diff --git a/sys/pic32/_startup.S b/sys/pic32/_startup.S
index dddbdee..df81a95 100644
--- a/sys/pic32/_startup.S
+++ b/sys/pic32/_startup.S
@@ -1,8 +1,9 @@
 #
 # Startup code for Microchip PIC32 microcontrollers.
-# Using HID bootloader.
+# Based on the MIPS application note:
+# "Boot-CPS: Example Boot Code for MIPS® Cores".
 #
-# Copyright (C) 2010 Serge Vakulenko, <serge@vak.ru>
+# Copyright (C) 2010-2014 Serge Vakulenko, <serge@vak.ru>
 #
 # Permission to use, copy, modify, and distribute this software
 # and its documentation for any purpose and without fee is hereby
@@ -24,86 +25,350 @@
 #
 #include "machine/io.h"
 
-#define UBASE	0x7f008000	/* User space base address */
+#define UBASE   0x7f008000      /* User space base address */
 
-		.set	noreorder
-		.set	mips32r2
-		.set	nomips16
+/*
+ * MIPS Coprocessor 0 register numbers
+ */
+#define C0_INDEX        $0
+#define C0_ENTRYLO0     $2
+#define C0_ENTRYLO1     $3
+#define C0_PAGEMASK     $5
+#define C0_WIRED        $6
+#define C0_COUNT        $9
+#define C0_ENTRYHI      $10
+#define C0_COMPARE      $11
+#define C0_STATUS       $12
+#define C0_SRSCTL       $12,2
+#define C0_CAUSE        $13
+#define C0_EPC          $14
+#define C0_CONFIG       $16
+#define C0_CONFIG1      $16,1
+#define C0_CONFIG7      $16,7
+#define C0_WATCHLO      $18
+#define C0_WATCHHI      $19
+#define C0_DEBUG        $23
+#define C0_DEPC         $24
+#define C0_ITAGLO       $28
+#define C0_DTAGLO       $28,2
+#define C0_ERRPC        $30
 
-		.extern	u
-		.extern	u_end
-		.extern	u0
-		.extern	main
-		.extern exception
+/*
+ * MIPS Config1 register
+ */
+#define CFG1_MMUSSHIFT  25              // mmu size - 1
+#define CFG1_ISSHIFT    22              // icache lines 64<<n
+#define CFG1_ILSHIFT    19              // icache line size 2<<n
+#define CFG1_IASHIFT    16              // icache ways - 1
+#define CFG1_DSSHIFT    13              // dcache lines 64<<n
+#define CFG1_DLSHIFT    10              // dcache line size 2<<n
+#define CFG1_DASHIFT    7               // dcache ways - 1
+
+                .set    noreorder
+                .set    mips32r2
+                .set    nomips16
+
+                .extern u
+                .extern u_end
+                .extern u0
+                .extern main
+                .extern exception
 
 #---------------------------------------
 # Reset vector: main entry point
 #
-		.section .startup,"ax",@progbits
-		.org	0
-		.type	_reset_vector_, @function
-_reset_vector_:	.globl	_reset_vector_
+                .section .startup,"ax",@progbits
+                .org    0
+                .type   _reset_vector_, @function
+_reset_vector_: .globl  _reset_vector_
 
-		.set	noat
-		move    $1, $zero                       # Clear all regs
-		move    $2, $zero
-		move    $3, $zero
-		move    $4, $zero
-		move    $5, $zero
-		move    $6, $zero
-		move    $7, $zero
-		move    $8, $zero
-		move    $9, $zero
-		move    $10, $zero
-		move    $11, $zero
-		move    $12, $zero
-		move    $13, $zero
-		move    $14, $zero
-		move    $15, $zero
-		move    $16, $zero
-		move    $17, $zero
-		move    $18, $zero
-		move    $19, $zero
-		move    $20, $zero
-		move    $21, $zero
-		move    $22, $zero
-		move    $23, $zero
-		move    $24, $zero
-		move    $25, $zero
-		move    $26, $zero
-		move    $27, $zero
-		move    $28, $zero
-		move    $29, $zero
-		move    $30, $zero
-		move    $31, $zero
-		mtlo	$zero
-		mthi	$zero
-		.set	at
+                .set    noat
+                mtc0    $0, C0_COUNT            # Clear cp0 Count (Used to measure boot time.)
+#if 0
+                //
+                // Check for NMI exception.
+                //
+check_nmi:                                      # Check whether we are here due to a reset or NMI.
+                mfc0    $s1, C0_STATUS          # Read Status
+                ext     $s1, $s1, 19, 1         # extract NMI
+                beqz    $s1, init_gpr           # Branch if this is NOT an NMI exception.
+                nop
 
-		la	$sp, u_end - 16			# Stack at end of U area
+                # Call nmi_exception().
+                la      $sp, _stack-16          # Set up stack base.
+                la      $gp, _gp                # GP register value defined by linker script.
+                la      $s1, nmi_exception      # Call user-defined NMI handler.
+                jalr    $s1
+                nop
+#endif
+                //
+                // Set all GPRs of all register sets to predefined state.
+                //
+init_gpr:
+                li      $1, 0xdeadbeef          # 0xdeadbeef stands out, kseg2 mapped, odd.
+
+                # Determine how many shadow sets are implemented (in addition to the base register set.)
+                # the first time thru the loop it will initialize using $1 set above.
+                # At the bottom og the loop, 1 is  subtract from $30
+                # and loop back to next_shadow_set to start the next loop and the next lowest set number.
+                mfc0    $29, C0_SRSCTL          # read SRSCtl
+                ext     $30, $29, 26, 4         # extract HSS
+
+next_shadow_set:                        # set PSS to shadow set to be initialized
+                ins     $29, $30, 6, 4          # insert PSS
+                mtc0    $29, C0_SRSCTL          # write SRSCtl
+
+                wrpgpr  $1, $1
+                wrpgpr  $2, $1
+                wrpgpr  $3, $1
+                wrpgpr  $4, $1
+                wrpgpr  $5, $1
+                wrpgpr  $6, $1
+                wrpgpr  $7, $1
+                wrpgpr  $8, $1
+                wrpgpr  $9, $1
+                wrpgpr  $10, $1
+                wrpgpr  $11, $1
+                wrpgpr  $12, $1
+                wrpgpr  $13, $1
+                wrpgpr  $14, $1
+                wrpgpr  $15, $1
+                wrpgpr  $16, $1
+                wrpgpr  $17, $1
+                wrpgpr  $18, $1
+                wrpgpr  $19, $1
+                wrpgpr  $20, $1
+                wrpgpr  $21, $1
+                wrpgpr  $22, $1
+                wrpgpr  $23, $1
+                wrpgpr  $24, $1
+                wrpgpr  $25, $1
+                wrpgpr  $26, $1
+                wrpgpr  $27, $1
+                wrpgpr  $28, $1
+                beqz    $30, init_cp0
+                wrpgpr  $29, $1
+
+                wrpgpr  $30, $1
+                wrpgpr  $31, $1
+                b       next_shadow_set
+                add     $30, -1                 # Decrement to the next lower number
+
+                //
+                // Init CP0 Status, Count, Compare, Watch*, and Cause.
+                //
+init_cp0:
+                # Initialize Status
+                li      $v1, 0x00400404         # (M_StatusIM | M_StatusERL | M_StatusBEV)
+                mtc0    $v1, C0_STATUS          # write Status
+
+                # Initialize Watch registers if implemented.
+                mfc0    $v0, C0_CONFIG1         # read Config1
+                ext     $v1, $v0, 3, 1          # extract bit 3 WR (Watch registers implemented)
+                beq     $v1, $zero, done_wr
+                li      $v1, 0x7                # (M_WatchHiI | M_WatchHiR | M_WatchHiW)
+
+                # Clear Watch Status bits and disable watch exceptions
+                mtc0    $v1, C0_WATCHHI         # write WatchHi0
+                mfc0    $v0, C0_WATCHHI         # read WatchHi0
+                bgez    $v0, done_wr            # Check for bit 31 (sign bit) for more Watch registers
+                mtc0    $zero, C0_WATCHLO       # clear WatchLo0
+
+                mtc0    $v1, C0_WATCHHI, 1      # write WatchHi1
+                mfc0    $v0, C0_WATCHHI, 1      # read WatchHi1
+                bgez    $v0, done_wr            # Check for bit 31 (sign bit) for more Watch registers
+                mtc0    $zero, C0_WATCHLO, 1    # clear WatchLo1
+
+                mtc0    $v1, C0_WATCHHI, 2      # write WatchHi2
+                mfc0    $v0, C0_WATCHHI, 2      # read WatchHi2
+                bgez    $v0, done_wr            # Check for bit 31 (sign bit) for more Watch registers
+                mtc0    $zero, C0_WATCHLO, 2    # clear WatchLo2
+
+                mtc0    $v1, C0_WATCHHI, 3      # write WatchHi3
+                mfc0    $v0, C0_WATCHHI, 3      # read WatchHi3
+                bgez    $v0, done_wr            # Check for bit 31 (sign bit) for more Watch registers
+                mtc0    $zero, C0_WATCHLO, 3    # clear WatchLo3
+
+                mtc0    $v1, C0_WATCHHI, 4      # write WatchHi4
+                mfc0    $v0, C0_WATCHHI, 4      # read WatchHi4
+                bgez    $v0, done_wr            # Check for bit 31 (sign bit) for more Watch registers
+                mtc0    $zero, C0_WATCHLO, 4    # clear WatchLo4
+
+                mtc0    $v1, C0_WATCHHI, 5      # write WatchHi5
+                mfc0    $v0, C0_WATCHHI, 5      # read WatchHi5
+                bgez    $v0, done_wr            # Check for bit 31 (sign bit) for more Watch registers
+                mtc0    $zero, C0_WATCHLO, 5    # clear WatchLo5
+
+                mtc0    $v1, C0_WATCHHI, 6      # write WatchHi6
+                mfc0    $v0, C0_WATCHHI, 6      # read WatchHi6
+                bgez    $v0, done_wr            # Check for bit 31 (sign bit) for more Watch registers
+                mtc0    $zero, C0_WATCHLO, 6    # clear WatchLo6
+
+                mtc0    $v1, C0_WATCHHI, 7      # write WatchHi7
+                mtc0    $zero, C0_WATCHLO, 7    # clear WatchLo7
+
+done_wr:
+                # Clear WP bit to avoid watch exception upon user code entry, IV, and software interrupts.
+                mtc0    $zero, C0_CAUSE         # clear Cause: init AFTER init of WatchHi/Lo registers.
+
+                # Clear timer interrupt. (Count was cleared at the reset vector to allow timing boot.)
+                b       init_tlb                # Jump over exception vectors
+                mtc0    $zero, C0_COMPARE       # clear Compare
+
+                //
+                // Clear TLB: generate unique EntryHi contents per entry pair.
+                //
+init_tlb:
+                # Determine if we have a TLB
+                mfc0    $v1, C0_CONFIG          # read Config
+                ext     $v1, $v1, 7, 3          # extract MT field
+                li      $a3, 0x1                # load a 1 to check against
+                bne     $v1, $a3, init_icache
+
+                # Config1MMUSize == Number of TLB entries - 1
+                mfc0    $v0, C0_CONFIG1         # Config1
+                ext     $v1, $v0, CFG1_MMUSSHIFT, 6     # extract MMU Size
+                mtc0    $zero, C0_ENTRYLO0      # clear EntryLo0
+                mtc0    $zero, C0_ENTRYLO1      # clear EntryLo1
+                mtc0    $zero, C0_PAGEMASK      # clear PageMask
+                mtc0    $zero, C0_WIRED         # clear Wired
+                li      $a0, 0x80000000
+
+next_tlb_entry:
+                mtc0    $v1, C0_INDEX           # write Index
+                mtc0    $a0, C0_ENTRYHI         # write EntryHi
+                ehb
+                tlbwi
+                add     $a0, 2<<13              # Add 8K to the address to avoid TLB conflict with previous entry
+
+                bne     $v1, $zero, next_tlb_entry
+                add     $v1, -1
+
+                //
+                // Clear L1 instruction cache.
+                //
+init_icache:
+                # Determine how big the I-cache is
+                mfc0    $v0, C0_CONFIG1         # read Config1
+                ext     $v1, $v0, CFG1_ILSHIFT, 3       # extract I-cache line size
+                beq     $v1, $zero, done_icache # Skip ahead if no I-cache
+                nop
+
+                mfc0    $s1, C0_CONFIG7         # Read Config7
+                ext     $s1, $s1, 18, 1         # extract HCI
+                bnez    $s1, done_icache        # Skip when Hardware Cache Initialization bit set
+
+                li      $a2, 2
+                sllv    $v1, $a2, $v1           # Now have true I-cache line size in bytes
+
+                ext     $a0, $v0, CFG1_ISSHIFT, 3       # extract IS
+                li      $a2, 64
+                sllv    $a0, $a2, $a0           # I-cache sets per way
+
+                ext     $a1, $v0, CFG1_IASHIFT, 3       # extract I-cache Assoc - 1
+                add     $a1, 1
+                mul     $a0, $a0, $a1           # Total number of sets
+                lui     $a2, 0x8000             # Get a KSeg0 address for cacheops
+
+                mtc0    $zero, C0_ITAGLO        # Clear ITagLo register
+                move    $a3, $a0
+
+next_icache_tag:
+                # Index Store Tag Cache Op
+                # Will invalidate the tag entry, clear the lock bit, and clear the LRF bit
+                cache   0x8, 0($a2)             # ICIndexStTag
+                add     $a3, -1                 # Decrement set counter
+                bne     $a3, $zero, next_icache_tag
+                add     $a2, $v1                # Get next line address
+done_icache:
+
+                //
+                // Enable cacheability of kseg0 segment.
+                // Need to switch to kseg1, modify kseg0 CCA, then switch back.
+                //
+                la      $a2, enable_k0_cache
+                li      $a1, 0xf
+                ins     $a2, $a1, 29, 1         # changed to KSEG1 address by setting bit 29
+                jr      $a2
+                nop
+
+enable_k0_cache:
+                # Set CCA for kseg0 to cacheable.
+                # NOTE! This code must be executed in KSEG1 (not KSEG0 uncached)
+                mfc0    $v0, C0_CONFIG          # read C0_Config
+                li      $v1, 3                  # CCA for single-core processors
+                ins     $v0, $v1, 0, 3          # instert K0
+                mtc0    $v0, C0_CONFIG          # write C0_Config
+
+                la      $a2, init_dcache
+                jr      $a2                     # switch back to KSEG0
+                ehb
+
+                //
+                // Initialize the L1 data cache
+                //
+init_dcache:
+                mfc0    $v0, C0_CONFIG1         # read C0_Config1
+                ext     $v1, $v0, CFG1_DLSHIFT, 3       # extract D-cache line size
+                beq     $v1, $zero, done_dcache # Skip ahead if no D-cache
+                nop
+
+                mfc0    $s1, C0_CONFIG7         # Read Config7
+                ext     $s1, $s1, 18, 1         # extract HCI
+                bnez    $s1, done_dcache        # Skip when Hardware Cache Initialization bit set
+
+                li      $a2, 2
+                sllv    $v1, $a2, $v1           # Now have true D-cache line size in bytes
+
+                ext     $a0, $v0, CFG1_DSSHIFT, 3       # extract DS
+                li      $a2, 64
+                sllv    $a0, $a2, $a0           # D-cache sets per way
+
+                ext     $a1, $v0, CFG1_DASHIFT, 3       # extract D-cache Assoc - 1
+                add     $a1, 1
+                mul     $a0, $a0, $a1           # Get total number of sets
+                lui     $a2, 0x8000             # Get a KSeg0 address for cacheops
+
+                mtc0    $zero, C0_ITAGLO        # Clear ITagLo/DTagLo registers
+                mtc0    $zero, C0_DTAGLO
+                move    $a3, $a0
+
+next_dcache_tag:
+                # Index Store Tag Cache Op
+                # Will invalidate the tag entry, clear the lock bit, and clear the LRF bit
+                cache   0x9, 0($a2)             # DCIndexStTag
+                add     $a3, -1                 # Decrement set counter
+                bne     $a3, $zero, next_dcache_tag
+                add     $a2, $v1                # Get next line address
+done_dcache:
+
+                //
+                // Call main() function.
+                //
+                la      $sp, u_end - 16                 # Stack at end of U area
                 la      $a0, main
-		jalr	$a0				# Jump to main()
-		lui	$gp, 0x8000			# Set global pointer (delay slot)
+                jalr    $a0                             # Jump to main()
+                lui     $gp, 0x8000                     # Set global pointer (delay slot)
 
                 la      $k0, UBASE
-		mtc0	$k0, $C0_EPC                    # Entry to user code.
+                mtc0    $k0, C0_EPC                     # Entry to user code.
 
-		mfc0	$k0, $C0_STATUS
-		ori	$k0, $k0, ST_UM | ST_EXL | ST_IE # Set user mode and enable interrupts
-		mtc0	$k0, $C0_STATUS			# Put SR back
-		ehb
-		eret					# PC <= EPC; EXL <= 0
-		nop					# just to be safe
+                mfc0    $k0, C0_STATUS
+                ori     $k0, $k0, ST_UM | ST_EXL | ST_IE # Set user mode and enable interrupts
+                mtc0    $k0, C0_STATUS                  # Put SR back
+                ehb
+                eret                                    # PC <= EPC; EXL <= 0
+                nop                                     # just to be safe
 
 
 #---------------------------------------
 # Secondary entry point for RetroBSD bootloader.
 #
-		.section .exception,"ax",@progbits
+                .section .exception,"ax",@progbits
 _exception_base_: .globl _exception_base_
 
-		.org	0
-		.type	_entry_vector_, @function
+                .org    0
+                .type   _entry_vector_, @function
 _entry_vector_: .globl  _entry_vector_
                 la      $k0, _reset_vector_
                 jr      $k0
@@ -112,204 +377,204 @@ _entry_vector_: .globl  _entry_vector_
 #---------------------------------------
 # Exception vector: handle interrupts and exceptions
 #
-		.org	0x200
-		.type	_exception_vector_, @function
+                .org    0x200
+                .type   _exception_vector_, @function
 _exception_vector_: .globl _exception_vector_
 
-		mfc0	$k0, $C0_STATUS
-		andi	$k1, $k0, ST_UM			# Check user mode
-		beqz	$k1, kernel_exception
-		move	$k1, $sp
+                mfc0    $k0, C0_STATUS
+                andi    $k1, $k0, ST_UM                 # Check user mode
+                beqz    $k1, kernel_exception
+                move    $k1, $sp
 
-		#
-		# Exception in user mode: switch stack.
-		#
+                #
+                # Exception in user mode: switch stack.
+                #
 user_exception:
-		la	$sp, u_end                      # Stack at end of U area
+                la      $sp, u_end                      # Stack at end of U area
 kernel_exception:
-		addi    $sp, -16-FRAME_WORDS*4		# Allocate space for registers
+                addi    $sp, -16-FRAME_WORDS*4          # Allocate space for registers
 save_regs:
-		sw	$k0, (16+FRAME_STATUS*4) ($sp)
+                sw      $k0, (16+FRAME_STATUS*4) ($sp)
                 sw      $k1, (16+FRAME_SP*4) ($sp)
 
-		.set	noat
-		sw	$1, (16+FRAME_R1*4) ($sp)	# Save general registers
-		sw	$2, (16+FRAME_R2*4) ($sp)
-		sw	$3, (16+FRAME_R3*4) ($sp)
-		sw	$4, (16+FRAME_R4*4) ($sp)
-		sw	$5, (16+FRAME_R5*4) ($sp)
-		sw	$6, (16+FRAME_R6*4) ($sp)
-		sw	$7, (16+FRAME_R7*4) ($sp)
-		sw	$8, (16+FRAME_R8*4) ($sp)
-		sw	$9, (16+FRAME_R9*4) ($sp)
-		sw	$10, (16+FRAME_R10*4) ($sp)
-		sw	$11, (16+FRAME_R11*4) ($sp)
-		sw	$12, (16+FRAME_R12*4) ($sp)
-		sw	$13, (16+FRAME_R13*4) ($sp)
-		sw	$14, (16+FRAME_R14*4) ($sp)
-		sw	$15, (16+FRAME_R15*4) ($sp)
-		sw	$16, (16+FRAME_R16*4) ($sp)
-		sw	$17, (16+FRAME_R17*4) ($sp)
-		sw	$18, (16+FRAME_R18*4) ($sp)
-		sw	$19, (16+FRAME_R19*4) ($sp)
-		sw	$20, (16+FRAME_R20*4) ($sp)
-		sw	$21, (16+FRAME_R21*4) ($sp)
-		sw	$22, (16+FRAME_R22*4) ($sp)
-		sw	$23, (16+FRAME_R23*4) ($sp)
-		sw	$24, (16+FRAME_R24*4) ($sp)
-		sw	$25, (16+FRAME_R25*4) ($sp)
-		# Skip $26 - K0
-		# Skip $27 - K1
-		sw	$28, (16+FRAME_GP*4) ($sp)
-		# Skip $29 - SP
-		sw	$30, (16+FRAME_FP*4) ($sp)
-		sw	$31, (16+FRAME_RA*4) ($sp)
-		.set	at
+                .set    noat
+                sw      $1, (16+FRAME_R1*4) ($sp)       # Save general registers
+                sw      $2, (16+FRAME_R2*4) ($sp)
+                sw      $3, (16+FRAME_R3*4) ($sp)
+                sw      $4, (16+FRAME_R4*4) ($sp)
+                sw      $5, (16+FRAME_R5*4) ($sp)
+                sw      $6, (16+FRAME_R6*4) ($sp)
+                sw      $7, (16+FRAME_R7*4) ($sp)
+                sw      $8, (16+FRAME_R8*4) ($sp)
+                sw      $9, (16+FRAME_R9*4) ($sp)
+                sw      $10, (16+FRAME_R10*4) ($sp)
+                sw      $11, (16+FRAME_R11*4) ($sp)
+                sw      $12, (16+FRAME_R12*4) ($sp)
+                sw      $13, (16+FRAME_R13*4) ($sp)
+                sw      $14, (16+FRAME_R14*4) ($sp)
+                sw      $15, (16+FRAME_R15*4) ($sp)
+                sw      $16, (16+FRAME_R16*4) ($sp)
+                sw      $17, (16+FRAME_R17*4) ($sp)
+                sw      $18, (16+FRAME_R18*4) ($sp)
+                sw      $19, (16+FRAME_R19*4) ($sp)
+                sw      $20, (16+FRAME_R20*4) ($sp)
+                sw      $21, (16+FRAME_R21*4) ($sp)
+                sw      $22, (16+FRAME_R22*4) ($sp)
+                sw      $23, (16+FRAME_R23*4) ($sp)
+                sw      $24, (16+FRAME_R24*4) ($sp)
+                sw      $25, (16+FRAME_R25*4) ($sp)
+                # Skip $26 - K0
+                # Skip $27 - K1
+                sw      $28, (16+FRAME_GP*4) ($sp)
+                # Skip $29 - SP
+                sw      $30, (16+FRAME_FP*4) ($sp)
+                sw      $31, (16+FRAME_RA*4) ($sp)
+                .set    at
 
-		mfhi	$k0				# Save special registers
-		sw	$k0, (16+FRAME_HI*4) ($sp)
+                mfhi    $k0                             # Save special registers
+                sw      $k0, (16+FRAME_HI*4) ($sp)
 
-		mflo	$k0
-		sw	$k0, (16+FRAME_LO*4) ($sp)
+                mflo    $k0
+                sw      $k0, (16+FRAME_LO*4) ($sp)
 
-		mfc0	$k0, $C0_EPC
-		sw	$k0, (16+FRAME_PC*4) ($sp)
+                mfc0    $k0, C0_EPC
+                sw      $k0, (16+FRAME_PC*4) ($sp)
 
-		move	$a0, $sp
-		addi	$a0, 16				# Arg 0: saved regs.
-		jal	exception			# Call C code.
-		lui	$gp, 0x8000			# Set global pointer (delay slot)
+                move    $a0, $sp
+                addi    $a0, 16                         # Arg 0: saved regs.
+                jal     exception                       # Call C code.
+                lui     $gp, 0x8000                     # Set global pointer (delay slot)
 
-		#
-		# Restore CPU state and return from interrupt.
-		#
+                #
+                # Restore CPU state and return from interrupt.
+                #
 restore_regs:
-		lw	$a0, (16+FRAME_LO*4) ($sp)	# Load HI, LO registers
-		mtlo	$a0
-		lw	$a0, (16+FRAME_HI*4) ($sp)
-		mthi	$a0
+                lw      $a0, (16+FRAME_LO*4) ($sp)      # Load HI, LO registers
+                mtlo    $a0
+                lw      $a0, (16+FRAME_HI*4) ($sp)
+                mthi    $a0
 
-		.set	noat
-		lw	$1, (16+FRAME_R1*4) ($sp)	# Load general registers
-		lw	$2, (16+FRAME_R2*4) ($sp)
-		lw	$3, (16+FRAME_R3*4) ($sp)
-		lw	$4, (16+FRAME_R4*4) ($sp)
-		lw	$5, (16+FRAME_R5*4) ($sp)
-		lw	$6, (16+FRAME_R6*4) ($sp)
-		lw	$7, (16+FRAME_R7*4) ($sp)
-		lw	$8, (16+FRAME_R8*4) ($sp)
-		lw	$9, (16+FRAME_R9*4) ($sp)
-		lw	$10, (16+FRAME_R10*4) ($sp)
-		lw	$11, (16+FRAME_R11*4) ($sp)
-		lw	$12, (16+FRAME_R12*4) ($sp)
-		lw	$13, (16+FRAME_R13*4) ($sp)
-		lw	$14, (16+FRAME_R14*4) ($sp)
-		lw	$15, (16+FRAME_R15*4) ($sp)
-		lw	$16, (16+FRAME_R16*4) ($sp)
-		lw	$17, (16+FRAME_R17*4) ($sp)
-		lw	$18, (16+FRAME_R18*4) ($sp)
-		lw	$19, (16+FRAME_R19*4) ($sp)
-		lw	$20, (16+FRAME_R20*4) ($sp)
-		lw	$21, (16+FRAME_R21*4) ($sp)
-		lw	$22, (16+FRAME_R22*4) ($sp)
-		lw	$23, (16+FRAME_R23*4) ($sp)
-		lw	$24, (16+FRAME_R24*4) ($sp)
-		lw	$25, (16+FRAME_R25*4) ($sp)
-		# Skip $26 - K0
-		# Skip $27 - K1
-		lw	$28, (16+FRAME_GP*4) ($sp)
-		# Skip $29 - SP
-		lw	$30, (16+FRAME_FP*4) ($sp)
-		.set	at
+                .set    noat
+                lw      $1, (16+FRAME_R1*4) ($sp)       # Load general registers
+                lw      $2, (16+FRAME_R2*4) ($sp)
+                lw      $3, (16+FRAME_R3*4) ($sp)
+                lw      $4, (16+FRAME_R4*4) ($sp)
+                lw      $5, (16+FRAME_R5*4) ($sp)
+                lw      $6, (16+FRAME_R6*4) ($sp)
+                lw      $7, (16+FRAME_R7*4) ($sp)
+                lw      $8, (16+FRAME_R8*4) ($sp)
+                lw      $9, (16+FRAME_R9*4) ($sp)
+                lw      $10, (16+FRAME_R10*4) ($sp)
+                lw      $11, (16+FRAME_R11*4) ($sp)
+                lw      $12, (16+FRAME_R12*4) ($sp)
+                lw      $13, (16+FRAME_R13*4) ($sp)
+                lw      $14, (16+FRAME_R14*4) ($sp)
+                lw      $15, (16+FRAME_R15*4) ($sp)
+                lw      $16, (16+FRAME_R16*4) ($sp)
+                lw      $17, (16+FRAME_R17*4) ($sp)
+                lw      $18, (16+FRAME_R18*4) ($sp)
+                lw      $19, (16+FRAME_R19*4) ($sp)
+                lw      $20, (16+FRAME_R20*4) ($sp)
+                lw      $21, (16+FRAME_R21*4) ($sp)
+                lw      $22, (16+FRAME_R22*4) ($sp)
+                lw      $23, (16+FRAME_R23*4) ($sp)
+                lw      $24, (16+FRAME_R24*4) ($sp)
+                lw      $25, (16+FRAME_R25*4) ($sp)
+                # Skip $26 - K0
+                # Skip $27 - K1
+                lw      $28, (16+FRAME_GP*4) ($sp)
+                # Skip $29 - SP
+                lw      $30, (16+FRAME_FP*4) ($sp)
+                .set    at
 
                 # Do not use k0/k1 here, as interrupts are still enabled
-		lw	$31, (16+FRAME_STATUS*4) ($sp)	# K0 = saved status
-		ori	$31, ST_EXL                     # Set EXL
-		mtc0	$31, $C0_STATUS			# put SR back: disable interrupts
+                lw      $31, (16+FRAME_STATUS*4) ($sp)  # K0 = saved status
+                ori     $31, ST_EXL                     # Set EXL
+                mtc0    $31, C0_STATUS                  # put SR back: disable interrupts
                 ehb
 
-		lw	$k0, (16+FRAME_PC*4) ($sp)	# K0 = EPC
-		mtc0	$k0, $C0_EPC			# put PC in EPC
+                lw      $k0, (16+FRAME_PC*4) ($sp)      # K0 = EPC
+                mtc0    $k0, C0_EPC                     # put PC in EPC
                 ext     $k1, $31, 27, 1                 # get RP bit: single-step request
 
-		lw	$31, (16+FRAME_RA*4) ($sp)
-		lw	$sp, (16+FRAME_SP*4) ($sp)	# Restore stack
+                lw      $31, (16+FRAME_RA*4) ($sp)
+                lw      $sp, (16+FRAME_SP*4) ($sp)      # Restore stack
 
-		# Return from exception
+                # Return from exception
                 bnez    $k1, debug_request              # single-step request
-		ehb
-		eret					# PC <= EPC; EXL <= 0
+                ehb
+                eret                                    # PC <= EPC; EXL <= 0
 debug_request:
-		sdbbp					# enter debug mode
+                sdbbp                                   # enter debug mode
 
 #---------------------------------------
 # Debug exception processing.
 #
-		.org	0x480
-		.type	_debug_vector_, @function
+                .org    0x480
+                .type   _debug_vector_, @function
 _debug_vector_: .globl _debug_vector_
 
-		mfc0	$k0, $C0_DEPC
-		la	$k1, debug_request
-		bne     $k0, $k1, single_step_done
-		nop
+                mfc0    $k0, C0_DEPC
+                la      $k1, debug_request
+                bne     $k0, $k1, single_step_done
+                nop
 
                 # single step request
-		mfc0	$k0, $C0_DEBUG
-		ori     $k0, DB_SST                     # set SST bit
-		mtc0	$k0, $C0_DEBUG
+                mfc0    $k0, C0_DEBUG
+                ori     $k0, DB_SST                     # set SST bit
+                mtc0    $k0, C0_DEBUG
 
-		mfc0	$k1, $C0_EPC
-		mtc0	$k1, $C0_DEPC                   # DEPC <= EPC
-		mfc0	$k0, $C0_STATUS
-		xori	$k0, ST_EXL                     # Clear EXL
-		mtc0	$k0, $C0_STATUS
-		ehb
-		deret					# PC <= DEPC; DM <= 0
+                mfc0    $k1, C0_EPC
+                mtc0    $k1, C0_DEPC                    # DEPC <= EPC
+                mfc0    $k0, C0_STATUS
+                xori    $k0, ST_EXL                     # Clear EXL
+                mtc0    $k0, C0_STATUS
+                ehb
+                deret                                   # PC <= DEPC; DM <= 0
 
 single_step_done:
-		mtc0	$k0, $C0_EPC                    # EPC <= DEPC
+                mtc0    $k0, C0_EPC                     # EPC <= DEPC
 
-		la      $k1, _exception_vector_
-		mtc0	$k1, $C0_DEPC                   # DEPC <= exception handler
+                la      $k1, _exception_vector_
+                mtc0    $k1, C0_DEPC                    # DEPC <= exception handler
 
-		mfc0	$k0, $C0_DEBUG
-		sw	$k0, c0_debug                   # save Debug register
-		ori     $k0, DB_SST
-		xori    $k0, DB_SST                     # clear SST bit
-		mtc0	$k0, $C0_DEBUG
+                mfc0    $k0, C0_DEBUG
+                sw      $k0, c0_debug                   # save Debug register
+                ori     $k0, DB_SST
+                xori    $k0, DB_SST                     # clear SST bit
+                mtc0    $k0, C0_DEBUG
 
-		mfc0	$k1, $C0_STATUS
-		ori	$k1, ST_EXL                     # Set EXL
-		mtc0	$k1, $C0_STATUS
-		ehb
-		deret					# PC <= DEPC; DM <= 0
+                mfc0    $k1, C0_STATUS
+                ori     $k1, ST_EXL                     # Set EXL
+                mtc0    $k1, C0_STATUS
+                ehb
+                deret                                   # PC <= DEPC; DM <= 0
 
 #---------------------------------------
 # Icode is copied out to process 1 to exec /sbin/init.
 # If the exec fails, process 1 exits.
 #
-		.globl	icode, icodeend
-		.type	icode, @function
-		.type	icodeend, @function
+                .globl  icode, icodeend
+                .type   icode, @function
+                .type   icodeend, @function
 icode:
-		la	$a0, UBASE
-		move	$a1, $a0
-		addi	$a0, etcinit - icode
-		addi	$a1, argv - icode
-		syscall	11				# SYS_execv
-		move	$a0, $v0
-		syscall	1				# SYS_exit
+                la      $a0, UBASE
+                move    $a1, $a0
+                addi    $a0, etcinit - icode
+                addi    $a1, argv - icode
+                syscall 11                              # SYS_execv
+                move    $a0, $v0
+                syscall 1                               # SYS_exit
 etcinit:
-		.ascii	"/sbin/init\0"
+                .ascii  "/sbin/init\0"
 initflags:
-		.ascii	"-\0"                          # ASCII initflags
+                .ascii  "-\0"                          # ASCII initflags
 argv:
-		.word	etcinit + 6 - icode + UBASE	# address of "init\0"
-		.word	initflags - icode + UBASE	# init options
-		.word	0
+                .word   etcinit + 6 - icode + UBASE     # address of "init\0"
+                .word   initflags - icode + UBASE       # init options
+                .word   0
 
-icodeend:	nop
+icodeend:       nop
 
 #---------------------------------------
 # int setjmp (label_t *env);
@@ -317,22 +582,22 @@ icodeend:	nop
 # Setjmp(env) will save the process' current register variables, stack
 # and program counter context and return a zero.
 #
-		.type	setjmp, @function
-setjmp:		.globl	setjmp
-		sw	$s0, (0 * 4) ($a0)		# save register variables s0-s8
-		sw	$s1, (1 * 4) ($a0)
-		sw	$s2, (2 * 4) ($a0)
-		sw	$s3, (3 * 4) ($a0)
-		sw	$s4, (4 * 4) ($a0)
-		sw	$s5, (5 * 4) ($a0)
-		sw	$s6, (6 * 4) ($a0)
-		sw	$s7, (7 * 4) ($a0)
-		sw	$s8, (8 * 4) ($a0)		# frame pointer
-		sw	$ra, (9 * 4) ($a0)		# return address
-		sw	$gp, (10 * 4) ($a0)		# global data pointer
-		sw	$sp, (11 * 4) ($a0)		# stack pointer
-		j	$ra
-		move	$v0, $zero			# return a zero for the setjmp call
+                .type   setjmp, @function
+setjmp:         .globl  setjmp
+                sw      $s0, (0 * 4) ($a0)              # save register variables s0-s8
+                sw      $s1, (1 * 4) ($a0)
+                sw      $s2, (2 * 4) ($a0)
+                sw      $s3, (3 * 4) ($a0)
+                sw      $s4, (4 * 4) ($a0)
+                sw      $s5, (5 * 4) ($a0)
+                sw      $s6, (6 * 4) ($a0)
+                sw      $s7, (7 * 4) ($a0)
+                sw      $s8, (8 * 4) ($a0)              # frame pointer
+                sw      $ra, (9 * 4) ($a0)              # return address
+                sw      $gp, (10 * 4) ($a0)             # global data pointer
+                sw      $sp, (11 * 4) ($a0)             # stack pointer
+                j       $ra
+                move    $v0, $zero                      # return a zero for the setjmp call
 
 #---------------------------------------
 # void longjmp (memaddr uaddr, label_t *env);
@@ -346,60 +611,60 @@ setjmp:		.globl	setjmp
 # This longjmp differs from the longjmp found in the standard library -
 # it's actually closer to the resume routine of the 4.3BSD kernel.
 #
-		.type	longjmp, @function
-longjmp:	.globl	longjmp
+                .type   longjmp, @function
+longjmp:        .globl  longjmp
 
-		di				# can't let anything in till we get a valid stack...
+                di                              # can't let anything in till we get a valid stack...
 
-		la	$v0, u			# pointer to &u
-		beq	$v0, $a0, 2f		# if uaddr == &u...
-		nop				# ...no need to remap U area
+                la      $v0, u                  # pointer to &u
+                beq     $v0, $a0, 2f            # if uaddr == &u...
+                nop                             # ...no need to remap U area
 
-		la	$a3, u_end		# pointer to &u + USIZE
-		la	$a0, u0			# pointer to &u0
+                la      $a3, u_end              # pointer to &u + USIZE
+                la      $a0, u0                 # pointer to &u0
 
-		lw	$v1, 0($v0)		# u.u_procp
-		sw	$a0, 60($v1)		# u.u_procp->p_addr = &u0
+                lw      $v1, 0($v0)             # u.u_procp
+                sw      $a0, 60($v1)            # u.u_procp->p_addr = &u0
 
-		# exchange contents of u and u0
-		move	$v1, $v0
+                # exchange contents of u and u0
+                move    $v1, $v0
 1:
-                lw	$t1, 0($v1)
-                lw	$t0, 0($a0)
-                sw	$t0, 0($v1)
-                sw	$t1, 0($a0)
-                lw	$t1, 4($v1)
-                lw	$t0, 4($a0)
-                sw	$t0, 4($v1)
-                sw	$t1, 4($a0)
-                lw	$t1, 8($v1)
-                lw	$t0, 8($a0)
-                sw	$t0, 8($v1)
-                sw	$t1, 8($a0)
-                lw	$t1, 12($v1)
-                lw	$t0, 12($a0)
-                sw	$t0, 12($v1)
-                sw	$t1, 12($a0)
-                addiu	$v1, $v1, 16
-                bne	$a3, $v1, 1b
-                addiu	$a0, $a0, 16
+                lw      $t1, 0($v1)
+                lw      $t0, 0($a0)
+                sw      $t0, 0($v1)
+                sw      $t1, 0($a0)
+                lw      $t1, 4($v1)
+                lw      $t0, 4($a0)
+                sw      $t0, 4($v1)
+                sw      $t1, 4($a0)
+                lw      $t1, 8($v1)
+                lw      $t0, 8($a0)
+                sw      $t0, 8($v1)
+                sw      $t1, 8($a0)
+                lw      $t1, 12($v1)
+                lw      $t0, 12($a0)
+                sw      $t0, 12($v1)
+                sw      $t1, 12($a0)
+                addiu   $v1, $v1, 16
+                bne     $a3, $v1, 1b
+                addiu   $a0, $a0, 16
 
-		lw	$v1, 0($v0)		# u.u_procp
-		sw	$v0, 60($v1)		# u.u_procp->p_addr = &u
+                lw      $v1, 0($v0)             # u.u_procp
+                sw      $v0, 60($v1)            # u.u_procp->p_addr = &u
 2:
-		lw	$s0, (0 * 4) ($a1)	# restore register variables s0-s8
-		lw	$s1, (1 * 4) ($a1)
-		lw	$s2, (2 * 4) ($a1)
-		lw	$s3, (3 * 4) ($a1)
-		lw	$s4, (4 * 4) ($a1)
-		lw	$s5, (5 * 4) ($a1)
-		lw	$s6, (6 * 4) ($a1)
-		lw	$s7, (7 * 4) ($a1)
-		lw	$s8, (8 * 4) ($a1)	# frame pointer
-		lw	$ra, (9 * 4) ($a1)	# return address
-		lw	$gp, (10 * 4) ($a1)	# global data pointer
-		lw	$sp, (11 * 4) ($a1)	# stack pointer
+                lw      $s0, (0 * 4) ($a1)      # restore register variables s0-s8
+                lw      $s1, (1 * 4) ($a1)
+                lw      $s2, (2 * 4) ($a1)
+                lw      $s3, (3 * 4) ($a1)
+                lw      $s4, (4 * 4) ($a1)
+                lw      $s5, (5 * 4) ($a1)
+                lw      $s6, (6 * 4) ($a1)
+                lw      $s7, (7 * 4) ($a1)
+                lw      $s8, (8 * 4) ($a1)      # frame pointer
+                lw      $ra, (9 * 4) ($a1)      # return address
+                lw      $gp, (10 * 4) ($a1)     # global data pointer
+                lw      $sp, (11 * 4) ($a1)     # stack pointer
 
-		ei				# release interrupts
-		j	$ra			# transfer back to setjmp()
-		li	$v0, 1			# return value of 1
+                ei                              # release interrupts
+                j       $ra                     # transfer back to setjmp()
+                li      $v0, 1                  # return value of 1
diff --git a/sys/pic32/pic32mx.h b/sys/pic32/pic32mx.h
index cc79e7a..4e69fc8 100644
--- a/sys/pic32/pic32mx.h
+++ b/sys/pic32/pic32mx.h
@@ -27,6 +27,7 @@
 /*--------------------------------------
  * Coprocessor 0 registers.
  */
+#ifndef __ASSEMBLER__
 #define C0_HWRENA	7	/* Enable RDHWR in non-privileged mode */
 #define C0_BADVADDR	8	/* Virtual address of last exception */
 #define C0_COUNT	9	/* Processor cycle count */
@@ -47,6 +48,7 @@
 #define C0_DEPC		24	/* Program counter at last debug exception */
 #define C0_ERROREPC	30	/* Program counter at last error */
 #define C0_DESAVE	31	/* Debug handler scratchpad register */
+#endif
 
 /*
  * Status register.