Merge git://git.kernel.org/pub/scm/linux/kernel/git/cmetcalf/linux-tile

* git://git.kernel.org/pub/scm/linux/kernel/git/cmetcalf/linux-tile:
  arch/tile: convert a BUG_ON to BUILD_BUG_ON
  arch/tile: make ptrace() work properly for TILE-Gx COMPAT mode
  arch/tile: support new info op generated by compiler
  arch/tile: minor whitespace/naming changes for string support files
  arch/tile: enable single-step support for TILE-Gx
  arch/tile: parameterize system PLs to support KVM port
  arch/tile: add Tilera's <arch/sim.h> header as an open-source header
  arch/tile: Bomb C99 comments to C89 comments in tile's <arch/sim_def.h>
  arch/tile: prevent corrupt top frame from causing backtracer runaway
  arch/tile: various top-level Makefile cleanups
  arch/tile: change lower bound on syscall error return to -4095
  arch/tile: properly export __mb_incoherent for modules
  arch/tile: provide a definition of MAP_STACK
  kmemleak: add TILE to the list of supported architectures.
  char: hvc: check for error case
  arch/tile: Add a warning if we try to allocate too much vmalloc memory.
  arch/tile: update some comments to clarify register usage.
  arch/tile: use better "punctuation" for VMSPLIT_3_5G and friends
  arch/tile: Use <asm-generic/syscalls.h>
  tile: replace some BUG_ON checks with BUILD_BUG_ON checks

15 files changed, 1128 insertions, 411 deletions

diff --git a/arch/tile/include/arch/sim.h b/arch/tile/include/arch/sim.h
new file mode 100644
index 000000000000..74b7c1624d34
--- /dev/null
+++ b/arch/tile/include/arch/sim.h
@@ -0,0 +1,619 @@
+/*
+ * Copyright 2010 Tilera 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, version 2.
+ *
+ *   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, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file
+ *
+ * Provides an API for controlling the simulator at runtime.
+ */
+
+/**
+ * @addtogroup arch_sim
+ * @{
+ *
+ * An API for controlling the simulator at runtime.
+ *
+ * The simulator's behavior can be modified while it is running.
+ * For example, human-readable trace output can be enabled and disabled
+ * around code of interest.
+ *
+ * There are two ways to modify simulator behavior:
+ * programmatically, by calling various sim_* functions, and
+ * interactively, by entering commands like "sim set functional true"
+ * at the tile-monitor prompt.  Typing "sim help" at that prompt provides
+ * a list of interactive commands.
+ *
+ * All interactive commands can also be executed programmatically by
+ * passing a string to the sim_command function.
+ */
+
+#ifndef __ARCH_SIM_H__
+#define __ARCH_SIM_H__
+
+#include <arch/sim_def.h>
+#include <arch/abi.h>
+
+#ifndef __ASSEMBLER__
+
+#include <arch/spr_def.h>
+
+
+/**
+ * Return true if the current program is running under a simulator,
+ * rather than on real hardware.  If running on hardware, other "sim_xxx()"
+ * calls have no useful effect.
+ */
+static inline int
+sim_is_simulator(void)
+{
+  return __insn_mfspr(SPR_SIM_CONTROL) != 0;
+}
+
+
+/**
+ * Checkpoint the simulator state to a checkpoint file.
+ *
+ * The checkpoint file name is either the default or the name specified
+ * on the command line with "--checkpoint-file".
+ */
+static __inline void
+sim_checkpoint(void)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_CHECKPOINT);
+}
+
+
+/**
+ * Report whether or not various kinds of simulator tracing are enabled.
+ *
+ * @return The bitwise OR of these values:
+ *
+ * SIM_TRACE_CYCLES (--trace-cycles),
+ * SIM_TRACE_ROUTER (--trace-router),
+ * SIM_TRACE_REGISTER_WRITES (--trace-register-writes),
+ * SIM_TRACE_DISASM (--trace-disasm),
+ * SIM_TRACE_STALL_INFO (--trace-stall-info)
+ * SIM_TRACE_MEMORY_CONTROLLER (--trace-memory-controller)
+ * SIM_TRACE_L2_CACHE (--trace-l2)
+ * SIM_TRACE_LINES (--trace-lines)
+ */
+static __inline unsigned int
+sim_get_tracing(void)
+{
+  return __insn_mfspr(SPR_SIM_CONTROL) & SIM_TRACE_FLAG_MASK;
+}
+
+
+/**
+ * Turn on or off different kinds of simulator tracing.
+ *
+ * @param mask Either one of these special values:
+ *
+ * SIM_TRACE_NONE (turns off tracing),
+ * SIM_TRACE_ALL (turns on all possible tracing).
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_TRACE_CYCLES (--trace-cycles),
+ * SIM_TRACE_ROUTER (--trace-router),
+ * SIM_TRACE_REGISTER_WRITES (--trace-register-writes),
+ * SIM_TRACE_DISASM (--trace-disasm),
+ * SIM_TRACE_STALL_INFO (--trace-stall-info)
+ * SIM_TRACE_MEMORY_CONTROLLER (--trace-memory-controller)
+ * SIM_TRACE_L2_CACHE (--trace-l2)
+ * SIM_TRACE_LINES (--trace-lines)
+ */
+static __inline void
+sim_set_tracing(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_TRACE_SPR_ARG(mask));
+}
+
+
+/**
+ * Request dumping of different kinds of simulator state.
+ *
+ * @param mask Either this special value:
+ *
+ * SIM_DUMP_ALL (dump all known state)
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_DUMP_REGS (the register file),
+ * SIM_DUMP_SPRS (the SPRs),
+ * SIM_DUMP_ITLB (the iTLB),
+ * SIM_DUMP_DTLB (the dTLB),
+ * SIM_DUMP_L1I (the L1 I-cache),
+ * SIM_DUMP_L1D (the L1 D-cache),
+ * SIM_DUMP_L2 (the L2 cache),
+ * SIM_DUMP_SNREGS (the switch register file),
+ * SIM_DUMP_SNITLB (the switch iTLB),
+ * SIM_DUMP_SNL1I (the switch L1 I-cache),
+ * SIM_DUMP_BACKTRACE (the current backtrace)
+ */
+static __inline void
+sim_dump(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_DUMP_SPR_ARG(mask));
+}
+
+
+/**
+ * Print a string to the simulator stdout.
+ *
+ * @param str The string to be written; a newline is automatically added.
+ */
+static __inline void
+sim_print_string(const char* str)
+{
+  int i;
+  for (i = 0; str[i] != 0; i++)
+  {
+    __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
+                 (str[i] << _SIM_CONTROL_OPERATOR_BITS));
+  }
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
+               (SIM_PUTC_FLUSH_STRING << _SIM_CONTROL_OPERATOR_BITS));
+}
+
+
+/**
+ * Execute a simulator command string.
+ *
+ * Type 'sim help' at the tile-monitor prompt to learn what commands
+ * are available.  Note the use of the tile-monitor "sim" command to
+ * pass commands to the simulator.
+ *
+ * The argument to sim_command() does not include the leading "sim"
+ * prefix used at the tile-monitor prompt; for example, you might call
+ * sim_command("trace disasm").
+ */
+static __inline void
+sim_command(const char* str)
+{
+  int c;
+  do
+  {
+    c = *str++;
+    __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_COMMAND |
+                 (c << _SIM_CONTROL_OPERATOR_BITS));
+  }
+  while (c);
+}
+
+
+
+#ifndef __DOXYGEN__
+
+/**
+ * The underlying implementation of "_sim_syscall()".
+ *
+ * We use extra "and" instructions to ensure that all the values
+ * we are passing to the simulator are actually valid in the registers
+ * (i.e. returned from memory) prior to the SIM_CONTROL spr.
+ */
+static __inline int _sim_syscall0(int val)
+{
+  long result;
+  __asm__ __volatile__ ("mtspr SIM_CONTROL, r0"
+                        : "=R00" (result) : "R00" (val));
+  return result;
+}
+
+static __inline int _sim_syscall1(int val, long arg1)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r1, r1; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result) : "R00" (val), "R01" (arg1));
+  return result;
+}
+
+static __inline int _sim_syscall2(int val, long arg1, long arg2)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r1, r2; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result)
+                        : "R00" (val), "R01" (arg1), "R02" (arg2));
+  return result;
+}
+
+/* Note that _sim_syscall3() and higher are technically at risk of
+   receiving an interrupt right before the mtspr bundle, in which case
+   the register values for arguments 3 and up may still be in flight
+   to the core from a stack frame reload. */
+
+static __inline int _sim_syscall3(int val, long arg1, long arg2, long arg3)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r3, r3 };"
+                        "{ and zero, r1, r2; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result)
+                        : "R00" (val), "R01" (arg1), "R02" (arg2),
+                          "R03" (arg3));
+  return result;
+}
+
+static __inline int _sim_syscall4(int val, long arg1, long arg2, long arg3,
+                                  long arg4)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r3, r4 };"
+                        "{ and zero, r1, r2; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result)
+                        : "R00" (val), "R01" (arg1), "R02" (arg2),
+                          "R03" (arg3), "R04" (arg4));
+  return result;
+}
+
+static __inline int _sim_syscall5(int val, long arg1, long arg2, long arg3,
+                                  long arg4, long arg5)
+{
+  long result;
+  __asm__ __volatile__ ("{ and zero, r3, r4; and zero, r5, r5 };"
+                        "{ and zero, r1, r2; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (result)
+                        : "R00" (val), "R01" (arg1), "R02" (arg2),
+                          "R03" (arg3), "R04" (arg4), "R05" (arg5));
+  return result;
+}
+
+
+/**
+ * Make a special syscall to the simulator itself, if running under
+ * simulation. This is used as the implementation of other functions
+ * and should not be used outside this file.
+ *
+ * @param syscall_num The simulator syscall number.
+ * @param nr The number of additional arguments provided.
+ *
+ * @return Varies by syscall.
+ */
+#define _sim_syscall(syscall_num, nr, args...) \
+  _sim_syscall##nr( \
+    ((syscall_num) << _SIM_CONTROL_OPERATOR_BITS) | SIM_CONTROL_SYSCALL, args)
+
+
+/* Values for the "access_mask" parameters below. */
+#define SIM_WATCHPOINT_READ    1
+#define SIM_WATCHPOINT_WRITE   2
+#define SIM_WATCHPOINT_EXECUTE 4
+
+
+static __inline int
+sim_add_watchpoint(unsigned int process_id,
+                   unsigned long address,
+                   unsigned long size,
+                   unsigned int access_mask,
+                   unsigned long user_data)
+{
+  return _sim_syscall(SIM_SYSCALL_ADD_WATCHPOINT, 5, process_id,
+                     address, size, access_mask, user_data);
+}
+
+
+static __inline int
+sim_remove_watchpoint(unsigned int process_id,
+                      unsigned long address,
+                      unsigned long size,
+                      unsigned int access_mask,
+                      unsigned long user_data)
+{
+  return _sim_syscall(SIM_SYSCALL_REMOVE_WATCHPOINT, 5, process_id,
+                     address, size, access_mask, user_data);
+}
+
+
+/**
+ * Return value from sim_query_watchpoint.
+ */
+struct SimQueryWatchpointStatus
+{
+  /**
+   * 0 if a watchpoint fired, 1 if no watchpoint fired, or -1 for
+   * error (meaning a bad process_id).
+   */
+  int syscall_status;
+
+  /**
+   * The address of the watchpoint that fired (this is the address
+   * passed to sim_add_watchpoint, not an address within that range
+   * that actually triggered the watchpoint).
+   */
+  unsigned long address;
+
+  /** The arbitrary user_data installed by sim_add_watchpoint. */
+  unsigned long user_data;
+};
+
+
+static __inline struct SimQueryWatchpointStatus
+sim_query_watchpoint(unsigned int process_id)
+{
+  struct SimQueryWatchpointStatus status;
+  long val = SIM_CONTROL_SYSCALL |
+    (SIM_SYSCALL_QUERY_WATCHPOINT << _SIM_CONTROL_OPERATOR_BITS);
+  __asm__ __volatile__ ("{ and zero, r1, r1; mtspr SIM_CONTROL, r0 }"
+                        : "=R00" (status.syscall_status),
+                          "=R01" (status.address),
+                          "=R02" (status.user_data)
+                        : "R00" (val), "R01" (process_id));
+  return status;
+}
+
+
+/* On the simulator, confirm lines have been evicted everywhere. */
+static __inline void
+sim_validate_lines_evicted(unsigned long long pa, unsigned long length)
+{
+#ifdef __LP64__
+  _sim_syscall(SIM_SYSCALL_VALIDATE_LINES_EVICTED, 2, pa, length);
+#else
+  _sim_syscall(SIM_SYSCALL_VALIDATE_LINES_EVICTED, 4,
+               0 /* dummy */, (long)(pa), (long)(pa >> 32), length);
+#endif
+}
+
+
+#endif /* !__DOXYGEN__ */
+
+
+
+
+/**
+ * Modify the shaping parameters of a shim.
+ *
+ * @param shim The shim to modify. One of:
+ *   SIM_CONTROL_SHAPING_GBE_0
+ *   SIM_CONTROL_SHAPING_GBE_1
+ *   SIM_CONTROL_SHAPING_GBE_2
+ *   SIM_CONTROL_SHAPING_GBE_3
+ *   SIM_CONTROL_SHAPING_XGBE_0
+ *   SIM_CONTROL_SHAPING_XGBE_1
+ *
+ * @param type The type of shaping. This should be the same type of
+ * shaping that is already in place on the shim. One of:
+ *   SIM_CONTROL_SHAPING_MULTIPLIER
+ *   SIM_CONTROL_SHAPING_PPS
+ *   SIM_CONTROL_SHAPING_BPS
+ *
+ * @param units The magnitude of the rate. One of:
+ *   SIM_CONTROL_SHAPING_UNITS_SINGLE
+ *   SIM_CONTROL_SHAPING_UNITS_KILO
+ *   SIM_CONTROL_SHAPING_UNITS_MEGA
+ *   SIM_CONTROL_SHAPING_UNITS_GIGA
+ *
+ * @param rate The rate to which to change it. This must fit in
+ * SIM_CONTROL_SHAPING_RATE_BITS bits or a warning is issued and
+ * the shaping is not changed.
+ *
+ * @return 0 if no problems were detected in the arguments to sim_set_shaping
+ * or 1 if problems were detected (for example, rate does not fit in 17 bits).
+ */
+static __inline int
+sim_set_shaping(unsigned shim,
+                unsigned type,
+                unsigned units,
+                unsigned rate)
+{
+  if ((rate & ~((1 << SIM_CONTROL_SHAPING_RATE_BITS) - 1)) != 0)
+    return 1;
+
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_SHAPING_SPR_ARG(shim, type, units, rate));
+  return 0;
+}
+
+#ifdef __tilegx__
+
+/** Enable a set of mPIPE links.  Pass a -1 link_mask to enable all links. */
+static __inline void
+sim_enable_mpipe_links(unsigned mpipe, unsigned long link_mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL,
+               (SIM_CONTROL_ENABLE_MPIPE_LINK_MAGIC_BYTE |
+                (mpipe << 8) | (1 << 16) | ((uint_reg_t)link_mask << 32)));
+}
+
+/** Disable a set of mPIPE links.  Pass a -1 link_mask to disable all links. */
+static __inline void
+sim_disable_mpipe_links(unsigned mpipe, unsigned long link_mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL,
+               (SIM_CONTROL_ENABLE_MPIPE_LINK_MAGIC_BYTE |
+                (mpipe << 8) | (0 << 16) | ((uint_reg_t)link_mask << 32)));
+}
+
+#endif /* __tilegx__ */
+
+
+/*
+ * An API for changing "functional" mode.
+ */
+
+#ifndef __DOXYGEN__
+
+#define sim_enable_functional() \
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_ENABLE_FUNCTIONAL)
+
+#define sim_disable_functional() \
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_DISABLE_FUNCTIONAL)
+
+#endif /* __DOXYGEN__ */
+
+
+/*
+ * Profiler support.
+ */
+
+/**
+ * Turn profiling on for the current task.
+ *
+ * Note that this has no effect if run in an environment without
+ * profiling support (thus, the proper flags to the simulator must
+ * be supplied).
+ */
+static __inline void
+sim_profiler_enable(void)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PROFILER_ENABLE);
+}
+
+
+/** Turn profiling off for the current task. */
+static __inline void
+sim_profiler_disable(void)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PROFILER_DISABLE);
+}
+
+
+/**
+ * Turn profiling on or off for the current task.
+ *
+ * @param enabled If true, turns on profiling. If false, turns it off.
+ *
+ * Note that this has no effect if run in an environment without
+ * profiling support (thus, the proper flags to the simulator must
+ * be supplied).
+ */
+static __inline void
+sim_profiler_set_enabled(int enabled)
+{
+  int val =
+    enabled ? SIM_CONTROL_PROFILER_ENABLE : SIM_CONTROL_PROFILER_DISABLE;
+  __insn_mtspr(SPR_SIM_CONTROL, val);
+}
+
+
+/**
+ * Return true if and only if profiling is currently enabled
+ * for the current task.
+ *
+ * This returns false even if sim_profiler_enable() was called
+ * if the current execution environment does not support profiling.
+ */
+static __inline int
+sim_profiler_is_enabled(void)
+{
+  return ((__insn_mfspr(SPR_SIM_CONTROL) & SIM_PROFILER_ENABLED_MASK) != 0);
+}
+
+
+/**
+ * Reset profiling counters to zero for the current task.
+ *
+ * Resetting can be done while profiling is enabled.  It does not affect
+ * the chip-wide profiling counters.
+ */
+static __inline void
+sim_profiler_clear(void)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PROFILER_CLEAR);
+}
+
+
+/**
+ * Enable specified chip-level profiling counters.
+ *
+ * Does not affect the per-task profiling counters.
+ *
+ * @param mask Either this special value:
+ *
+ * SIM_CHIP_ALL (enables all chip-level components).
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_CHIP_MEMCTL (enable all memory controllers)
+ * SIM_CHIP_XAUI (enable all XAUI controllers)
+ * SIM_CHIP_MPIPE (enable all MPIPE controllers)
+ */
+static __inline void
+sim_profiler_chip_enable(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_PROFILER_CHIP_ENABLE_SPR_ARG(mask));
+}
+
+
+/**
+ * Disable specified chip-level profiling counters.
+ *
+ * Does not affect the per-task profiling counters.
+ *
+ * @param mask Either this special value:
+ *
+ * SIM_CHIP_ALL (disables all chip-level components).
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_CHIP_MEMCTL (disable all memory controllers)
+ * SIM_CHIP_XAUI (disable all XAUI controllers)
+ * SIM_CHIP_MPIPE (disable all MPIPE controllers)
+ */
+static __inline void
+sim_profiler_chip_disable(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_PROFILER_CHIP_DISABLE_SPR_ARG(mask));
+}
+
+
+/**
+ * Reset specified chip-level profiling counters to zero.
+ *
+ * Does not affect the per-task profiling counters.
+ *
+ * @param mask Either this special value:
+ *
+ * SIM_CHIP_ALL (clears all chip-level components).
+ *
+ * or the bitwise OR of these values:
+ *
+ * SIM_CHIP_MEMCTL (clear all memory controllers)
+ * SIM_CHIP_XAUI (clear all XAUI controllers)
+ * SIM_CHIP_MPIPE (clear all MPIPE controllers)
+ */
+static __inline void
+sim_profiler_chip_clear(unsigned int mask)
+{
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_PROFILER_CHIP_CLEAR_SPR_ARG(mask));
+}
+
+
+/*
+ * Event support.
+ */
+
+#ifndef __DOXYGEN__
+
+static __inline void
+sim_event_begin(unsigned int x)
+{
+#if defined(__tile__) && !defined(__NO_EVENT_SPR__)
+  __insn_mtspr(SPR_EVENT_BEGIN, x);
+#endif
+}
+
+static __inline void
+sim_event_end(unsigned int x)
+{
+#if defined(__tile__) && !defined(__NO_EVENT_SPR__)
+  __insn_mtspr(SPR_EVENT_END, x);
+#endif
+}
+
+#endif /* !__DOXYGEN__ */
+
+#endif /* !__ASSEMBLER__ */
+
+#endif /* !__ARCH_SIM_H__ */
+
+/** @} */
diff --git a/arch/tile/include/arch/sim_def.h b/arch/tile/include/arch/sim_def.h
index 6418fbde063e..7a17082c3773 100644
--- a/arch/tile/include/arch/sim_def.h
+++ b/arch/tile/include/arch/sim_def.h
@@ -1,477 +1,461 @@
-// Copyright 2010 Tilera 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, version 2.
-//
-//   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, GOOD TITLE or
-//   NON INFRINGEMENT.  See the GNU General Public License for
-//   more details.
-
-//! @file
-//!
-//! Some low-level simulator definitions.
-//!
+/*
+ * Copyright 2010 Tilera 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, version 2.
+ *
+ *   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, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file
+ *
+ * Some low-level simulator definitions.
+ */
 
 #ifndef __ARCH_SIM_DEF_H__
 #define __ARCH_SIM_DEF_H__
 
 
-//! Internal: the low bits of the SIM_CONTROL_* SPR values specify
-//! the operation to perform, and the remaining bits are
-//! an operation-specific parameter (often unused).
-//!
+/**
+ * Internal: the low bits of the SIM_CONTROL_* SPR values specify
+ * the operation to perform, and the remaining bits are
+ * an operation-specific parameter (often unused).
+ */
 #define _SIM_CONTROL_OPERATOR_BITS 8
 
 
-//== Values which can be written to SPR_SIM_CONTROL.
+/*
+ * Values which can be written to SPR_SIM_CONTROL.
+ */
 
-//! If written to SPR_SIM_CONTROL, stops profiling.
-//!
+/** If written to SPR_SIM_CONTROL, stops profiling. */
 #define SIM_CONTROL_PROFILER_DISABLE 0
 
-//! If written to SPR_SIM_CONTROL, starts profiling.
-//!
+/** If written to SPR_SIM_CONTROL, starts profiling. */
 #define SIM_CONTROL_PROFILER_ENABLE 1
 
-//! If written to SPR_SIM_CONTROL, clears profiling counters.
-//!
+/** If written to SPR_SIM_CONTROL, clears profiling counters. */
 #define SIM_CONTROL_PROFILER_CLEAR 2
 
-//! If written to SPR_SIM_CONTROL, checkpoints the simulator.
-//!
+/** If written to SPR_SIM_CONTROL, checkpoints the simulator. */
 #define SIM_CONTROL_CHECKPOINT 3
 
-//! If written to SPR_SIM_CONTROL, combined with a mask (shifted by 8),
-//! sets the tracing mask to the given mask. See "sim_set_tracing()".
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mask (shifted by 8),
+ * sets the tracing mask to the given mask. See "sim_set_tracing()".
+ */
 #define SIM_CONTROL_SET_TRACING 4
 
-//! If written to SPR_SIM_CONTROL, combined with a mask (shifted by 8),
-//! dumps the requested items of machine state to the log.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mask (shifted by 8),
+ * dumps the requested items of machine state to the log.
+ */
 #define SIM_CONTROL_DUMP 5
 
-//! If written to SPR_SIM_CONTROL, clears chip-level profiling counters.
-//!
+/** If written to SPR_SIM_CONTROL, clears chip-level profiling counters. */
 #define SIM_CONTROL_PROFILER_CHIP_CLEAR 6
 
-//! If written to SPR_SIM_CONTROL, disables chip-level profiling.
-//!
+/** If written to SPR_SIM_CONTROL, disables chip-level profiling. */
 #define SIM_CONTROL_PROFILER_CHIP_DISABLE 7
 
-//! If written to SPR_SIM_CONTROL, enables chip-level profiling.
-//!
+/** If written to SPR_SIM_CONTROL, enables chip-level profiling. */
 #define SIM_CONTROL_PROFILER_CHIP_ENABLE 8
 
-//! If written to SPR_SIM_CONTROL, enables chip-level functional mode
-//!
+/** If written to SPR_SIM_CONTROL, enables chip-level functional mode */
 #define SIM_CONTROL_ENABLE_FUNCTIONAL 9
 
-//! If written to SPR_SIM_CONTROL, disables chip-level functional mode.
-//!
+/** If written to SPR_SIM_CONTROL, disables chip-level functional mode. */
 #define SIM_CONTROL_DISABLE_FUNCTIONAL 10
 
-//! If written to SPR_SIM_CONTROL, enables chip-level functional mode.
-//! All tiles must perform this write for functional mode to be enabled.
-//! Ignored in naked boot mode unless --functional is specified.
-//! WARNING: Only the hypervisor startup code should use this!
-//!
+/**
+ * If written to SPR_SIM_CONTROL, enables chip-level functional mode.
+ * All tiles must perform this write for functional mode to be enabled.
+ * Ignored in naked boot mode unless --functional is specified.
+ * WARNING: Only the hypervisor startup code should use this!
+ */
 #define SIM_CONTROL_ENABLE_FUNCTIONAL_BARRIER 11
 
-//! If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
-//! writes a string directly to the simulator output.  Written to once for
-//! each character in the string, plus a final NUL.  Instead of NUL,
-//! you can also use "SIM_PUTC_FLUSH_STRING" or "SIM_PUTC_FLUSH_BINARY".
-//!
-// ISSUE: Document the meaning of "newline", and the handling of NUL.
-//
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * writes a string directly to the simulator output.  Written to once for
+ * each character in the string, plus a final NUL.  Instead of NUL,
+ * you can also use "SIM_PUTC_FLUSH_STRING" or "SIM_PUTC_FLUSH_BINARY".
+ */
+/* ISSUE: Document the meaning of "newline", and the handling of NUL. */
 #define SIM_CONTROL_PUTC 12
 
-//! If written to SPR_SIM_CONTROL, clears the --grind-coherence state for
-//! this core.  This is intended to be used before a loop that will
-//! invalidate the cache by loading new data and evicting all current data.
-//! Generally speaking, this API should only be used by system code.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, clears the --grind-coherence state for
+ * this core.  This is intended to be used before a loop that will
+ * invalidate the cache by loading new data and evicting all current data.
+ * Generally speaking, this API should only be used by system code.
+ */
 #define SIM_CONTROL_GRINDER_CLEAR 13
 
-//! If written to SPR_SIM_CONTROL, shuts down the simulator.
-//!
+/** If written to SPR_SIM_CONTROL, shuts down the simulator. */
 #define SIM_CONTROL_SHUTDOWN 14
 
-//! If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
-//! indicates that a fork syscall just created the given process.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
+ * indicates that a fork syscall just created the given process.
+ */
 #define SIM_CONTROL_OS_FORK 15
 
-//! If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
-//! indicates that an exit syscall was just executed by the given process.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
+ * indicates that an exit syscall was just executed by the given process.
+ */
 #define SIM_CONTROL_OS_EXIT 16
 
-//! If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
-//! indicates that the OS just switched to the given process.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
+ * indicates that the OS just switched to the given process.
+ */
 #define SIM_CONTROL_OS_SWITCH 17
 
-//! If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
-//! indicates that an exec syscall was just executed. Written to once for
-//! each character in the executable name, plus a final NUL.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * indicates that an exec syscall was just executed. Written to once for
+ * each character in the executable name, plus a final NUL.
+ */
 #define SIM_CONTROL_OS_EXEC 18
 
-//! If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
-//! indicates that an interpreter (PT_INTERP) was loaded.  Written to once
-//! for each character in "ADDR:PATH", plus a final NUL, where "ADDR" is a
-//! hex load address starting with "0x", and "PATH" is the executable name.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * indicates that an interpreter (PT_INTERP) was loaded.  Written to once
+ * for each character in "ADDR:PATH", plus a final NUL, where "ADDR" is a
+ * hex load address starting with "0x", and "PATH" is the executable name.
+ */
 #define SIM_CONTROL_OS_INTERP 19
 
-//! If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
-//! indicates that a dll was loaded.  Written to once for each character
-//! in "ADDR:PATH", plus a final NUL, where "ADDR" is a hexadecimal load
-//! address starting with "0x", and "PATH" is the executable name.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * indicates that a dll was loaded.  Written to once for each character
+ * in "ADDR:PATH", plus a final NUL, where "ADDR" is a hexadecimal load
+ * address starting with "0x", and "PATH" is the executable name.
+ */
 #define SIM_CONTROL_DLOPEN 20
 
-//! If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
-//! indicates that a dll was unloaded.  Written to once for each character
-//! in "ADDR", plus a final NUL, where "ADDR" is a hexadecimal load
-//! address starting with "0x".
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a character (shifted by 8),
+ * indicates that a dll was unloaded.  Written to once for each character
+ * in "ADDR", plus a final NUL, where "ADDR" is a hexadecimal load
+ * address starting with "0x".
+ */
 #define SIM_CONTROL_DLCLOSE 21
 
-//! If written to SPR_SIM_CONTROL, combined with a flag (shifted by 8),
-//! indicates whether to allow data reads to remotely-cached
-//! dirty cache lines to be cached locally without grinder warnings or
-//! assertions (used by Linux kernel fast memcpy).
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a flag (shifted by 8),
+ * indicates whether to allow data reads to remotely-cached
+ * dirty cache lines to be cached locally without grinder warnings or
+ * assertions (used by Linux kernel fast memcpy).
+ */
 #define SIM_CONTROL_ALLOW_MULTIPLE_CACHING 22
 
-//! If written to SPR_SIM_CONTROL, enables memory tracing.
-//!
+/** If written to SPR_SIM_CONTROL, enables memory tracing. */
 #define SIM_CONTROL_ENABLE_MEM_LOGGING 23
 
-//! If written to SPR_SIM_CONTROL, disables memory tracing.
-//!
+/** If written to SPR_SIM_CONTROL, disables memory tracing. */
 #define SIM_CONTROL_DISABLE_MEM_LOGGING 24
 
-//! If written to SPR_SIM_CONTROL, changes the shaping parameters of one of
-//! the gbe or xgbe shims. Must specify the shim id, the type, the units, and
-//! the rate, as defined in SIM_SHAPING_SPR_ARG.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, changes the shaping parameters of one of
+ * the gbe or xgbe shims. Must specify the shim id, the type, the units, and
+ * the rate, as defined in SIM_SHAPING_SPR_ARG.
+ */
 #define SIM_CONTROL_SHAPING 25
 
-//! If written to SPR_SIM_CONTROL, combined with character (shifted by 8),
-//! requests that a simulator command be executed.  Written to once for each
-//! character in the command, plus a final NUL.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with character (shifted by 8),
+ * requests that a simulator command be executed.  Written to once for each
+ * character in the command, plus a final NUL.
+ */
 #define SIM_CONTROL_COMMAND 26
 
-//! If written to SPR_SIM_CONTROL, indicates that the simulated system
-//! is panicking, to allow debugging via --debug-on-panic.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, indicates that the simulated system
+ * is panicking, to allow debugging via --debug-on-panic.
+ */
 #define SIM_CONTROL_PANIC 27
 
-//! If written to SPR_SIM_CONTROL, triggers a simulator syscall.
-//! See "sim_syscall()" for more info.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, triggers a simulator syscall.
+ * See "sim_syscall()" for more info.
+ */
 #define SIM_CONTROL_SYSCALL 32
 
-//! If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
-//! provides the pid that subsequent SIM_CONTROL_OS_FORK writes should
-//! use as the pid, rather than the default previous SIM_CONTROL_OS_SWITCH.
-//!
+/**
+ * If written to SPR_SIM_CONTROL, combined with a pid (shifted by 8),
+ * provides the pid that subsequent SIM_CONTROL_OS_FORK writes should
+ * use as the pid, rather than the default previous SIM_CONTROL_OS_SWITCH.
+ */
 #define SIM_CONTROL_OS_FORK_PARENT 33
 
-//! If written to SPR_SIM_CONTROL, combined with a mPIPE shim number
-//! (shifted by 8), clears the pending magic data section.  The cleared
-//! pending magic data section and any subsequently appended magic bytes
-//! will only take effect when the classifier blast programmer is run.
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mPIPE shim number
+ * (shifted by 8), clears the pending magic data section.  The cleared
+ * pending magic data section and any subsequently appended magic bytes
+ * will only take effect when the classifier blast programmer is run.
+ */
 #define SIM_CONTROL_CLEAR_MPIPE_MAGIC_BYTES 34
 
-//! If written to SPR_SIM_CONTROL, combined with a mPIPE shim number
-//! (shifted by 8) and a byte of data (shifted by 16), appends that byte
-//! to the shim's pending magic data section.  The pending magic data
-//! section takes effect when the classifier blast programmer is run.
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mPIPE shim number
+ * (shifted by 8) and a byte of data (shifted by 16), appends that byte
+ * to the shim's pending magic data section.  The pending magic data
+ * section takes effect when the classifier blast programmer is run.
+ */
 #define SIM_CONTROL_APPEND_MPIPE_MAGIC_BYTE 35
 
-//! If written to SPR_SIM_CONTROL, combined with a mPIPE shim number
-//! (shifted by 8), an enable=1/disable=0 bit (shifted by 16), and a
-//! mask of links (shifted by 32), enable or disable the corresponding
-//! mPIPE links.
+/**
+ * If written to SPR_SIM_CONTROL, combined with a mPIPE shim number
+ * (shifted by 8), an enable=1/disable=0 bit (shifted by 16), and a
+ * mask of links (shifted by 32), enable or disable the corresponding
+ * mPIPE links.
+ */
 #define SIM_CONTROL_ENABLE_MPIPE_LINK_MAGIC_BYTE 36
 
-//== Syscall numbers for use with "sim_syscall()".
 
-//! Syscall number for sim_add_watchpoint().
-//!
+/*
+ * Syscall numbers for use with "sim_syscall()".
+ */
+
+/** Syscall number for sim_add_watchpoint(). */
 #define SIM_SYSCALL_ADD_WATCHPOINT 2
 
-//! Syscall number for sim_remove_watchpoint().
-//!
+/** Syscall number for sim_remove_watchpoint(). */
 #define SIM_SYSCALL_REMOVE_WATCHPOINT 3
 
-//! Syscall number for sim_query_watchpoint().
-//!
+/** Syscall number for sim_query_watchpoint(). */
 #define SIM_SYSCALL_QUERY_WATCHPOINT 4
 
-//! Syscall number that asserts that the cache lines whose 64-bit PA
-//! is passed as the second argument to sim_syscall(), and over a
-//! range passed as the third argument, are no longer in cache.
-//! The simulator raises an error if this is not the case.
-//!
+/**
+ * Syscall number that asserts that the cache lines whose 64-bit PA
+ * is passed as the second argument to sim_syscall(), and over a
+ * range passed as the third argument, are no longer in cache.
+ * The simulator raises an error if this is not the case.
+ */
 #define SIM_SYSCALL_VALIDATE_LINES_EVICTED 5
 
 
-//== Bit masks which can be shifted by 8, combined with
-//== SIM_CONTROL_SET_TRACING, and written to SPR_SIM_CONTROL.
+/*
+ * Bit masks which can be shifted by 8, combined with
+ * SIM_CONTROL_SET_TRACING, and written to SPR_SIM_CONTROL.
+ */
 
-//! @addtogroup arch_sim
-//! @{
+/**
+ * @addtogroup arch_sim
+ * @{
+ */
 
-//! Enable --trace-cycle when passed to simulator_set_tracing().
-//!
+/** Enable --trace-cycle when passed to simulator_set_tracing(). */
 #define SIM_TRACE_CYCLES          0x01
 
-//! Enable --trace-router when passed to simulator_set_tracing().
-//!
+/** Enable --trace-router when passed to simulator_set_tracing(). */
 #define SIM_TRACE_ROUTER          0x02
 
-//! Enable --trace-register-writes when passed to simulator_set_tracing().
-//!
+/** Enable --trace-register-writes when passed to simulator_set_tracing(). */
 #define SIM_TRACE_REGISTER_WRITES 0x04
 
-//! Enable --trace-disasm when passed to simulator_set_tracing().
-//!
+/** Enable --trace-disasm when passed to simulator_set_tracing(). */
 #define SIM_TRACE_DISASM          0x08
 
-//! Enable --trace-stall-info when passed to simulator_set_tracing().
-//!
+/** Enable --trace-stall-info when passed to simulator_set_tracing(). */
 #define SIM_TRACE_STALL_INFO      0x10
 
-//! Enable --trace-memory-controller when passed to simulator_set_tracing().
-//!
+/** Enable --trace-memory-controller when passed to simulator_set_tracing(). */
 #define SIM_TRACE_MEMORY_CONTROLLER 0x20
 
-//! Enable --trace-l2 when passed to simulator_set_tracing().
-//!
+/** Enable --trace-l2 when passed to simulator_set_tracing(). */
 #define SIM_TRACE_L2_CACHE 0x40
 
-//! Enable --trace-lines when passed to simulator_set_tracing().
-//!
+/** Enable --trace-lines when passed to simulator_set_tracing(). */
 #define SIM_TRACE_LINES 0x80
 
-//! Turn off all tracing when passed to simulator_set_tracing().
-//!
+/** Turn off all tracing when passed to simulator_set_tracing(). */
 #define SIM_TRACE_NONE 0
 
-//! Turn on all tracing when passed to simulator_set_tracing().
-//!
+/** Turn on all tracing when passed to simulator_set_tracing(). */
 #define SIM_TRACE_ALL (-1)
 
-//! @}
+/** @} */
 
-//! Computes the value to write to SPR_SIM_CONTROL to set tracing flags.
-//!
+/** Computes the value to write to SPR_SIM_CONTROL to set tracing flags. */
 #define SIM_TRACE_SPR_ARG(mask) \
   (SIM_CONTROL_SET_TRACING | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
 
 
-//== Bit masks which can be shifted by 8, combined with
-//== SIM_CONTROL_DUMP, and written to SPR_SIM_CONTROL.
+/*
+ * Bit masks which can be shifted by 8, combined with
+ * SIM_CONTROL_DUMP, and written to SPR_SIM_CONTROL.
+ */
 
-//! @addtogroup arch_sim
-//! @{
+/**
+ * @addtogroup arch_sim
+ * @{
+ */
 
-//! Dump the general-purpose registers.
-//!
+/** Dump the general-purpose registers. */
 #define SIM_DUMP_REGS          0x001
 
-//! Dump the SPRs.
-//!
+/** Dump the SPRs. */
 #define SIM_DUMP_SPRS          0x002
 
-//! Dump the ITLB.
-//!
+/** Dump the ITLB. */
 #define SIM_DUMP_ITLB          0x004
 
-//! Dump the DTLB.
-//!
+/** Dump the DTLB. */
 #define SIM_DUMP_DTLB          0x008
 
-//! Dump the L1 I-cache.
-//!
+/** Dump the L1 I-cache. */
 #define SIM_DUMP_L1I           0x010
 
-//! Dump the L1 D-cache.
-//!
+/** Dump the L1 D-cache. */
 #define SIM_DUMP_L1D           0x020
 
-//! Dump the L2 cache.
-//!
+/** Dump the L2 cache. */
 #define SIM_DUMP_L2            0x040
 
-//! Dump the switch registers.
-//!
+/** Dump the switch registers. */
 #define SIM_DUMP_SNREGS        0x080
 
-//! Dump the switch ITLB.
-//!
+/** Dump the switch ITLB. */
 #define SIM_DUMP_SNITLB        0x100
 
-//! Dump the switch L1 I-cache.
-//!
+/** Dump the switch L1 I-cache. */
 #define SIM_DUMP_SNL1I         0x200
 
-//! Dump the current backtrace.
-//!
+/** Dump the current backtrace. */
 #define SIM_DUMP_BACKTRACE     0x400
 
-//! Only dump valid lines in caches.
-//!
+/** Only dump valid lines in caches. */
 #define SIM_DUMP_VALID_LINES   0x800
 
-//! Dump everything that is dumpable.
-//!
+/** Dump everything that is dumpable. */
 #define SIM_DUMP_ALL (-1 & ~SIM_DUMP_VALID_LINES)
 
-// @}
+/** @} */
 
-//! Computes the value to write to SPR_SIM_CONTROL to dump machine state.
-//!
+/** Computes the value to write to SPR_SIM_CONTROL to dump machine state. */
 #define SIM_DUMP_SPR_ARG(mask) \
   (SIM_CONTROL_DUMP | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
 
 
-//== Bit masks which can be shifted by 8, combined with
-//== SIM_CONTROL_PROFILER_CHIP_xxx, and written to SPR_SIM_CONTROL.
+/*
+ * Bit masks which can be shifted by 8, combined with
+ * SIM_CONTROL_PROFILER_CHIP_xxx, and written to SPR_SIM_CONTROL.
+ */
 
-//! @addtogroup arch_sim
-//! @{
+/**
+ * @addtogroup arch_sim
+ * @{
+ */
 
-//! Use with with SIM_PROFILER_CHIP_xxx to control the memory controllers.
-//!
+/** Use with with SIM_PROFILER_CHIP_xxx to control the memory controllers. */
 #define SIM_CHIP_MEMCTL        0x001
 
-//! Use with with SIM_PROFILER_CHIP_xxx to control the XAUI interface.
-//!
+/** Use with with SIM_PROFILER_CHIP_xxx to control the XAUI interface. */
 #define SIM_CHIP_XAUI          0x002
 
-//! Use with with SIM_PROFILER_CHIP_xxx to control the PCIe interface.
-//!
+/** Use with with SIM_PROFILER_CHIP_xxx to control the PCIe interface. */
 #define SIM_CHIP_PCIE          0x004
 
-//! Use with with SIM_PROFILER_CHIP_xxx to control the MPIPE interface.
-//!
+/** Use with with SIM_PROFILER_CHIP_xxx to control the MPIPE interface. */
 #define SIM_CHIP_MPIPE         0x008
 
-//! Reference all chip devices.
-//!
+/** Use with with SIM_PROFILER_CHIP_xxx to control the TRIO interface. */
+#define SIM_CHIP_TRIO          0x010
+
+/** Reference all chip devices. */
 #define SIM_CHIP_ALL (-1)
 
-//! @}
+/** @} */
 
-//! Computes the value to write to SPR_SIM_CONTROL to clear chip statistics.
-//!
+/** Computes the value to write to SPR_SIM_CONTROL to clear chip statistics. */
 #define SIM_PROFILER_CHIP_CLEAR_SPR_ARG(mask) \
   (SIM_CONTROL_PROFILER_CHIP_CLEAR | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
 
-//! Computes the value to write to SPR_SIM_CONTROL to disable chip statistics.
-//!
+/** Computes the value to write to SPR_SIM_CONTROL to disable chip statistics.*/
 #define SIM_PROFILER_CHIP_DISABLE_SPR_ARG(mask) \
   (SIM_CONTROL_PROFILER_CHIP_DISABLE | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
 
-//! Computes the value to write to SPR_SIM_CONTROL to enable chip statistics.
-//!
+/** Computes the value to write to SPR_SIM_CONTROL to enable chip statistics. */
 #define SIM_PROFILER_CHIP_ENABLE_SPR_ARG(mask) \
   (SIM_CONTROL_PROFILER_CHIP_ENABLE | ((mask) << _SIM_CONTROL_OPERATOR_BITS))
 
 
 
-// Shim bitrate controls.
+/* Shim bitrate controls. */
 
-//! The number of bits used to store the shim id.
-//!
+/** The number of bits used to store the shim id. */
 #define SIM_CONTROL_SHAPING_SHIM_ID_BITS 3
 
-//! @addtogroup arch_sim
-//! @{
+/**
+ * @addtogroup arch_sim
+ * @{
+ */
 
-//! Change the gbe 0 bitrate.
-//!
+/** Change the gbe 0 bitrate. */
 #define SIM_CONTROL_SHAPING_GBE_0 0x0
 
-//! Change the gbe 1 bitrate.
-//!
+/** Change the gbe 1 bitrate. */
 #define SIM_CONTROL_SHAPING_GBE_1 0x1
 
-//! Change the gbe 2 bitrate.
-//!
+/** Change the gbe 2 bitrate. */
 #define SIM_CONTROL_SHAPING_GBE_2 0x2
 
-//! Change the gbe 3 bitrate.
-//!
+/** Change the gbe 3 bitrate. */
 #define SIM_CONTROL_SHAPING_GBE_3 0x3
 
-//! Change the xgbe 0 bitrate.
-//!
+/** Change the xgbe 0 bitrate. */
 #define SIM_CONTROL_SHAPING_XGBE_0 0x4
 
-//! Change the xgbe 1 bitrate.
-//!
+/** Change the xgbe 1 bitrate. */
 #define SIM_CONTROL_SHAPING_XGBE_1 0x5
 
-//! The type of shaping to do.
-//!
+/** The type of shaping to do. */
 #define SIM_CONTROL_SHAPING_TYPE_BITS 2
 
-//! Control the multiplier.
-//!
+/** Control the multiplier. */
 #define SIM_CONTROL_SHAPING_MULTIPLIER 0
 
-//! Control the PPS.
-//!
+/** Control the PPS. */
 #define SIM_CONTROL_SHAPING_PPS 1
 
-//! Control the BPS.
-//!
+/** Control the BPS. */
 #define SIM_CONTROL_SHAPING_BPS 2
 
-//! The number of bits for the units for the shaping parameter.
-//!
+/** The number of bits for the units for the shaping parameter. */
 #define SIM_CONTROL_SHAPING_UNITS_BITS 2
 
-//! Provide a number in single units.
-//!
+/** Provide a number in single units. */
 #define SIM_CONTROL_SHAPING_UNITS_SINGLE 0
 
-//! Provide a number in kilo units.
-//!
+/** Provide a number in kilo units. */
 #define SIM_CONTROL_SHAPING_UNITS_KILO 1
 
-//! Provide a number in mega units.
-//!
+/** Provide a number in mega units. */
 #define SIM_CONTROL_SHAPING_UNITS_MEGA 2
 
-//! Provide a number in giga units.
-//!
+/** Provide a number in giga units. */
 #define SIM_CONTROL_SHAPING_UNITS_GIGA 3
 
-// @}
+/** @} */
 
-//! How many bits are available for the rate.
-//!
+/** How many bits are available for the rate. */
 #define SIM_CONTROL_SHAPING_RATE_BITS \
   (32 - (_SIM_CONTROL_OPERATOR_BITS + \
          SIM_CONTROL_SHAPING_SHIM_ID_BITS + \
          SIM_CONTROL_SHAPING_TYPE_BITS + \
          SIM_CONTROL_SHAPING_UNITS_BITS))
 
-//! Computes the value to write to SPR_SIM_CONTROL to change a bitrate.
-//!
+/** Computes the value to write to SPR_SIM_CONTROL to change a bitrate. */
 #define SIM_SHAPING_SPR_ARG(shim, type, units, rate) \
   (SIM_CONTROL_SHAPING | \
    ((shim) | \
@@ -483,30 +467,36 @@
                SIM_CONTROL_SHAPING_UNITS_BITS))) << _SIM_CONTROL_OPERATOR_BITS)
 
 
-//== Values returned when reading SPR_SIM_CONTROL.
-// ISSUE: These names should share a longer common prefix.
+/*
+ * Values returned when reading SPR_SIM_CONTROL.
+ * ISSUE: These names should share a longer common prefix.
+ */
 
-//! When reading SPR_SIM_CONTROL, the mask of simulator tracing bits
-//! (SIM_TRACE_xxx values).
-//!
+/**
+ * When reading SPR_SIM_CONTROL, the mask of simulator tracing bits
+ * (SIM_TRACE_xxx values).
+ */
 #define SIM_TRACE_FLAG_MASK 0xFFFF
 
-//! When reading SPR_SIM_CONTROL, the mask for whether profiling is enabled.
-//!
+/** When reading SPR_SIM_CONTROL, the mask for whether profiling is enabled. */
 #define SIM_PROFILER_ENABLED_MASK 0x10000
 
 
-//== Special arguments for "SIM_CONTROL_PUTC".
+/*
+ * Special arguments for "SIM_CONTROL_PUTC".
+ */
 
-//! Flag value for forcing a PUTC string-flush, including
-//! coordinate/cycle prefix and newline.
-//!
+/**
+ * Flag value for forcing a PUTC string-flush, including
+ * coordinate/cycle prefix and newline.
+ */
 #define SIM_PUTC_FLUSH_STRING 0x100
 
-//! Flag value for forcing a PUTC binary-data-flush, which skips the
-//! prefix and does not append a newline.
-//!
+/**
+ * Flag value for forcing a PUTC binary-data-flush, which skips the
+ * prefix and does not append a newline.
+ */
 #define SIM_PUTC_FLUSH_BINARY 0x101
 
 
-#endif //__ARCH_SIM_DEF_H__
+#endif /* __ARCH_SIM_DEF_H__ */
diff --git a/arch/tile/include/arch/spr_def.h b/arch/tile/include/arch/spr_def.h
index c8fdbd9a45e6..442fcba0d122 100644
--- a/arch/tile/include/arch/spr_def.h
+++ b/arch/tile/include/arch/spr_def.h
@@ -12,8 +12,93 @@
  *   more details.
  */
 
+/*
+ * In addition to including the proper base SPR definition file, depending
+ * on machine architecture, this file defines several macros which allow
+ * kernel code to use protection-level dependent SPRs without worrying
+ * about which PL it's running at.  In these macros, the PL that the SPR
+ * or interrupt number applies to is replaced by K.
+ */
+
+#if CONFIG_KERNEL_PL != 1 && CONFIG_KERNEL_PL != 2
+#error CONFIG_KERNEL_PL must be 1 or 2
+#endif
+
+/* Concatenate 4 strings. */
+#define __concat4(a, b, c, d) a ## b ## c ## d
+#define _concat4(a, b, c, d)  __concat4(a, b, c, d)
+
 #ifdef __tilegx__
 #include <arch/spr_def_64.h>
+
+/* TILE-Gx dependent, protection-level dependent SPRs. */
+
+#define SPR_INTERRUPT_MASK_K \
+	_concat4(SPR_INTERRUPT_MASK_, CONFIG_KERNEL_PL,,)
+#define SPR_INTERRUPT_MASK_SET_K \
+	_concat4(SPR_INTERRUPT_MASK_SET_, CONFIG_KERNEL_PL,,)
+#define SPR_INTERRUPT_MASK_RESET_K \
+	_concat4(SPR_INTERRUPT_MASK_RESET_, CONFIG_KERNEL_PL,,)
+#define SPR_INTERRUPT_VECTOR_BASE_K \
+	_concat4(SPR_INTERRUPT_VECTOR_BASE_, CONFIG_KERNEL_PL,,)
+
+#define SPR_IPI_MASK_K \
+	_concat4(SPR_IPI_MASK_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_MASK_RESET_K \
+	_concat4(SPR_IPI_MASK_RESET_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_MASK_SET_K \
+	_concat4(SPR_IPI_MASK_SET_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_EVENT_K \
+	_concat4(SPR_IPI_EVENT_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_EVENT_RESET_K \
+	_concat4(SPR_IPI_EVENT_RESET_, CONFIG_KERNEL_PL,,)
+#define SPR_IPI_MASK_SET_K \
+	_concat4(SPR_IPI_MASK_SET_, CONFIG_KERNEL_PL,,)
+#define INT_IPI_K \
+	_concat4(INT_IPI_, CONFIG_KERNEL_PL,,)
+
+#define SPR_SINGLE_STEP_CONTROL_K \
+	_concat4(SPR_SINGLE_STEP_CONTROL_, CONFIG_KERNEL_PL,,)
+#define SPR_SINGLE_STEP_EN_K_K \
+	_concat4(SPR_SINGLE_STEP_EN_, CONFIG_KERNEL_PL, _, CONFIG_KERNEL_PL)
+#define INT_SINGLE_STEP_K \
+	_concat4(INT_SINGLE_STEP_, CONFIG_KERNEL_PL,,)
+
 #else
 #include <arch/spr_def_32.h>
+
+/* TILEPro dependent, protection-level dependent SPRs. */
+
+#define SPR_INTERRUPT_MASK_K_0 \
+	_concat4(SPR_INTERRUPT_MASK_, CONFIG_KERNEL_PL, _0,)
+#define SPR_INTERRUPT_MASK_K_1 \
+	_concat4(SPR_INTERRUPT_MASK_, CONFIG_KERNEL_PL, _1,)
+#define SPR_INTERRUPT_MASK_SET_K_0 \
+	_concat4(SPR_INTERRUPT_MASK_SET_, CONFIG_KERNEL_PL, _0,)
+#define SPR_INTERRUPT_MASK_SET_K_1 \
+	_concat4(SPR_INTERRUPT_MASK_SET_, CONFIG_KERNEL_PL, _1,)
+#define SPR_INTERRUPT_MASK_RESET_K_0 \
+	_concat4(SPR_INTERRUPT_MASK_RESET_, CONFIG_KERNEL_PL, _0,)
+#define SPR_INTERRUPT_MASK_RESET_K_1 \
+	_concat4(SPR_INTERRUPT_MASK_RESET_, CONFIG_KERNEL_PL, _1,)
+
 #endif
+
+/* Generic protection-level dependent SPRs. */
+
+#define SPR_SYSTEM_SAVE_K_0 \
+	_concat4(SPR_SYSTEM_SAVE_, CONFIG_KERNEL_PL, _0,)
+#define SPR_SYSTEM_SAVE_K_1 \
+	_concat4(SPR_SYSTEM_SAVE_, CONFIG_KERNEL_PL, _1,)
+#define SPR_SYSTEM_SAVE_K_2 \
+	_concat4(SPR_SYSTEM_SAVE_, CONFIG_KERNEL_PL, _2,)
+#define SPR_SYSTEM_SAVE_K_3 \
+	_concat4(SPR_SYSTEM_SAVE_, CONFIG_KERNEL_PL, _3,)
+#define SPR_EX_CONTEXT_K_0 \
+	_concat4(SPR_EX_CONTEXT_, CONFIG_KERNEL_PL, _0,)
+#define SPR_EX_CONTEXT_K_1 \
+	_concat4(SPR_EX_CONTEXT_, CONFIG_KERNEL_PL, _1,)
+#define SPR_INTCTRL_K_STATUS \
+	_concat4(SPR_INTCTRL_, CONFIG_KERNEL_PL, _STATUS,)
+#define INT_INTCTRL_K \
+	_concat4(INT_INTCTRL_, CONFIG_KERNEL_PL,,)
diff --git a/arch/tile/include/arch/spr_def_32.h b/arch/tile/include/arch/spr_def_32.h
index b4fc06864df6..bbc1f4c924ee 100644
--- a/arch/tile/include/arch/spr_def_32.h
+++ b/arch/tile/include/arch/spr_def_32.h
@@ -56,58 +56,93 @@
 #define SPR_EX_CONTEXT_1_1__ICS_SHIFT 2
 #define SPR_EX_CONTEXT_1_1__ICS_RMASK 0x1
 #define SPR_EX_CONTEXT_1_1__ICS_MASK  0x4
+#define SPR_EX_CONTEXT_2_0 0x4605
+#define SPR_EX_CONTEXT_2_1 0x4606
+#define SPR_EX_CONTEXT_2_1__PL_SHIFT 0
+#define SPR_EX_CONTEXT_2_1__PL_RMASK 0x3
+#define SPR_EX_CONTEXT_2_1__PL_MASK  0x3
+#define SPR_EX_CONTEXT_2_1__ICS_SHIFT 2
+#define SPR_EX_CONTEXT_2_1__ICS_RMASK 0x1
+#define SPR_EX_CONTEXT_2_1__ICS_MASK  0x4
 #define SPR_FAIL 0x4e09
 #define SPR_INTCTRL_0_STATUS 0x4a07
 #define SPR_INTCTRL_1_STATUS 0x4807
+#define SPR_INTCTRL_2_STATUS 0x4607
 #define SPR_INTERRUPT_CRITICAL_SECTION 0x4e0a
 #define SPR_INTERRUPT_MASK_0_0 0x4a08
 #define SPR_INTERRUPT_MASK_0_1 0x4a09
 #define SPR_INTERRUPT_MASK_1_0 0x4809
 #define SPR_INTERRUPT_MASK_1_1 0x480a
+#define SPR_INTERRUPT_MASK_2_0 0x4608
+#define SPR_INTERRUPT_MASK_2_1 0x4609
 #define SPR_INTERRUPT_MASK_RESET_0_0 0x4a0a
 #define SPR_INTERRUPT_MASK_RESET_0_1 0x4a0b
 #define SPR_INTERRUPT_MASK_RESET_1_0 0x480b
 #define SPR_INTERRUPT_MASK_RESET_1_1 0x480c
+#define SPR_INTERRUPT_MASK_RESET_2_0 0x460a
+#define SPR_INTERRUPT_MASK_RESET_2_1 0x460b
 #define SPR_INTERRUPT_MASK_SET_0_0 0x4a0c
 #define SPR_INTERRUPT_MASK_SET_0_1 0x4a0d
 #define SPR_INTERRUPT_MASK_SET_1_0 0x480d
 #define SPR_INTERRUPT_MASK_SET_1_1 0x480e
+#define SPR_INTERRUPT_MASK_SET_2_0 0x460c
+#define SPR_INTERRUPT_MASK_SET_2_1 0x460d
 #define SPR_MPL_DMA_CPL_SET_0 0x5800
 #define SPR_MPL_DMA_CPL_SET_1 0x5801
+#define SPR_MPL_DMA_CPL_SET_2 0x5802
 #define SPR_MPL_DMA_NOTIFY_SET_0 0x3800
 #define SPR_MPL_DMA_NOTIFY_SET_1 0x3801
+#define SPR_MPL_DMA_NOTIFY_SET_2 0x3802
 #define SPR_MPL_INTCTRL_0_SET_0 0x4a00
 #define SPR_MPL_INTCTRL_0_SET_1 0x4a01
+#define SPR_MPL_INTCTRL_0_SET_2 0x4a02
 #define SPR_MPL_INTCTRL_1_SET_0 0x4800
 #define SPR_MPL_INTCTRL_1_SET_1 0x4801
+#define SPR_MPL_INTCTRL_1_SET_2 0x4802
+#define SPR_MPL_INTCTRL_2_SET_0 0x4600
+#define SPR_MPL_INTCTRL_2_SET_1 0x4601
+#define SPR_MPL_INTCTRL_2_SET_2 0x4602
 #define SPR_MPL_SN_ACCESS_SET_0 0x0800
 #define SPR_MPL_SN_ACCESS_SET_1 0x0801
+#define SPR_MPL_SN_ACCESS_SET_2 0x0802
 #define SPR_MPL_SN_CPL_SET_0 0x5a00
 #define SPR_MPL_SN_CPL_SET_1 0x5a01
+#define SPR_MPL_SN_CPL_SET_2 0x5a02
 #define SPR_MPL_SN_FIREWALL_SET_0 0x2c00
 #define SPR_MPL_SN_FIREWALL_SET_1 0x2c01
+#define SPR_MPL_SN_FIREWALL_SET_2 0x2c02
 #define SPR_MPL_SN_NOTIFY_SET_0 0x2a00
 #define SPR_MPL_SN_NOTIFY_SET_1 0x2a01
+#define SPR_MPL_SN_NOTIFY_SET_2 0x2a02
 #define SPR_MPL_UDN_ACCESS_SET_0 0x0c00
 #define SPR_MPL_UDN_ACCESS_SET_1 0x0c01
+#define SPR_MPL_UDN_ACCESS_SET_2 0x0c02
 #define SPR_MPL_UDN_AVAIL_SET_0 0x4000
 #define SPR_MPL_UDN_AVAIL_SET_1 0x4001
+#define SPR_MPL_UDN_AVAIL_SET_2 0x4002
 #define SPR_MPL_UDN_CA_SET_0 0x3c00
 #define SPR_MPL_UDN_CA_SET_1 0x3c01
+#define SPR_MPL_UDN_CA_SET_2 0x3c02
 #define SPR_MPL_UDN_COMPLETE_SET_0 0x1400
 #define SPR_MPL_UDN_COMPLETE_SET_1 0x1401
+#define SPR_MPL_UDN_COMPLETE_SET_2 0x1402
 #define SPR_MPL_UDN_FIREWALL_SET_0 0x3000
 #define SPR_MPL_UDN_FIREWALL_SET_1 0x3001
+#define SPR_MPL_UDN_FIREWALL_SET_2 0x3002
 #define SPR_MPL_UDN_REFILL_SET_0 0x1000
 #define SPR_MPL_UDN_REFILL_SET_1 0x1001
+#define SPR_MPL_UDN_REFILL_SET_2 0x1002
 #define SPR_MPL_UDN_TIMER_SET_0 0x3600
 #define SPR_MPL_UDN_TIMER_SET_1 0x3601
+#define SPR_MPL_UDN_TIMER_SET_2 0x3602
 #define SPR_MPL_WORLD_ACCESS_SET_0 0x4e00
 #define SPR_MPL_WORLD_ACCESS_SET_1 0x4e01
+#define SPR_MPL_WORLD_ACCESS_SET_2 0x4e02
 #define SPR_PASS 0x4e0b
 #define SPR_PERF_COUNT_0 0x4205
 #define SPR_PERF_COUNT_1 0x4206
 #define SPR_PERF_COUNT_CTL 0x4207
+#define SPR_PERF_COUNT_DN_CTL 0x4210
 #define SPR_PERF_COUNT_STS 0x4208
 #define SPR_PROC_STATUS 0x4f00
 #define SPR_SIM_CONTROL 0x4e0c
@@ -124,6 +159,10 @@
 #define SPR_SYSTEM_SAVE_1_1 0x4901
 #define SPR_SYSTEM_SAVE_1_2 0x4902
 #define SPR_SYSTEM_SAVE_1_3 0x4903
+#define SPR_SYSTEM_SAVE_2_0 0x4700
+#define SPR_SYSTEM_SAVE_2_1 0x4701
+#define SPR_SYSTEM_SAVE_2_2 0x4702
+#define SPR_SYSTEM_SAVE_2_3 0x4703
 #define SPR_TILE_COORD 0x4c17
 #define SPR_TILE_RTF_HWM 0x4e10
 #define SPR_TILE_TIMER_CONTROL 0x3205
diff --git a/arch/tile/include/asm/backtrace.h b/arch/tile/include/asm/backtrace.h
index 758ca4619d50..f18887d82399 100644
--- a/arch/tile/include/asm/backtrace.h
+++ b/arch/tile/include/asm/backtrace.h
@@ -146,7 +146,10 @@ enum {
 
 	CALLER_SP_IN_R52_BASE = 4,
 
-	CALLER_SP_OFFSET_BASE = 8
+	CALLER_SP_OFFSET_BASE = 8,
+
+	/* Marks the entry point of certain functions. */
+	ENTRY_POINT_INFO_OP = 16
 };
 
 
diff --git a/arch/tile/include/asm/compat.h b/arch/tile/include/asm/compat.h
index 8b60ec8b2d19..c3ae570c0a5d 100644
--- a/arch/tile/include/asm/compat.h
+++ b/arch/tile/include/asm/compat.h
@@ -216,15 +216,16 @@ struct compat_siginfo;
 struct compat_sigaltstack;
 long compat_sys_execve(const char __user *path,
 		       const compat_uptr_t __user *argv,
-		       const compat_uptr_t __user *envp);
+		       const compat_uptr_t __user *envp, struct pt_regs *);
 long compat_sys_rt_sigaction(int sig, struct compat_sigaction __user *act,
 			     struct compat_sigaction __user *oact,
 			     size_t sigsetsize);
 long compat_sys_rt_sigqueueinfo(int pid, int sig,
 				struct compat_siginfo __user *uinfo);
-long compat_sys_rt_sigreturn(void);
+long compat_sys_rt_sigreturn(struct pt_regs *);
 long compat_sys_sigaltstack(const struct compat_sigaltstack __user *uss_ptr,
-			    struct compat_sigaltstack __user *uoss_ptr);
+			    struct compat_sigaltstack __user *uoss_ptr,
+			    struct pt_regs *);
 long compat_sys_truncate64(char __user *filename, u32 dummy, u32 low, u32 high);
 long compat_sys_ftruncate64(unsigned int fd, u32 dummy, u32 low, u32 high);
 long compat_sys_pread64(unsigned int fd, char __user *ubuf, size_t count,
@@ -255,4 +256,12 @@ long tile_compat_sys_ptrace(compat_long_t request, compat_long_t pid,
 /* Tilera Linux syscalls that don't have "compat" versions. */
 #define compat_sys_flush_cache sys_flush_cache
 
+/* These are the intvec_64.S trampolines. */
+long _compat_sys_execve(const char __user *path,
+			const compat_uptr_t __user *argv,
+			const compat_uptr_t __user *envp);
+long _compat_sys_sigaltstack(const struct compat_sigaltstack __user *uss_ptr,
+			    struct compat_sigaltstack __user *uoss_ptr);
+long _compat_sys_rt_sigreturn(void);
+
 #endif /* _ASM_TILE_COMPAT_H */
diff --git a/arch/tile/include/asm/irqflags.h b/arch/tile/include/asm/irqflags.h
index a11d4837ee4d..641e4ff3d805 100644
--- a/arch/tile/include/asm/irqflags.h
+++ b/arch/tile/include/asm/irqflags.h
@@ -47,53 +47,53 @@
 	int __n = (n); \
 	int __mask = 1 << (__n & 0x1f); \
 	if (__n < 32) \
-		__insn_mtspr(SPR_INTERRUPT_MASK_SET_1_0, __mask); \
+		__insn_mtspr(SPR_INTERRUPT_MASK_SET_K_0, __mask); \
 	else \
-		__insn_mtspr(SPR_INTERRUPT_MASK_SET_1_1, __mask); \
+		__insn_mtspr(SPR_INTERRUPT_MASK_SET_K_1, __mask); \
 } while (0)
 #define interrupt_mask_reset(n) do { \
 	int __n = (n); \
 	int __mask = 1 << (__n & 0x1f); \
 	if (__n < 32) \
-		__insn_mtspr(SPR_INTERRUPT_MASK_RESET_1_0, __mask); \
+		__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K_0, __mask); \
 	else \
-		__insn_mtspr(SPR_INTERRUPT_MASK_RESET_1_1, __mask); \
+		__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K_1, __mask); \
 } while (0)
 #define interrupt_mask_check(n) ({ \
 	int __n = (n); \
 	(((__n < 32) ? \
-	 __insn_mfspr(SPR_INTERRUPT_MASK_1_0) : \
-	 __insn_mfspr(SPR_INTERRUPT_MASK_1_1)) \
+	 __insn_mfspr(SPR_INTERRUPT_MASK_K_0) : \
+	 __insn_mfspr(SPR_INTERRUPT_MASK_K_1)) \
 	  >> (__n & 0x1f)) & 1; \
 })
 #define interrupt_mask_set_mask(mask) do { \
 	unsigned long long __m = (mask); \
-	__insn_mtspr(SPR_INTERRUPT_MASK_SET_1_0, (unsigned long)(__m)); \
-	__insn_mtspr(SPR_INTERRUPT_MASK_SET_1_1, (unsigned long)(__m>>32)); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_SET_K_0, (unsigned long)(__m)); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_SET_K_1, (unsigned long)(__m>>32)); \
 } while (0)
 #define interrupt_mask_reset_mask(mask) do { \
 	unsigned long long __m = (mask); \
-	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_1_0, (unsigned long)(__m)); \
-	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_1_1, (unsigned long)(__m>>32)); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K_0, (unsigned long)(__m)); \
+	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K_1, (unsigned long)(__m>>32)); \
 } while (0)
 #else
 #define interrupt_mask_set(n) \
-	__insn_mtspr(SPR_INTERRUPT_MASK_SET_1, (1UL << (n)))
+	__insn_mtspr(SPR_INTERRUPT_MASK_SET_K, (1UL << (n)))
 #define interrupt_mask_reset(n) \
-	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_1, (1UL << (n)))
+	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K, (1UL << (n)))
 #define interrupt_mask_check(n) \
-	((__insn_mfspr(SPR_INTERRUPT_MASK_1) >> (n)) & 1)
+	((__insn_mfspr(SPR_INTERRUPT_MASK_K) >> (n)) & 1)
 #define interrupt_mask_set_mask(mask) \
-	__insn_mtspr(SPR_INTERRUPT_MASK_SET_1, (mask))
+	__insn_mtspr(SPR_INTERRUPT_MASK_SET_K, (mask))
 #define interrupt_mask_reset_mask(mask) \
-	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_1, (mask))
+	__insn_mtspr(SPR_INTERRUPT_MASK_RESET_K, (mask))
 #endif
 
 /*
  * The set of interrupts we want active if irqs are enabled.
  * Note that in particular, the tile timer interrupt comes and goes
  * from this set, since we have no other way to turn off the timer.
- * Likewise, INTCTRL_1 is removed and re-added during device
+ * Likewise, INTCTRL_K is removed and re-added during device
  * interrupts, as is the the hardwall UDN_FIREWALL interrupt.
  * We use a low bit (MEM_ERROR) as our sentinel value and make sure it
  * is always claimed as an "active interrupt" so we can query that bit
@@ -170,14 +170,14 @@ DECLARE_PER_CPU(unsigned long long, interrupts_enabled_mask);
 
 /* Return 0 or 1 to indicate whether interrupts are currently disabled. */
 #define IRQS_DISABLED(tmp)					\
-	mfspr   tmp, INTERRUPT_MASK_1;				\
+	mfspr   tmp, SPR_INTERRUPT_MASK_K;			\
 	andi    tmp, tmp, 1
 
 /* Load up a pointer to &interrupts_enabled_mask. */
 #define GET_INTERRUPTS_ENABLED_MASK_PTR(reg)			\
-	moveli reg, hw2_last(interrupts_enabled_mask); \
-	shl16insli reg, reg, hw1(interrupts_enabled_mask); \
-	shl16insli reg, reg, hw0(interrupts_enabled_mask); \
+	moveli reg, hw2_last(interrupts_enabled_mask);		\
+	shl16insli reg, reg, hw1(interrupts_enabled_mask);	\
+	shl16insli reg, reg, hw0(interrupts_enabled_mask);	\
 	add     reg, reg, tp
 
 /* Disable interrupts. */
@@ -185,18 +185,18 @@ DECLARE_PER_CPU(unsigned long long, interrupts_enabled_mask);
 	moveli  tmp0, hw2_last(LINUX_MASKABLE_INTERRUPTS);	\
 	shl16insli tmp0, tmp0, hw1(LINUX_MASKABLE_INTERRUPTS);	\
 	shl16insli tmp0, tmp0, hw0(LINUX_MASKABLE_INTERRUPTS);	\
-	mtspr   INTERRUPT_MASK_SET_1, tmp0
+	mtspr   SPR_INTERRUPT_MASK_SET_K, tmp0
 
 /* Disable ALL synchronous interrupts (used by NMI entry). */
 #define IRQ_DISABLE_ALL(tmp)					\
 	movei   tmp, -1;					\
-	mtspr   INTERRUPT_MASK_SET_1, tmp
+	mtspr   SPR_INTERRUPT_MASK_SET_K, tmp
 
 /* Enable interrupts. */
 #define IRQ_ENABLE(tmp0, tmp1)					\
 	GET_INTERRUPTS_ENABLED_MASK_PTR(tmp0);			\
 	ld      tmp0, tmp0;					\
-	mtspr   INTERRUPT_MASK_RESET_1, tmp0
+	mtspr   SPR_INTERRUPT_MASK_RESET_K, tmp0
 
 #else /* !__tilegx__ */
 
@@ -210,14 +210,14 @@ DECLARE_PER_CPU(unsigned long long, interrupts_enabled_mask);
  * (making the original code's write of the "high" mask word idempotent).
  */
 #define IRQS_DISABLED(tmp)					\
-	mfspr   tmp, INTERRUPT_MASK_1_0;			\
+	mfspr   tmp, SPR_INTERRUPT_MASK_K_0;			\
 	shri    tmp, tmp, INT_MEM_ERROR;			\
 	andi    tmp, tmp, 1
 
 /* Load up a pointer to &interrupts_enabled_mask. */
 #define GET_INTERRUPTS_ENABLED_MASK_PTR(reg)			\
-	moveli  reg, lo16(interrupts_enabled_mask);	\
-	auli    reg, reg, ha16(interrupts_enabled_mask);\
+	moveli  reg, lo16(interrupts_enabled_mask);		\
+	auli    reg, reg, ha16(interrupts_enabled_mask);	\
 	add     reg, reg, tp
 
 /* Disable interrupts. */
@@ -227,16 +227,16 @@ DECLARE_PER_CPU(unsigned long long, interrupts_enabled_mask);
 	 moveli tmp1, lo16(LINUX_MASKABLE_INTERRUPTS)		\
 	};							\
 	{							\
-	 mtspr  INTERRUPT_MASK_SET_1_0, tmp0;			\
+	 mtspr  SPR_INTERRUPT_MASK_SET_K_0, tmp0;		\
 	 auli   tmp1, tmp1, ha16(LINUX_MASKABLE_INTERRUPTS)	\
 	};							\
-	mtspr   INTERRUPT_MASK_SET_1_1, tmp1
+	mtspr   SPR_INTERRUPT_MASK_SET_K_1, tmp1
 
 /* Disable ALL synchronous interrupts (used by NMI entry). */
 #define IRQ_DISABLE_ALL(tmp)					\
 	movei   tmp, -1;					\
-	mtspr   INTERRUPT_MASK_SET_1_0, tmp;			\
-	mtspr   INTERRUPT_MASK_SET_1_1, tmp
+	mtspr   SPR_INTERRUPT_MASK_SET_K_0, tmp;		\
+	mtspr   SPR_INTERRUPT_MASK_SET_K_1, tmp
 
 /* Enable interrupts. */
 #define IRQ_ENABLE(tmp0, tmp1)					\
@@ -246,8 +246,8 @@ DECLARE_PER_CPU(unsigned long long, interrupts_enabled_mask);
 	 addi   tmp1, tmp0, 4					\
 	};							\
 	lw      tmp1, tmp1;					\
-	mtspr   INTERRUPT_MASK_RESET_1_0, tmp0;			\
-	mtspr   INTERRUPT_MASK_RESET_1_1, tmp1
+	mtspr   SPR_INTERRUPT_MASK_RESET_K_0, tmp0;		\
+	mtspr   SPR_INTERRUPT_MASK_RESET_K_1, tmp1
 #endif
 
 /*
diff --git a/arch/tile/include/asm/mman.h b/arch/tile/include/asm/mman.h
index 4c6811e3e8dc..81b8fc348d63 100644
--- a/arch/tile/include/asm/mman.h
+++ b/arch/tile/include/asm/mman.h
@@ -23,6 +23,7 @@
 #define MAP_POPULATE	0x0040		/* populate (prefault) pagetables */
 #define MAP_NONBLOCK	0x0080		/* do not block on IO */
 #define MAP_GROWSDOWN	0x0100		/* stack-like segment */
+#define MAP_STACK	MAP_GROWSDOWN	/* provide convenience alias */
 #define MAP_LOCKED	0x0200		/* pages are locked */
 #define MAP_NORESERVE	0x0400		/* don't check for reservations */
 #define MAP_DENYWRITE	0x0800		/* ETXTBSY */
diff --git a/arch/tile/include/asm/page.h b/arch/tile/include/asm/page.h
index 7d90641cf18d..7979a45430d3 100644
--- a/arch/tile/include/asm/page.h
+++ b/arch/tile/include/asm/page.h
@@ -199,17 +199,17 @@ static inline __attribute_const__ int get_order(unsigned long size)
  * If you want more physical memory than this then see the CONFIG_HIGHMEM
  * option in the kernel configuration.
  *
- * The top two 16MB chunks in the table below (VIRT and HV) are
- * unavailable to Linux.  Since the kernel interrupt vectors must live
- * at 0xfd000000, we map all of the bottom of RAM at this address with
- * a huge page table entry to minimize its ITLB footprint (as well as
- * at PAGE_OFFSET).  The last architected requirement is that user
- * interrupt vectors live at 0xfc000000, so we make that range of
- * memory available to user processes.  The remaining regions are sized
- * as shown; after the first four addresses, we show "typical" values,
- * since the actual addresses depend on kernel #defines.
+ * The top 16MB chunk in the table below is unavailable to Linux.  Since
+ * the kernel interrupt vectors must live at ether 0xfe000000 or 0xfd000000
+ * (depending on whether the kernel is at PL2 or Pl1), we map all of the
+ * bottom of RAM at this address with a huge page table entry to minimize
+ * its ITLB footprint (as well as at PAGE_OFFSET).  The last architected
+ * requirement is that user interrupt vectors live at 0xfc000000, so we
+ * make that range of memory available to user processes.  The remaining
+ * regions are sized as shown; the first four addresses use the PL 1
+ * values, and after that, we show "typical" values, since the actual
+ * addresses depend on kernel #defines.
  *
- * MEM_VIRT_INTRPT                 0xff000000
  * MEM_HV_INTRPT                   0xfe000000
  * MEM_SV_INTRPT (kernel code)     0xfd000000
  * MEM_USER_INTRPT (user vector)   0xfc000000
@@ -221,9 +221,14 @@ static inline __attribute_const__ int get_order(unsigned long size)
  */
 
 #define MEM_USER_INTRPT		_AC(0xfc000000, UL)
+#if CONFIG_KERNEL_PL == 1
 #define MEM_SV_INTRPT		_AC(0xfd000000, UL)
 #define MEM_HV_INTRPT		_AC(0xfe000000, UL)
-#define MEM_VIRT_INTRPT		_AC(0xff000000, UL)
+#else
+#define MEM_GUEST_INTRPT	_AC(0xfd000000, UL)
+#define MEM_SV_INTRPT		_AC(0xfe000000, UL)
+#define MEM_HV_INTRPT		_AC(0xff000000, UL)
+#endif
 
 #define INTRPT_SIZE		0x4000
 
diff --git a/arch/tile/include/asm/processor.h b/arch/tile/include/asm/processor.h
index ccd5f8425688..1747ff3946b2 100644
--- a/arch/tile/include/asm/processor.h
+++ b/arch/tile/include/asm/processor.h
@@ -328,18 +328,21 @@ extern int kdata_huge;
  * Note that assembly code assumes that USER_PL is zero.
  */
 #define USER_PL 0
-#define KERNEL_PL 1
+#if CONFIG_KERNEL_PL == 2
+#define GUEST_PL 1
+#endif
+#define KERNEL_PL CONFIG_KERNEL_PL
 
-/* SYSTEM_SAVE_1_0 holds the current cpu number ORed with ksp0. */
+/* SYSTEM_SAVE_K_0 holds the current cpu number ORed with ksp0. */
 #define CPU_LOG_MASK_VALUE 12
 #define CPU_MASK_VALUE ((1 << CPU_LOG_MASK_VALUE) - 1)
 #if CONFIG_NR_CPUS > CPU_MASK_VALUE
 # error Too many cpus!
 #endif
 #define raw_smp_processor_id() \
-	((int)__insn_mfspr(SPR_SYSTEM_SAVE_1_0) & CPU_MASK_VALUE)
+	((int)__insn_mfspr(SPR_SYSTEM_SAVE_K_0) & CPU_MASK_VALUE)
 #define get_current_ksp0() \
-	(__insn_mfspr(SPR_SYSTEM_SAVE_1_0) & ~CPU_MASK_VALUE)
+	(__insn_mfspr(SPR_SYSTEM_SAVE_K_0) & ~CPU_MASK_VALUE)
 #define next_current_ksp0(task) ({ \
 	unsigned long __ksp0 = task_ksp0(task); \
 	int __cpu = raw_smp_processor_id(); \
diff --git a/arch/tile/include/asm/ptrace.h b/arch/tile/include/asm/ptrace.h
index 4a02bb073979..ac6d343129d3 100644
--- a/arch/tile/include/asm/ptrace.h
+++ b/arch/tile/include/asm/ptrace.h
@@ -62,8 +62,8 @@ struct pt_regs {
 	pt_reg_t lr;		/* aliases regs[TREG_LR] */
 
 	/* Saved special registers. */
-	pt_reg_t pc;		/* stored in EX_CONTEXT_1_0 */
-	pt_reg_t ex1;		/* stored in EX_CONTEXT_1_1 (PL and ICS bit) */
+	pt_reg_t pc;		/* stored in EX_CONTEXT_K_0 */
+	pt_reg_t ex1;		/* stored in EX_CONTEXT_K_1 (PL and ICS bit) */
 	pt_reg_t faultnum;	/* fault number (INT_SWINT_1 for syscall) */
 	pt_reg_t orig_r0;	/* r0 at syscall entry, else zero */
 	pt_reg_t flags;		/* flags (see below) */
diff --git a/arch/tile/include/asm/syscalls.h b/arch/tile/include/asm/syscalls.h
index ce99ffefeacf..3b5507c31eae 100644
--- a/arch/tile/include/asm/syscalls.h
+++ b/arch/tile/include/asm/syscalls.h
@@ -32,8 +32,9 @@ extern void *compat_sys_call_table[];
 
 /*
  * Note that by convention, any syscall which requires the current
- * register set takes an additional "struct pt_regs *" pointer; the
- * sys_xxx() function just adds the pointer and tail-calls to _sys_xxx().
+ * register set takes an additional "struct pt_regs *" pointer; a
+ * _sys_xxx() trampoline in intvec*.S just sets up the pointer and
+ * jumps to sys_xxx().
  */
 
 /* kernel/sys.c */
@@ -43,66 +44,17 @@ long sys32_fadvise64(int fd, u32 offset_lo, u32 offset_hi,
 int sys32_fadvise64_64(int fd, u32 offset_lo, u32 offset_hi,
 		       u32 len_lo, u32 len_hi, int advice);
 long sys_flush_cache(void);
-long sys_mmap2(unsigned long addr, unsigned long len,
-	       unsigned long prot, unsigned long flags,
-	       unsigned long fd, unsigned long pgoff);
-#ifdef __tilegx__
-long sys_mmap(unsigned long addr, unsigned long len,
-	      unsigned long prot, unsigned long flags,
-	      unsigned long fd, off_t pgoff);
+#ifndef __tilegx__  /* No mmap() in the 32-bit kernel. */
+#define sys_mmap sys_mmap
 #endif
 
-/* kernel/process.c */
-long sys_clone(unsigned long clone_flags, unsigned long newsp,
-	       void __user *parent_tid, void __user *child_tid);
-long _sys_clone(unsigned long clone_flags, unsigned long newsp,
-		void __user *parent_tid, void __user *child_tid,
-		struct pt_regs *regs);
-long sys_fork(void);
-long _sys_fork(struct pt_regs *regs);
-long sys_vfork(void);
-long _sys_vfork(struct pt_regs *regs);
-long sys_execve(const char __user *filename,
-		const char __user *const __user *argv,
-		const char __user *const __user *envp);
-long _sys_execve(const char __user *filename,
-		 const char __user *const __user *argv,
-		 const char __user *const __user *envp, struct pt_regs *regs);
-
-/* kernel/signal.c */
-long sys_sigaltstack(const stack_t __user *, stack_t __user *);
-long _sys_sigaltstack(const stack_t __user *, stack_t __user *,
-		      struct pt_regs *);
-long sys_rt_sigreturn(void);
-long _sys_rt_sigreturn(struct pt_regs *regs);
-
-/* platform-independent functions */
-long sys_rt_sigsuspend(sigset_t __user *unewset, size_t sigsetsize);
-long sys_rt_sigaction(int sig, const struct sigaction __user *act,
-		      struct sigaction __user *oact, size_t sigsetsize);
-
 #ifndef __tilegx__
 /* mm/fault.c */
-int sys_cmpxchg_badaddr(unsigned long address);
-int _sys_cmpxchg_badaddr(unsigned long address, struct pt_regs *);
+long sys_cmpxchg_badaddr(unsigned long address, struct pt_regs *);
+long _sys_cmpxchg_badaddr(unsigned long address);
 #endif
 
 #ifdef CONFIG_COMPAT
-long compat_sys_execve(const char __user *path,
-		       const compat_uptr_t __user *argv,
-		       const compat_uptr_t __user *envp);
-long _compat_sys_execve(const char __user *path,
-			const compat_uptr_t __user *argv,
-			const compat_uptr_t __user *envp,
-			struct pt_regs *regs);
-long compat_sys_sigaltstack(const struct compat_sigaltstack __user *uss_ptr,
-			    struct compat_sigaltstack __user *uoss_ptr);
-long _compat_sys_sigaltstack(const struct compat_sigaltstack __user *uss_ptr,
-			     struct compat_sigaltstack __user *uoss_ptr,
-			     struct pt_regs *regs);
-long compat_sys_rt_sigreturn(void);
-long _compat_sys_rt_sigreturn(struct pt_regs *regs);
-
 /* These four are not defined for 64-bit, but serve as "compat" syscalls. */
 long sys_fcntl64(unsigned int fd, unsigned int cmd, unsigned long arg);
 long sys_fstat64(unsigned long fd, struct stat64 __user *statbuf);
@@ -110,4 +62,15 @@ long sys_truncate64(const char __user *path, loff_t length);
 long sys_ftruncate64(unsigned int fd, loff_t length);
 #endif
 
+/* These are the intvec*.S trampolines. */
+long _sys_sigaltstack(const stack_t __user *, stack_t __user *);
+long _sys_rt_sigreturn(void);
+long _sys_clone(unsigned long clone_flags, unsigned long newsp,
+		void __user *parent_tid, void __user *child_tid);
+long _sys_execve(const char __user *filename,
+		 const char __user *const __user *argv,
+		 const char __user *const __user *envp);
+
+#include <asm-generic/syscalls.h>
+
 #endif /* _ASM_TILE_SYSCALLS_H */
diff --git a/arch/tile/include/asm/system.h b/arch/tile/include/asm/system.h
index f749be327ce0..5388850deeb2 100644
--- a/arch/tile/include/asm/system.h
+++ b/arch/tile/include/asm/system.h
@@ -89,6 +89,10 @@
 #define get_cycles_low() __insn_mfspr(SPR_CYCLE)   /* just get all 64 bits */
 #endif
 
+#if !CHIP_HAS_MF_WAITS_FOR_VICTIMS()
+int __mb_incoherent(void);  /* Helper routine for mb_incoherent(). */
+#endif
+
 /* Fence to guarantee visibility of stores to incoherent memory. */
 static inline void
 mb_incoherent(void)
@@ -97,7 +101,6 @@ mb_incoherent(void)
 
 #if !CHIP_HAS_MF_WAITS_FOR_VICTIMS()
 	{
-		int __mb_incoherent(void);
 #if CHIP_HAS_TILE_WRITE_PENDING()
 		const unsigned long WRITE_TIMEOUT_CYCLES = 400;
 		unsigned long start = get_cycles_low();
@@ -161,7 +164,7 @@ extern struct task_struct *_switch_to(struct task_struct *prev,
 /* Helper function for _switch_to(). */
 extern struct task_struct *__switch_to(struct task_struct *prev,
 				       struct task_struct *next,
-				       unsigned long new_system_save_1_0);
+				       unsigned long new_system_save_k_0);
 
 /* Address that switched-away from tasks are at. */
 extern unsigned long get_switch_to_pc(void);
@@ -214,13 +217,6 @@ int hardwall_deactivate(struct task_struct *task);
 } while (0)
 #endif
 
-/* Invoke the simulator "syscall" mechanism (see arch/tile/kernel/entry.S). */
-extern int _sim_syscall(int syscall_num, ...);
-#define sim_syscall(syscall_num, ...) \
-	_sim_syscall(SIM_CONTROL_SYSCALL + \
-		((syscall_num) << _SIM_CONTROL_OPERATOR_BITS), \
-		## __VA_ARGS__)
-
 /*
  * Kernel threads can check to see if they need to migrate their
  * stack whenever they return from a context switch; for user
diff --git a/arch/tile/include/asm/traps.h b/arch/tile/include/asm/traps.h
index 432a9c15c8a2..d06e35f57201 100644
--- a/arch/tile/include/asm/traps.h
+++ b/arch/tile/include/asm/traps.h
@@ -59,4 +59,8 @@ void do_hardwall_trap(struct pt_regs *, int fault_num);
 void do_breakpoint(struct pt_regs *, int fault_num);
 
 
+#ifdef __tilegx__
+void gx_singlestep_handle(struct pt_regs *, int fault_num);
+#endif
+
 #endif /* _ASM_TILE_SYSCALLS_H */
diff --git a/arch/tile/include/hv/hypervisor.h b/arch/tile/include/hv/hypervisor.h
index 9bd303a141b2..f672544cd4f9 100644
--- a/arch/tile/include/hv/hypervisor.h
+++ b/arch/tile/include/hv/hypervisor.h
@@ -1003,37 +1003,37 @@ int hv_console_write(HV_VirtAddr bytes, int len);
  *  when these occur in a client's interrupt critical section, they must
  *  be delivered through the downcall mechanism.
  *
- *  A downcall is initially delivered to the client as an INTCTRL_1
- *  interrupt.  Upon entry to the INTCTRL_1 vector, the client must
- *  immediately invoke the hv_downcall_dispatch service.  This service
- *  will not return; instead it will cause one of the client's actual
- *  downcall-handling interrupt vectors to be entered.  The EX_CONTEXT
- *  registers in the client will be set so that when the client irets,
- *  it will return to the code which was interrupted by the INTCTRL_1
- *  interrupt.
- *
- *  Under some circumstances, the firing of INTCTRL_1 can race with
+ *  A downcall is initially delivered to the client as an INTCTRL_CL
+ *  interrupt, where CL is the client's PL.  Upon entry to the INTCTRL_CL
+ *  vector, the client must immediately invoke the hv_downcall_dispatch
+ *  service.  This service will not return; instead it will cause one of
+ *  the client's actual downcall-handling interrupt vectors to be entered.
+ *  The EX_CONTEXT registers in the client will be set so that when the
+ *  client irets, it will return to the code which was interrupted by the
+ *  INTCTRL_CL interrupt.
+ *
+ *  Under some circumstances, the firing of INTCTRL_CL can race with
  *  the lowering of a device interrupt.  In such a case, the
  *  hv_downcall_dispatch service may issue an iret instruction instead
  *  of entering one of the client's actual downcall-handling interrupt
  *  vectors.  This will return execution to the location that was
- *  interrupted by INTCTRL_1.
+ *  interrupted by INTCTRL_CL.
  *
  *  Any saving of registers should be done by the actual handling
- *  vectors; no registers should be changed by the INTCTRL_1 handler.
+ *  vectors; no registers should be changed by the INTCTRL_CL handler.
  *  In particular, the client should not use a jal instruction to invoke
  *  the hv_downcall_dispatch service, as that would overwrite the client's
  *  lr register.  Note that the hv_downcall_dispatch service may overwrite
  *  one or more of the client's system save registers.
  *
- *  The client must not modify the INTCTRL_1_STATUS SPR.  The hypervisor
+ *  The client must not modify the INTCTRL_CL_STATUS SPR.  The hypervisor
  *  will set this register to cause a downcall to happen, and will clear
  *  it when no further downcalls are pending.
  *
- *  When a downcall vector is entered, the INTCTRL_1 interrupt will be
+ *  When a downcall vector is entered, the INTCTRL_CL interrupt will be
  *  masked.  When the client is done processing a downcall, and is ready
  *  to accept another, it must unmask this interrupt; if more downcalls
- *  are pending, this will cause the INTCTRL_1 vector to be reentered.
+ *  are pending, this will cause the INTCTRL_CL vector to be reentered.
  *  Currently the following interrupt vectors can be entered through a
  *  downcall:
  *