15, PT_R15(sp)
ldw ra, PT_RA(sp)
ldw fp, PT_FP(sp)
ldw gp, PT_GP(sp)
ldw et, PT_ESTATUS(sp)
wrctl estatus, et
ldw ea, PT_EA(sp)
ldw et, PT_SP(sp) /* backup sp in et */
addi sp, sp, PT_REGS_SIZE
/* INSTRUCTION EMULATION
* ---------------------
*
* Nios II processors generate exceptions for unimplemented instructions.
* The routines below emulate these instructions. Depending on the
* processor core, the only instructions that might need to be emulated
* are div, divu, mul, muli, mulxss, mulxsu, and mulxuu.
*
* The emulations match the instructions, except for the following
* limitations:
*
* 1) The emulation routines do not emulate the use of the exception
* temporary register (et) as a source operand because the exception
* handler already has modified it.
*
* 2) The routines do not emulate the use of the stack pointer (sp) or
* the exception return address register (ea) as a destination because
* modifying these registers crashes the exception handler or the
* interrupted routine.
*
* Detailed Design
* ---------------
*
* The emulation routines expect the contents of integer registers r0-r31
* to be on the stack at addresses sp, 4(sp), 8(sp), ... 124(sp). The
* routines retrieve source operands from the stack and modify the
* destination register's value on the stack prior to the end of the
* exception handler. Then all registers except the destination register
* are restored to their previous values.
*
* The instruction that causes the exception is found at address -4(ea).
* The instruction's OP and OPX fields identify the operation to be
* performed.
*
* One instruction, muli, is an I-type instruction that is identified by
* an OP field of 0x24.
*
* muli AAAAA,BBBBB,IIIIIIIIIIIIIIII,-0x24-
* 27 22 6 0 <-- LSB of field
*
* The remaining emulated instructions are R-type and have an OP field
* of 0x3a. Their OPX fields identify them.
*
* R-type AAAAA,BBBBB,CCCCC,XXXXXX,NNNNN,-0x3a-
* 27 22 17 11 6 0 <-- LSB of field
*
*
* Opcode Encoding. muli is identified by its OP value. Then OPX & 0x02
* is used to differentiate between the division opcodes and the
* remaining multiplication opcodes.
*
* Instruction OP OPX OPX & 0x02
* ----------- ---- ---- ----------
* muli 0x24
* divu 0x3a 0x24 0
* div 0x3a 0x25 0
* mul 0x3a 0x27 != 0
* mulxuu 0x3a 0x07 != 0
* mulxsu 0x3a 0x17 != 0
* mulxss 0x3a 0x1f != 0
*/
/*
* Save everything on the stack to make it easy for the emulation
* routines to retrieve the source register operands.
*/
addi sp, sp, -128
stw zero, 0(sp) /* Save zero on stack to avoid special case for r0. */
stw r1, 4(sp)
stw r2, 8(sp)
stw r3, 12(sp)
stw r4, 16(sp)
stw r5, 20(sp)
stw r6, 24(sp)
stw r7, 28(sp)
stw r8, 32(sp)
stw r9, 36(sp)
stw r10, 40(sp)
stw r11, 44(sp)
stw r12, 48(sp)
stw r13, 52(sp)
stw r14, 56(sp)
stw r15, 60(sp)
stw r16, 64(sp)
stw r17, 68(sp)
stw r18, 72(sp)
stw r19, 76(sp)
stw r20, 80(sp)
stw r21, 84(sp)
stw r22, 88(sp)
stw r23, 92(sp)
/* Don't bother to save et. It's already been changed. */
rdctl r5, estatus
stw r5, 100(sp)
stw gp, 104(sp)
stw et, 108(sp) /* et contains previous sp value. */
stw fp, 112(sp)
stw ea, 116(sp)
stw ra, 120(sp)
/*
* Split the instruction into its fields. We need 4*A, 4*B, and 4*C as
* offsets to the stack pointer for access to the stored register values.
*/
ldw r2,-4(ea) /* r2 = AAAAA,BBBBB,IIIIIIIIIIIIIIII,PPPPPP */
roli r3, r2, 7 /* r3 = BBB,IIIIIIIIIIIIIIII,PPPPPP,AAAAA,BB */
roli r4, r3, 3 /* r4 = IIIIIIIIIIIIIIII,PPPPPP,AAAAA,BBBBB */
roli r5, r4, 2 /* r5 = IIIIIIIIIIIIII,PPPPPP,AAAAA,BBBBB,II */
srai r4, r4, 16 /* r4 = (sign-extended) IMM16 */
roli r6, r5, 5 /* r6 = XXXX,NNNNN,PPPPPP,AAAAA,BBBBB,CCCCC,XX */
andi r2, r2, 0x3f /* r2 = 00000000000000000000000000,PPPPPP */
andi r3, r3, 0x7c /* r3 = 0000000000000000000000000,AAAAA,00 */
andi r5, r5, 0x7c /* r5 = 0000000000000000000000000,BBBBB,00 */
andi r6, r6, 0x7c /* r6 = 0000000000000000000000000,CCCCC,00 */
/* Now
* r2 = OP
* r3 = 4*A
* r4 = IMM16 (sign extended)
* r5 = 4*B
* r6 = 4*C
*/
/*
* Get the operands.
*
* It is necessary to check for muli because it uses an I-type
* instruction format, while the other instructions are have an R-type
* format.
*
* Prepare for either multiplication or division loop.
* They both loop 32 times.
*/
movi r14, 32
add r3, r3, sp /* r3 = address of A-operand. */
ldw r3, 0(r3) /* r3 = A-operand. */
movi r7, 0x24 /* muli opcode (I-type instruction format) */
beq r2, r7, mul_immed /* muli doesn't use the B register as a source */
add r5, r5, sp /* r5 = address of B-operand. */
ldw r5, 0(r5) /* r5 = B-operand. */
/* r4 = SSSSSSSSSSSSSSSS,-----IMM16------ */
/* IMM16 not needed, align OPX portion */
/* r4 = SSSSSSSSSSSSSSSS,CCCCC,-OPX--,00000 */
srli r4, r4, 5 /* r4 = 00000,SSSSSSSSSSSSSSSS,CCCCC,-OPX-- */
andi r4, r4, 0x3f /* r4 = 00000000000000000000000000,-OPX-- */
/* Now
* r2 = OP
* r3 = src1
* r5 = src2
* r4 = OPX (no longer can be muli)
* r6 = 4*C
*/
/*
* Multiply or Divide?
*/
andi r7, r4, 0x02 /* For R-type multiply instructions,
OPX & 0x02 != 0 */
bne r7, zero, multiply
/* DIVISION
*
* Divide an unsigned dividend by an unsigned divisor using
* a shift-and-subtract algorithm. The example below shows
* 43 div 7 = 6 for 8-bit integers. This classic algorithm uses a
* single register to store both the dividend and the quotient,
* allowing both values to be shifted with a single instruction.
*
* remainder dividend:quotient
* --------- -----------------
* initialize 00000000 00101011:
* shift 00000000 0101011:_
* remainder >= divisor? no 00000000 0101011:0
* shift 00000000 101011:0_
* remainder >= divisor? no 00000000 101011:00
* shift 00000001 01011:00_
* remainder >= divisor? no 00000001 01011:000
* shift 00000010 1011:000_
* remainder >= divisor? no 00000010 1011:0000
* shift 00000101 011:0000_
* remainder >= divisor? no 00000101 011:00000
* shift 00001010 11:00000_
* remainder >= divisor? yes 00001010 11:000001
* remainder -= divisor - 00000111
* ----------
* 00000011 11:000001
* shift 00000111 1:000001_
* remainder >= divisor? yes 00000111 1:0000011
* remainder -= divisor - 00000111
* ----------
* 00000000 1:0000011
* shift 00000001 :0000011_
* remainder >= divisor? no 00000001 :00000110
*
* The quotient is 00000110.
*/
divide:
/*
* Prepare for division by assuming the result
* is unsigned, and storing its "sign" as 0.
*/
movi r17, 0
/* Which division opcode? */
xori r7, r4, 0x25 /* OPX of div */
bne r7, zero, unsigned_division
/*
* OPX is div. Determine and store the sign of the quotient.
* Then take the absolute value of both operands.
*/
xor r17, r3, r5 /* MSB contains sign of quotient */
bge r3,zero,dividend_is_nonnegative
sub r3, zero, r3 /* -r3 */
dividend_is_nonnegative:
bge r5, zero, divisor_is_nonnegative
sub r5, zero, r5 /* -r5 */
divisor_is_nonnegative:
unsigned_division:
/* Initialize the unsigned-division loop. */
movi r13, 0 /* remainder = 0 */
/* Now
* r3 = dividend : quotient
* r4 = 0x25 for div, 0x24 for divu
* r5 = divisor
* r13 = remainder
* r14 = loop counter (already initialized to 32)
* r17 = MSB contains sign of quotient
*/
/*
* for (count = 32; count > 0; --count)
* {
*/
divide_loop:
/*
* Division:
*
* (remainder:dividend:quotient) <<= 1;
*/
slli r13, r13, 1
cmplt r7, r3, zero /* r7 = MSB of r3 */
or r13, r13, r7
slli r3, r3, 1
/*
* if (remainder >= divisor)
* {
* set LSB of quotient
* remainder -= divisor;
* }
*/
bltu r13, r5, div_skip
ori r3, r3, 1
sub r13, r13, r5
div_skip:
/*
* }
*/
subi r14, r14, 1
bne r14, zero, divide_loop
/* Now
* r3 = quotient
* r4 = 0x25 for div, 0x24 for divu
* r6 = 4*C
* r17 = MSB contains sign of quotient
*/
/*
* Conditionally negate signed quotient. If quotient is unsigned,
* the sign already is initialized to 0.
*/
bge r17, zero, quotient_is_nonnegative
sub r3, zero, r3 /* -r3 */
quotient_is_nonnegative:
/*
* Final quotient is in r3.
*/
add r6, r6, sp
stw r3, 0(r6) /* write quotient to stack */
br restore_registers
/* MULTIPLICATION
*
* A "product" is the number that one gets by summing a "multiplicand"
* several times. The "multiplier" specifies the number of copies of the
* multiplicand that are summed.
*
* Actual multiplication algorithms don't use repeated addition, however.
* Shift-and-add algorithms get the same answer as repeated addition, and
* they are faster. To compute the lower half of a product (pppp below)
* one shifts the product left before adding in each of the partial
* products (a * mmmm) through (d * mmmm).
*
* To compute the upper half of a product (PPPP below), one adds in the
* partial products (d * mmmm) through (a * mmmm), each time following
* the add by a right shift of the product.
*
* mmmm
* * abcd
* ------
* #### = d * mmmm
* #### = c * mmmm
* #### = b * mmmm
* #### = a * mmmm
* --------
* PPPPpppp
*
* The example above shows 4 partial products. Computing actual Nios II
* products requires 32 partials.
*
* It is possible to compute the result of mulxsu from the result of
* mulxuu because the only difference between the results of these two
* opcodes is the value of the partial product associated with the sign
* bit of rA.
*
* mulxsu = mulxuu - (rA < 0) ? rB : 0;
*
* It is possible to compute the result of mulxss from the result of
* mulxsu because the only difference between the results of these two
* opcodes is the value of the partial product associated with the sign
* bit of rB.
*
* mulxss = mulxsu - (rB < 0) ? rA : 0;
*
*/
mul_immed:
/* Opcode is muli. Change it into mul for remainder of algorithm. */
mov r6, r5 /* Field B is dest register, not field C. */
mov r5, r4 /* Field IMM16 is src2, not field B. */
movi r4, 0x27 /* OPX of mul is 0x27 */
multiply:
/* Initialize the multiplication loop. */
movi r9, 0 /* mul_product = 0 */
movi r10, 0 /* mulxuu_product = 0 */
mov r11, r5 /* save original multiplier for mulxsu and mulxss */
mov r12, r5 /* mulxuu_multiplier (will be shifted) */
movi r16, 1 /* used to create "rori B,A,1" from "ror B,A,r16" */
/* Now
* r3 = multiplicand
* r5 = mul_multiplier
* r6 = 4 * dest_register (used later as offset to sp)
* r7 = temp
* r9 = mul_product
* r10 = mulxuu_product
* r11 = original multiplier
* r12 = mulxuu_multiplier
* r14 = loop counter (already initialized)
* r16 = 1
*/
/*
* for (count = 32; count > 0; --count)
* {
*/
multiply_loop:
/*
* mul_product <<= 1;
* lsb = multiplier & 1;
*/
slli r9, r9, 1
andi r7, r12, 1
/*
* if (lsb == 1)
* {
* mulxuu_product += multiplicand;
* }
*/
beq r7, zero, mulx_skip
add r10, r10, r3
cmpltu r7, r10, r3 /* Save the carry from the MSB of mulxuu_product. */
ror r7, r7, r16 /* r7 = 0x80000000 on carry, or else 0x00000000 */
mulx_skip:
/*
* if (MSB of mul_multiplier == 1)
* {
* mul_product += multiplicand;
* }
*/
bge r5, zero, mul_skip
add r9, r9, r3
mul_skip:
/*
* mulxuu_product >>= 1; logical shift
* mul_multiplier <<= 1; done with MSB
* mulx_multiplier >>= 1; done with LSB
*/
srli r10, r10, 1
or r10, r10, r7 /* OR in the saved carry bit. */
slli r5, r5, 1
srli r12, r12, 1
/*
* }
*/
subi r14, r14, 1
bne r14, zero, multiply_loop
/*
* Multiply emulation loop done.
*/
/* Now
* r3 = multiplicand
* r4 = OPX
* r6 = 4 * dest_register (used later as offset to sp)
* r7 = temp
* r9 = mul_product
* r10 = mulxuu_product
* r11 = original multiplier
*/
/* Calculate address for result from 4 * dest_register */
add r6, r6, sp
/*
* Select/compute the result based on OPX.
*/
/* OPX == mul? Then store. */
xori r7, r4, 0x27
beq r7, zero, store_product
/* It's one of the mulx.. opcodes. Move over the result. */
mov r9, r10
/* OPX == mulxuu? Then store. */
xori r7, r4, 0x07
beq r7, zero, store_product
/* Compute mulxsu
*
* mulxsu = mulxuu - (rA < 0) ? rB : 0;
*/
bge r3, zero, mulxsu_skip
sub r9, r9, r11
mulxsu_skip:
/* OPX == mulxsu? Then store. */
xori r7, r4, 0x17
beq r7, zero, store_product
/* Compute mulxss
*
* mulxss = mulxsu - (rB < 0) ? rA : 0;
*/
bge r11,zero,mulxss_skip
sub r9, r9, r3
mulxss_skip:
/* At this point, assume that OPX is mulxss, so store*/
store_product:
stw r9, 0(r6)
restore_registers:
/* No need to restore r0. */
ldw r5, 100(sp)
wrctl estatus, r5
ldw r1, 4(sp)
ldw r2, 8(sp)
ldw r3, 12(sp)
ldw r4, 16(sp)
ldw r5, 20(sp)
ldw r6, 24(sp)
ldw r7, 28(sp)
ldw r8, 32(sp)
ldw r9, 36(sp)
ldw r10, 40(sp)
ldw r11, 44(sp)
ldw r12, 48(sp)
ldw r13, 52(sp)
ldw r14, 56(sp)
ldw r15, 60(sp)
ldw r16, 64(sp)
ldw r17, 68(sp)
ldw r18, 72(sp)
ldw r19, 76(sp)
ldw r20, 80(sp)
ldw r21, 84(sp)
ldw r22, 88(sp)
ldw r23, 92(sp)
/* Does not need to restore et */
ldw gp, 104(sp)
ldw fp, 112(sp)
ldw ea, 116(sp)
ldw ra, 120(sp)
ldw sp, 108(sp) /* last restore sp */
eret
.set at
.set break