#include "mini-gc.h"
#include <mono/metadata/gc-internal.h>
-/*
- * The code below does not work yet, and probably needs to be thrown out if we move
- * to GC safe points.
- */
-
//#if 0
-#ifdef HAVE_SGEN_GC
+#if defined(MONO_ARCH_GC_MAPS_SUPPORTED)
#include <mono/metadata/gc-internal.h>
#include <mono/utils/mono-counters.h>
+#if SIZEOF_VOID_P == 4
+typedef guint32 mword;
+#else
+typedef guint64 mword;
+#endif
+
+#define GC_BITS_PER_WORD (sizeof (mword) * 8)
+
/* Contains state needed by the GC Map construction code */
typedef struct {
/*
/* Number of slots in the map */
int nslots;
- /* The type of the slots */
- StackSlotType *slots;
- /* Live intervals for every slot */
- GSList **live_intervals;
- /* Whenever the slot starts out as SLOT_PIN, then changes to SLOT_REF */
- gboolean *starts_pinned;
/* The number of registers in the map */
int nregs;
- /*
- * GC Type of registers.
- * Registers might be shared between refs and non-refs, so we store a GC type
- * for each live interval.
- * FIXME: Do the same for slots too, i.e. make 'slots' a list.
- * FIXME: Add a struct for the type + interval pair.
- */
- GSList **reg_types;
- /*
- * Live intervals for registers.
- * This has width MONO_MAX_IREGS.
- * FIXME: Only store callee saved regs.
- */
- GSList **reg_live_intervals;
/* Min and Max offsets of the stack frame relative to fp */
int min_offset, max_offset;
/* Same for the locals area */
int locals_min_offset, locals_max_offset;
+
+ /* The call sites where this frame can be stopped during GC */
+ GCCallSite **callsites;
+ /* The number of call sites */
+ int ncallsites;
+
+ /*
+ * The width of the stack bitmaps in bytes. This is not equal to the bitmap width at
+ * runtime, since it includes columns which are 0.
+ */
+ int stack_bitmap_width;
+ /*
+ * A bitmap whose width equals nslots, and whose height equals ncallsites.
+ * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
+ * given callsite.
+ */
+ guint8 *stack_ref_bitmap;
+ /* Same for SLOT_PIN */
+ guint8 *stack_pin_bitmap;
+
+ /*
+ * Similar bitmaps for registers. These have width MONO_MAX_IREGS in bits.
+ */
+ int reg_bitmap_width;
+ guint8 *reg_ref_bitmap;
+ guint8 *reg_pin_bitmap;
} MonoCompileGC;
-#define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
+#define ALIGN_TO(val,align) ((((mgreg_t)val) + ((align) - 1)) & ~((align) - 1))
+
+#undef DEBUG
#if 0
-#define DEBUG(s) do { s; } while (0)
+/* We don't support debug levels, its all-or-nothing */
+#define DEBUG(s) do { s; fflush (logfile); } while (0)
+#define DEBUG_ENABLED 1
#else
#define DEBUG(s)
#endif
-#if 1
-#define DEBUG_GC_MAP(s) do { s; fflush (stdout); } while (0)
+#ifdef DEBUG_ENABLED
+//#if 1
+#define DEBUG_PRECISE(s) do { s; } while (0)
+#define DEBUG_PRECISE_ENABLED
#else
-#define DEBUG_GC_MAP(s)
+#define DEBUG_PRECISE(s)
+#endif
+
+/*
+ * Contains information collected during the conservative stack marking pass,
+ * used during the precise pass. This helps to avoid doing a stack walk twice, which
+ * is expensive.
+ */
+typedef struct {
+ guint8 *bitmap;
+ int nslots;
+ int frame_start_offset;
+ int nreg_locations;
+ /* Relative to stack_start */
+ int reg_locations [MONO_MAX_IREGS];
+#ifdef DEBUG_PRECISE_ENABLED
+ MonoJitInfo *ji;
+ gpointer fp;
+ int regs [MONO_MAX_IREGS];
#endif
+} FrameInfo;
-#define GC_BITS_PER_WORD (sizeof (gsize) * 8)
+/* Max number of frames stored in the TLS data */
+#define MAX_FRAMES 50
/*
* Per-thread data kept by this module. This is stored in the GC and passed to us as
MonoContext ctx;
gboolean has_context;
MonoJitTlsData *jit_tls;
+ /* For debugging */
+ mgreg_t tid;
+ gpointer ref_to_track;
+ /* Number of frames collected during the !precise pass */
+ int nframes;
+ FrameInfo frames [MAX_FRAMES];
} TlsData;
+/* These are constant so don't store them in the GC Maps */
+/* Number of registers stored in gc maps */
+#define NREGS MONO_MAX_IREGS
+
/*
+ * The GC Map itself.
* Contains information needed to mark a stack frame.
- * FIXME: Optimize the memory usage.
+ * This is a transient structure, created from a compressed representation on-demand.
*/
typedef struct {
- /* The frame pointer register */
- int frame_reg;
- /* The offset of the GC tracked area inside the stack frame relative to the frame pointer */
- int frame_offset;
- /* The number of stack slots */
+ /*
+ * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
+ * This includes memory which is NOREF thus doesn't need GC maps.
+ */
+ int start_offset;
+ int end_offset;
+ /*
+ * The offset relative to frame_offset where the the memory described by the GC maps
+ * begins.
+ */
+ int map_offset;
+ /* The number of stack slots in the map */
int nslots;
- /* The number of registers in the map */
- int nregs;
- /* Thw width of the stack bitmap in bytes */
- int bitmap_width;
- /* Thw width of the register bitmap in bytes */
- int reg_bitmap_width;
+ /* The frame pointer register */
+ guint8 frame_reg;
+ /* The size of each callsite table entry */
+ guint8 callsite_entry_size;
guint has_pin_slots : 1;
guint has_ref_slots : 1;
guint has_ref_regs : 1;
guint has_pin_regs : 1;
+
+ /* The offsets below are into an external bitmaps array */
+
/*
- * A bitmap whose width is equal to bitmap_width, and whose
- * height is equal to the number of possible PC offsets.
+ * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
* The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
- * given pc offset.
- * FIXME: Compress this.
- * FIXME: Embed this after the structure.
+ * given callsite.
*/
- guint8 *ref_bitmap;
+ guint32 stack_ref_bitmap_offset;
/*
* Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
- * different pc offsets, if the slot starts out as PIN, and later changes to REF.
+ * different callsites, if the slot starts out as PIN, and later changes to REF.
*/
- guint8 *pin_bitmap;
+ guint32 stack_pin_bitmap_offset;
/*
* Corresponding bitmaps for registers
- * These have width MONO_MAX_IREGS in bits.
+ * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
* FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
*/
- guint8 *reg_pin_bitmap;
- guint8 *reg_ref_bitmap;
+ guint32 reg_pin_bitmap_offset;
+ guint32 reg_ref_bitmap_offset;
+
+ guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
- /* The registers used by the method */
- guint64 used_regs;
+ /* The number of bits set in the two masks above */
+ guint8 nref_regs, npin_regs;
/*
* A bit array marking slots which contain refs.
* This is used only for debugging.
*/
- guint8 *ref_slots;
+ //guint8 *ref_slots;
+
+ /* Callsite offsets */
+ /* These can take up a lot of space, so encode them compactly */
+ union {
+ guint8 *offsets8;
+ guint16 *offsets16;
+ guint32 *offsets32;
+ } callsites;
+ int ncallsites;
} GCMap;
-/* Statistics */
-static guint32 gc_maps_size;
+/*
+ * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
+ */
+typedef struct {
+ //guint8 *ref_slots;
+ //guint8 encoded_size;
+
+ /*
+ * The arrays below are embedded after the struct.
+ * Their address needs to be computed.
+ */
+
+ /* The fixed fields of the GCMap encoded using LEB128 */
+ guint8 encoded [MONO_ZERO_LEN_ARRAY];
+
+ /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
+ guint8 callsites [MONO_ZERO_LEN_ARRAY];
+
+ /* The GC bitmaps */
+ guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
+} GCEncodedMap;
+
+static int precise_frame_count [2], precise_frame_limit = -1;
+static gboolean precise_frame_limit_inited;
+
+/* Stats */
+typedef struct {
+ int scanned_stacks;
+ int scanned;
+ int scanned_precisely;
+ int scanned_conservatively;
+ int scanned_registers;
+ int scanned_native;
+ int scanned_other;
+
+ int all_slots;
+ int noref_slots;
+ int ref_slots;
+ int pin_slots;
+
+ int gc_maps_size;
+ int gc_callsites_size;
+ int gc_callsites8_size;
+ int gc_callsites16_size;
+ int gc_callsites32_size;
+ int gc_bitmaps_size;
+ int gc_map_struct_size;
+ int tlsdata_size;
+} JITGCStats;
+
+static JITGCStats stats;
+
+static FILE *logfile;
+
+// FIXME: Move these to a shared place
+
+static inline void
+encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
+{
+ guint8 *p = buf;
+
+ do {
+ guint8 b = value & 0x7f;
+ value >>= 7;
+ if (value != 0) /* more bytes to come */
+ b |= 0x80;
+ *p ++ = b;
+ } while (value);
+
+ *endbuf = p;
+}
+
+static G_GNUC_UNUSED void
+encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
+{
+ gboolean more = 1;
+ gboolean negative = (value < 0);
+ guint32 size = 32;
+ guint8 byte;
+ guint8 *p = buf;
+
+ while (more) {
+ byte = value & 0x7f;
+ value >>= 7;
+ /* the following is unnecessary if the
+ * implementation of >>= uses an arithmetic rather
+ * than logical shift for a signed left operand
+ */
+ if (negative)
+ /* sign extend */
+ value |= - (1 <<(size - 7));
+ /* sign bit of byte is second high order bit (0x40) */
+ if ((value == 0 && !(byte & 0x40)) ||
+ (value == -1 && (byte & 0x40)))
+ more = 0;
+ else
+ byte |= 0x80;
+ *p ++= byte;
+ }
+
+ *endbuf = p;
+}
+
+static inline guint32
+decode_uleb128 (guint8 *buf, guint8 **endbuf)
+{
+ guint8 *p = buf;
+ guint32 res = 0;
+ int shift = 0;
+
+ while (TRUE) {
+ guint8 b = *p;
+ p ++;
+
+ res = res | (((int)(b & 0x7f)) << shift);
+ if (!(b & 0x80))
+ break;
+ shift += 7;
+ }
+
+ *endbuf = p;
+
+ return res;
+}
+
+static inline gint32
+decode_sleb128 (guint8 *buf, guint8 **endbuf)
+{
+ guint8 *p = buf;
+ gint32 res = 0;
+ int shift = 0;
+
+ while (TRUE) {
+ guint8 b = *p;
+ p ++;
+
+ res = res | (((int)(b & 0x7f)) << shift);
+ shift += 7;
+ if (!(b & 0x80)) {
+ if (shift < 32 && (b & 0x40))
+ res |= - (1 << shift);
+ break;
+ }
+ }
+
+ *endbuf = p;
+
+ return res;
+}
+
+static int
+encode_frame_reg (int frame_reg)
+{
+#ifdef TARGET_AMD64
+ if (frame_reg == AMD64_RSP)
+ return 0;
+ else if (frame_reg == AMD64_RBP)
+ return 1;
+#elif defined(TARGET_X86)
+ if (frame_reg == X86_EBP)
+ return 0;
+ else if (frame_reg == X86_ESP)
+ return 1;
+#else
+ NOT_IMPLEMENTED;
+#endif
+ g_assert_not_reached ();
+ return -1;
+}
+
+static int
+decode_frame_reg (int encoded)
+{
+#ifdef TARGET_AMD64
+ if (encoded == 0)
+ return AMD64_RSP;
+ else if (encoded == 1)
+ return AMD64_RBP;
+#elif defined(TARGET_X86)
+ if (encoded == 0)
+ return X86_EBP;
+ else if (encoded == 1)
+ return X86_ESP;
+#else
+ NOT_IMPLEMENTED;
+#endif
+ g_assert_not_reached ();
+ return -1;
+}
+
+#ifdef TARGET_AMD64
+#ifdef HOST_WIN32
+static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
+#else
+static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
+#endif
+#elif defined(TARGET_X86)
+static int callee_saved_regs [] = { X86_EBX, X86_ESI, X86_EDI };
+#endif
+
+static guint32
+encode_regmask (guint32 regmask)
+{
+ int i;
+ guint32 res;
+
+ res = 0;
+ for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
+ if (regmask & (1 << callee_saved_regs [i])) {
+ res |= (1 << i);
+ regmask -= (1 << callee_saved_regs [i]);
+ }
+ }
+ g_assert (regmask == 0);
+ return res;
+}
+
+static guint32
+decode_regmask (guint32 regmask)
+{
+ int i;
+ guint32 res;
+
+ res = 0;
+ for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
+ if (regmask & (1 << i))
+ res |= (1 << callee_saved_regs [i]);
+ return res;
+}
+
+/*
+ * encode_gc_map:
+ *
+ * Encode the fixed fields of MAP into a buffer pointed to by BUF.
+ */
+static void
+encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
+{
+ guint32 flags, freg;
+
+ encode_sleb128 (map->start_offset / sizeof (mgreg_t), buf, &buf);
+ encode_sleb128 (map->end_offset / sizeof (mgreg_t), buf, &buf);
+ encode_sleb128 (map->map_offset / sizeof (mgreg_t), buf, &buf);
+ encode_uleb128 (map->nslots, buf, &buf);
+ g_assert (map->callsite_entry_size <= 4);
+ freg = encode_frame_reg (map->frame_reg);
+ g_assert (freg < 2);
+ flags = (map->has_ref_slots ? 1 : 0) | (map->has_pin_slots ? 2 : 0) | (map->has_ref_regs ? 4 : 0) | (map->has_pin_regs ? 8 : 0) | ((map->callsite_entry_size - 1) << 4) | (freg << 6);
+ encode_uleb128 (flags, buf, &buf);
+ encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
+ if (map->has_ref_regs)
+ encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
+ if (map->has_pin_regs)
+ encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
+ encode_uleb128 (map->ncallsites, buf, &buf);
+
+ *endbuf = buf;
+}
+
+/*
+ * decode_gc_map:
+ *
+ * Decode the encoded GC map representation in BUF and store the result into MAP.
+ */
+static void
+decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
+{
+ guint32 flags;
+ int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
+ int i, n;
+
+ map->start_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
+ map->end_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
+ map->map_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
+ map->nslots = decode_uleb128 (buf, &buf);
+ flags = decode_uleb128 (buf, &buf);
+ map->has_ref_slots = (flags & 1) ? 1 : 0;
+ map->has_pin_slots = (flags & 2) ? 1 : 0;
+ map->has_ref_regs = (flags & 4) ? 1 : 0;
+ map->has_pin_regs = (flags & 8) ? 1 : 0;
+ map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
+ freg = flags >> 6;
+ map->frame_reg = decode_frame_reg (freg);
+ map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
+ if (map->has_ref_regs) {
+ map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
+ n = 0;
+ for (i = 0; i < NREGS; ++i)
+ if (map->reg_ref_mask & (1 << i))
+ n ++;
+ map->nref_regs = n;
+ }
+ if (map->has_pin_regs) {
+ map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
+ n = 0;
+ for (i = 0; i < NREGS; ++i)
+ if (map->reg_pin_mask & (1 << i))
+ n ++;
+ map->npin_regs = n;
+ }
+ map->ncallsites = decode_uleb128 (buf, &buf);
+
+ stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
+ reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
+ reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
+ offset = 0;
+ map->stack_ref_bitmap_offset = offset;
+ if (map->has_ref_slots)
+ offset += stack_bitmap_size;
+ map->stack_pin_bitmap_offset = offset;
+ if (map->has_pin_slots)
+ offset += stack_bitmap_size;
+ map->reg_ref_bitmap_offset = offset;
+ if (map->has_ref_regs)
+ offset += reg_ref_bitmap_size;
+ map->reg_pin_bitmap_offset = offset;
+ if (map->has_pin_regs)
+ offset += reg_pin_bitmap_size;
+
+ *endbuf = buf;
+}
static gpointer
thread_attach_func (void)
{
- return g_new0 (TlsData, 1);
+ TlsData *tls;
+
+ tls = g_new0 (TlsData, 1);
+ tls->tid = GetCurrentThreadId ();
+ stats.tlsdata_size += sizeof (TlsData);
+
+ return tls;
+}
+
+static void
+thread_detach_func (gpointer user_data)
+{
+ TlsData *tls = user_data;
+
+ g_free (tls);
}
static void
tls->jit_tls = TlsGetValue (mono_jit_tls_id);
}
-static int precise_frame_count [2], precise_frame_limit = -1;
-static gboolean precise_frame_limit_inited;
-
-/* Stats */
-typedef struct {
- int scanned_stacks;
- int scanned;
- int scanned_precisely;
- int scanned_conservatively;
- int scanned_registers;
-
- int all_slots;
- int noref_slots;
- int ref_slots;
- int pin_slots;
-} JITGCStats;
-
-static JITGCStats stats;
-
#define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
static inline void
}
static const char*
-slot_type_to_string (StackSlotType type)
+slot_type_to_string (GCSlotType type)
{
switch (type) {
case SLOT_REF:
}
}
+static inline mgreg_t
+get_frame_pointer (MonoContext *ctx, int frame_reg)
+{
+#if defined(TARGET_AMD64)
+ if (frame_reg == AMD64_RSP)
+ return ctx->rsp;
+ else if (frame_reg == AMD64_RBP)
+ return ctx->rbp;
+#elif defined(TARGET_X86)
+ if (frame_reg == X86_ESP)
+ return ctx->esp;
+ else if (frame_reg == X86_EBP)
+ return ctx->ebp;
+#endif
+ g_assert_not_reached ();
+ return 0;
+}
+
/*
- * thread_mark_func:
+ * conservatively_pass:
*
- * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
- * call, and TRUE at the second. USER_DATA points to a TlsData
- * structure filled up by thread_suspend_func.
+ * Mark a thread stack conservatively and collect information needed by the precise pass.
*/
static void
-thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
+conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
{
- TlsData *tls = user_data;
MonoJitInfo *ji;
MonoContext ctx, new_ctx;
MonoLMF *lmf;
guint8 *stack_limit;
gboolean last = TRUE;
GCMap *map;
- guint8* fp, *frame_start, *frame_end;
- int i, pc_offset;
+ GCMap map_tmp;
+ GCEncodedMap *emap;
+ guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
+ int i, pc_offset, cindex, bitmap_width;
int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
gboolean res;
StackFrameInfo frame;
mgreg_t *reg_locations [MONO_MAX_IREGS];
mgreg_t *new_reg_locations [MONO_MAX_IREGS];
+ guint8 *bitmaps;
+ FrameInfo *fi;
+ guint32 precise_regmask;
+
+ if (tls) {
+ tls->nframes = 0;
+ tls->ref_to_track = NULL;
+ }
/* tls == NULL can happen during startup */
if (mono_thread_internal_current () == NULL || !tls) {
- if (!precise) {
- mono_gc_conservatively_scan_area (stack_start, stack_end);
- stats.scanned_stacks += stack_end - stack_start;
- }
+ mono_gc_conservatively_scan_area (stack_start, stack_end);
+ stats.scanned_stacks += stack_end - stack_start;
return;
}
/* This is one past the last address which we have scanned */
stack_limit = stack_start;
- DEBUG (printf ("*** %s stack marking %p-%p ***\n", precise ? "Precise" : "Conservative", stack_start, stack_end));
-
if (!tls->has_context)
memset (&new_ctx, 0, sizeof (ctx));
else
/*
* If the current frame saves the register, it means it might modify its
* value, thus the old location might not contain the same value, so
- * we have to mark it conservatively. If we have precise info about the
- * register, reg_locations [i] is already cleared.
+ * we have to mark it conservatively.
*/
- if (!precise && reg_locations [i]) {
- DEBUG (printf ("\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
+ if (reg_locations [i]) {
+ DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
- // FIXME: This is not correct because the location might be in a frame
- // without a GC map
- // Use a separate stat for now
- //scanned_conservatively += sizeof (mgreg_t);
scanned_registers += sizeof (mgreg_t);
}
reg_locations [i] = new_reg_locations [i];
- if (!precise) {
- DEBUG (printf ("\treg %s is at location %p.\n", mono_arch_regname (i), reg_locations [i]));
- }
+ DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
}
}
- g_assert ((guint64)stack_limit % sizeof (mgreg_t) == 0);
+ g_assert ((mgreg_t)stack_limit % sizeof (mgreg_t) == 0);
-#ifdef MONO_ARCH_HAVE_FIND_JIT_INFO_EXT
res = mono_find_jit_info_ext (frame.domain ? frame.domain : mono_domain_get (), tls->jit_tls, NULL, &ctx, &new_ctx, NULL, &lmf, new_reg_locations, &frame);
if (!res)
break;
-#else
- break;
-#endif
+
+ ji = frame.ji;
+
+ if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
+ /*
+ * These frames are problematic for several reasons:
+ * - they are unwound through an LMF, and we have no precise register tracking for those.
+ * - the LMF might not contain a precise ip, so we can't compute the call site.
+ * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
+ */
+ DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
+ for (i = 0; i < MONO_MAX_IREGS; ++i) {
+ if (reg_locations [i]) {
+ DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
+ mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
+ scanned_registers += sizeof (mgreg_t);
+ }
+ reg_locations [i] = NULL;
+ new_reg_locations [i] = NULL;
+ }
+ ctx = new_ctx;
+ continue;
+ }
/* The last frame can be in any state so mark conservatively */
if (last) {
+ if (ji) {
+ DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); fprintf (logfile, "Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
+ }
+ DEBUG (fprintf (logfile, "\t <Last frame>\n"));
last = FALSE;
continue;
}
- /* These frames are returned by mono_find_jit_info () two times */
- if (!frame.managed)
+ pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
+
+ /* These frames are very problematic */
+ if (ji->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
+ DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); fprintf (logfile, "Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
+ DEBUG (fprintf (logfile, "\tSkip.\n"));
continue;
+ }
/* All the other frames are at a call site */
+ if (tls->nframes == MAX_FRAMES) {
+ /*
+ * Can't save information since the array is full. So scan the rest of the
+ * stack conservatively.
+ */
+ DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
+ break;
+ }
+
/* Scan the frame of this method */
/*
* - locals
* - spill area
* - localloc-ed memory
- * Currently, only the locals/args are scanned precisely.
*/
+ g_assert (pc_offset >= 0);
- ji = frame.ji;
- map = ji->gc_info;
-
- if (!map) {
- DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); printf ("Mark(%d): No GC map for %s\n", precise, fname); g_free (fname));
+ emap = ji->gc_info;
+ if (!emap) {
+ DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); fprintf (logfile, "Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
+ DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
continue;
}
+ /* The embedded callsite table requires this */
+ g_assert (((mgreg_t)emap % 4) == 0);
+
/*
* Debugging aid to control the number of frames scanned precisely
*/
}
if (precise_frame_limit != -1) {
- if (precise_frame_count [precise] == precise_frame_limit)
+ if (precise_frame_count [FALSE] == precise_frame_limit)
printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (ji->method, TRUE));
- if (precise_frame_count [precise] > precise_frame_limit)
+ if (precise_frame_count [FALSE] > precise_frame_limit)
continue;
}
- precise_frame_count [precise] ++;
-
-#ifdef __x86_64__
- if (map->frame_reg == AMD64_RSP)
- fp = (guint8*)ctx.rsp;
- else if (map->frame_reg == AMD64_RBP)
- fp = (guint8*)ctx.rbp;
- else
+ precise_frame_count [FALSE] ++;
+
+ /* Decode the encoded GC map */
+ map = &map_tmp;
+ memset (map, 0, sizeof (GCMap));
+ decode_gc_map (&emap->encoded [0], map, &p);
+ p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
+ map->callsites.offsets8 = p;
+ p += map->callsite_entry_size * map->ncallsites;
+ bitmaps = p;
+
+ fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
+
+ real_frame_start = fp + map->start_offset;
+ frame_start = fp + map->start_offset + map->map_offset;
+ frame_end = fp + map->end_offset;
+
+ DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); fprintf (logfile, "Mark(0): %s+0x%x (%p) limit=%p fp=%p frame=%p-%p (%d)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx), stack_limit, fp, frame_start, frame_end, (int)(frame_end - frame_start)); g_free (fname));
+
+ /* Find the callsite index */
+ if (map->callsite_entry_size == 1) {
+ for (i = 0; i < map->ncallsites; ++i)
+ /* ip points inside the call instruction */
+ if (map->callsites.offsets8 [i] == pc_offset + 1)
+ break;
+ } else if (map->callsite_entry_size == 2) {
+ // FIXME: Use a binary search
+ for (i = 0; i < map->ncallsites; ++i)
+ /* ip points inside the call instruction */
+ if (map->callsites.offsets16 [i] == pc_offset + 1)
+ break;
+ } else {
+ // FIXME: Use a binary search
+ for (i = 0; i < map->ncallsites; ++i)
+ /* ip points inside the call instruction */
+ if (map->callsites.offsets32 [i] == pc_offset + 1)
+ break;
+ }
+ if (i == map->ncallsites) {
+ printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (ji->method, TRUE));
g_assert_not_reached ();
-#else
- fp = NULL;
- g_assert_not_reached ();
-#endif
+ }
+ cindex = i;
- frame_start = fp + map->frame_offset;
- frame_end = frame_start + (map->nslots * sizeof (mgreg_t));
+ g_assert (real_frame_start >= stack_limit);
- pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
- g_assert (pc_offset >= 0);
+ if (real_frame_start > stack_limit) {
+ /* This scans the previously skipped frames as well */
+ DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
+ mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
+ stats.scanned_other += real_frame_start - stack_limit;
+ }
- DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); printf ("Mark(%d): %s+0x%x (%p) limit=%p fp=%p frame=%p-%p (%d)\n", precise, fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx), stack_limit, fp, frame_start, frame_end, (int)(frame_end - frame_start)); g_free (fname));
+ /* Mark stack slots */
+ if (map->has_pin_slots) {
+ int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
+ guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
+ guint8 *p;
+ gboolean pinned;
+
+ p = frame_start;
+ for (i = 0; i < map->nslots; ++i) {
+ pinned = pin_bitmap [i / 8] & (1 << (i % 8));
+ if (pinned) {
+ DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
+ mono_gc_conservatively_scan_area (p, p + sizeof (mgreg_t));
+ scanned_conservatively += sizeof (mgreg_t);
+ } else {
+ scanned_precisely += sizeof (mgreg_t);
+ }
+ p += sizeof (mgreg_t);
+ }
+ } else {
+ scanned_precisely += (map->nslots * sizeof (mgreg_t));
+ }
- /*
- * FIXME: Add a function to mark using a bitmap, to avoid doing a
- * call for each object.
- */
+ /* The area outside of start-end is NOREF */
+ scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * sizeof (mgreg_t));
+
+ /* Mark registers */
+ precise_regmask = map->used_int_regs | (1 << map->frame_reg);
+ if (map->has_pin_regs) {
+ int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
+ guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
+ int bindex = 0;
+ for (i = 0; i < NREGS; ++i) {
+ if (!(map->used_int_regs & (1 << i)))
+ continue;
+
+ if (!(map->reg_pin_mask & (1 << i)))
+ continue;
- /* Pinning needs to be done first, then the precise scan later */
+ if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
+ DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
+ precise_regmask &= ~(1 << i);
+ }
+ bindex ++;
+ }
+ }
- if (!precise) {
- g_assert (frame_start >= stack_limit);
+ scanned += map->end_offset - map->start_offset;
- if (frame_start > stack_limit) {
- /* This scans the previously skipped frames as well */
- DEBUG (printf ("\tscan area %p-%p.\n", stack_limit, frame_start));
- mono_gc_conservatively_scan_area (stack_limit, frame_start);
- }
+ g_assert (scanned == scanned_precisely + scanned_conservatively);
- /* Mark stack slots */
- if (map->has_pin_slots) {
- guint8 *pin_bitmap = &map->pin_bitmap [(map->bitmap_width * pc_offset)];
- guint8 *p;
- gboolean pinned;
-
- p = frame_start;
- for (i = 0; i < map->nslots; ++i) {
- pinned = pin_bitmap [i / 8] & (1 << (i % 8));
- if (pinned) {
- DEBUG (printf ("\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
- mono_gc_conservatively_scan_area (p, p + sizeof (mgreg_t));
- scanned_conservatively += sizeof (mgreg_t);
- } else {
- scanned_precisely += sizeof (mgreg_t);
- }
- p += sizeof (mgreg_t);
+ stack_limit = frame_end;
+
+ /* Save information for the precise pass */
+ fi = &tls->frames [tls->nframes];
+ fi->nslots = map->nslots;
+ bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
+ if (map->has_ref_slots)
+ fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
+ else
+ fi->bitmap = NULL;
+ fi->frame_start_offset = frame_start - stack_start;
+ fi->nreg_locations = 0;
+ DEBUG_PRECISE (fi->ji = ji);
+ DEBUG_PRECISE (fi->fp = fp);
+
+ if (map->has_ref_regs) {
+ int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
+ guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
+ int bindex = 0;
+ for (i = 0; i < NREGS; ++i) {
+ if (!(map->reg_ref_mask & (1 << i)))
+ continue;
+
+ if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
+ DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
+ DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
+ fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
+ fi->nreg_locations ++;
}
- } else {
- scanned_precisely += map->nslots * sizeof (mgreg_t);
+ bindex ++;
}
+ }
- /* Mark registers */
- if (map->has_pin_regs) {
- guint8 *pin_bitmap = &map->reg_pin_bitmap [(map->reg_bitmap_width * pc_offset)];
- for (i = 0; i < map->nregs; ++i) {
- if (!(map->used_regs & (1 << i)))
- continue;
+ /*
+ * Clear locations of precisely stacked registers.
+ */
+ if (precise_regmask) {
+ for (i = 0; i < NREGS; ++i) {
+ if (precise_regmask & (1 << i)) {
+ /*
+ * The method uses this register, and we have precise info for it.
+ * This means the location will be scanned precisely.
+ * Tell the code at the beginning of the loop that this location is
+ * processed.
+ */
+ if (reg_locations [i])
+ DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
+ reg_locations [i] = NULL;
+ }
+ }
+ }
- /* We treated the save slots as precise above */
- scanned_precisely -= sizeof (mgreg_t);
+ tls->nframes ++;
+ }
- if (!reg_locations [i])
- continue;
+ /* Scan the remaining register save locations */
+ for (i = 0; i < MONO_MAX_IREGS; ++i) {
+ if (reg_locations [i]) {
+ DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
+ mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
+ scanned_registers += sizeof (mgreg_t);
+ }
+ if (new_reg_locations [i]) {
+ DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
+ mono_gc_conservatively_scan_area (new_reg_locations [i], new_reg_locations [i] + sizeof (mgreg_t));
+ scanned_registers += sizeof (mgreg_t);
+ }
+ }
- if (!(pin_bitmap [i / 8] & (1 << (i % 8)))) {
- /*
- * The method uses this register, and we have precise info for it.
- * This means the location will be scanned precisely.
- * Tell the code at the beginning of the loop that this location is
- * processed.
- */
- DEBUG (printf ("\treg %s at location %p is precise.\n", mono_arch_regname (i), reg_locations [i]));
- reg_locations [i] = NULL;
- scanned_precisely += sizeof (mgreg_t);
- }
- }
- }
+ if (stack_limit < stack_end) {
+ DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
+ mono_gc_conservatively_scan_area (stack_limit, stack_end);
+ stats.scanned_native += stack_end - stack_limit;
+ }
- scanned += frame_end - frame_start;
+ DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
- /* Not == because registers are marked in a later frame */
- g_assert (scanned >= scanned_precisely + scanned_conservatively);
+ stats.scanned_stacks += stack_end - stack_start;
+ stats.scanned += scanned;
+ stats.scanned_precisely += scanned_precisely;
+ stats.scanned_conservatively += scanned_conservatively;
+ stats.scanned_registers += scanned_registers;
- stack_limit = frame_end;
- } else {
- /* Mark stack slots */
- if (map->has_ref_slots) {
- guint8 *ref_bitmap = &map->ref_bitmap [(map->bitmap_width * pc_offset)];
- gboolean live;
-
- for (i = 0; i < map->nslots; ++i) {
- MonoObject **ptr = (MonoObject**)(frame_start + (i * sizeof (mgreg_t)));
-
- live = ref_bitmap [i / 8] & (1 << (i % 8));
-
- if (live) {
- MonoObject *obj = *ptr;
- if (obj) {
- DEBUG (printf ("\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fp)), ptr, obj));
- *ptr = mono_gc_scan_object (obj);
- DEBUG (printf (" %p.\n", *ptr));
- } else {
- DEBUG (printf ("\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fp)), ptr, obj));
- }
- } else {
- if (map->ref_slots [i / 8] & (1 << (i % 8))) {
- DEBUG (printf ("\tref %s0x%x(fp)=%p: dead (%p)\n", (guint8*)ptr >= (guint8*)fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fp)), ptr, *ptr));
- /*
- * Fail fast if the live range is incorrect, and
- * the JITted code tries to access this object
- */
- *ptr = DEAD_REF;
- }
- }
- }
- }
+ //mono_gc_conservatively_scan_area (stack_start, stack_end);
+}
- /* Mark registers */
+/*
+ * precise_pass:
+ *
+ * Mark a thread stack precisely based on information saved during the conservative
+ * pass.
+ */
+static void
+precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
+{
+ int findex, i;
+ FrameInfo *fi;
+ guint8 *frame_start;
- /*
- * Registers are different from stack slots, they have no address where they
- * are stored. Instead, some frame below this frame in the stack saves them
- * in its prolog to the stack. We can mark this location precisely.
- */
- if (map->has_ref_regs) {
- guint8 *ref_bitmap = &map->reg_ref_bitmap [(map->reg_bitmap_width * pc_offset)];
- for (i = 0; i < map->nregs; ++i) {
- if (!reg_locations [i])
- continue;
+ if (!tls)
+ return;
+
+ for (findex = 0; findex < tls->nframes; findex ++) {
+ /* Load information saved by the !precise pass */
+ fi = &tls->frames [findex];
+ frame_start = stack_start + fi->frame_start_offset;
+
+ DEBUG (char *fname = mono_method_full_name (fi->ji->method, TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
+
+ /*
+ * FIXME: Add a function to mark using a bitmap, to avoid doing a
+ * call for each object.
+ */
+
+ /* Mark stack slots */
+ if (fi->bitmap) {
+ guint8 *ref_bitmap = fi->bitmap;
+ gboolean live;
+
+ for (i = 0; i < fi->nslots; ++i) {
+ MonoObject **ptr = (MonoObject**)(frame_start + (i * sizeof (mgreg_t)));
- if (ref_bitmap [i / 8] & (1 << (i % 8))) {
+ live = ref_bitmap [i / 8] & (1 << (i % 8));
+
+ if (live) {
+ MonoObject *obj = *ptr;
+ if (obj) {
+ DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
+ *ptr = mono_gc_scan_object (obj);
+ DEBUG (fprintf (logfile, " %p.\n", *ptr));
+ } else {
+ DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
+ }
+ } else {
+#if 0
+ /*
+ * This is disabled because the pointer takes up a lot of space.
+ * Stack slots might be shared between ref and non-ref variables ?
+ */
+ if (map->ref_slots [i / 8] & (1 << (i % 8))) {
+ DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: dead (%p)\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, *ptr));
/*
- * reg_locations [i] contains the address of the stack slot where
- * i was last saved, so mark that slot.
+ * Fail fast if the live range is incorrect, and
+ * the JITted code tries to access this object
*/
- MonoObject **ptr = (MonoObject**)reg_locations [i];
- MonoObject *obj = *ptr;
-
- if (obj) {
- DEBUG (printf ("\treg %s saved at %p: %p ->", mono_arch_regname (i), reg_locations [i], obj));
- *ptr = mono_gc_scan_object (obj);
- DEBUG (printf (" %p.\n", *ptr));
- } else {
- DEBUG (printf ("\treg %s saved at %p: %p", mono_arch_regname (i), reg_locations [i], obj));
- }
+ *ptr = DEAD_REF;
}
-
- /* Mark the save slot as processed */
- reg_locations [i] = NULL;
+#endif
}
- }
+ }
}
- }
- if (!precise) {
- /* Scan the remaining register save locations */
- for (i = 0; i < MONO_MAX_IREGS; ++i) {
- if (reg_locations [i]) {
- DEBUG (printf ("\tscan saved reg location %p.\n", reg_locations [i]));
- mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
- scanned_conservatively += sizeof (mgreg_t);
- }
- // FIXME: Is this needed ?
- if (new_reg_locations [i]) {
- DEBUG (printf ("\tscan saved reg location %p.\n", new_reg_locations [i]));
- mono_gc_conservatively_scan_area (new_reg_locations [i], new_reg_locations [i] + sizeof (mgreg_t));
- scanned_conservatively += sizeof (mgreg_t);
+ /* Mark registers */
+
+ /*
+ * Registers are different from stack slots, they have no address where they
+ * are stored. Instead, some frame below this frame in the stack saves them
+ * in its prolog to the stack. We can mark this location precisely.
+ */
+ for (i = 0; i < fi->nreg_locations; ++i) {
+ /*
+ * reg_locations [i] contains the address of the stack slot where
+ * a reg was last saved, so mark that slot.
+ */
+ MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
+ MonoObject *obj = *ptr;
+
+ if (obj) {
+ DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
+ *ptr = mono_gc_scan_object (obj);
+ DEBUG (fprintf (logfile, " %p.\n", *ptr));
+ } else {
+ DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
}
- }
+ }
}
- if (stack_limit < stack_end && !precise) {
- DEBUG (printf ("\tscan area %p-%p.\n", stack_limit, stack_end));
- mono_gc_conservatively_scan_area (stack_limit, stack_end);
+ /*
+ * Debugging aid to check for missed refs.
+ */
+ if (tls->ref_to_track) {
+ mgreg_t *p;
+
+ for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
+ if (*p == (mgreg_t)tls->ref_to_track)
+ printf ("REF AT %p.\n", p);
}
+}
- DEBUG (printf ("Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
+/*
+ * thread_mark_func:
+ *
+ * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
+ * call, and TRUE at the second. USER_DATA points to a TlsData
+ * structure filled up by thread_suspend_func.
+ */
+static void
+thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
+{
+ TlsData *tls = user_data;
- if (!precise) {
- stats.scanned_stacks += stack_end - stack_start;
- stats.scanned += scanned;
- stats.scanned_precisely += scanned_precisely;
- stats.scanned_conservatively += scanned_conservatively;
- stats.scanned_registers += scanned_registers;
- }
+ DEBUG (fprintf (logfile, "****************************************\n"));
+ DEBUG (fprintf (logfile, "*** %s stack marking for thread %p (%p-%p) ***\n", precise ? "Precise" : "Conservative", tls ? GUINT_TO_POINTER (tls->tid) : NULL, stack_start, stack_end));
+ DEBUG (fprintf (logfile, "****************************************\n"));
- //mono_gc_conservatively_scan_area (stack_start, stack_end);
+ if (!precise)
+ conservative_pass (tls, stack_start, stack_end);
+ else
+ precise_pass (tls, stack_start, stack_end);
}
static void
if (COMPILE_LLVM (cfg))
return;
+ if (!mono_gc_is_moving ())
+ return;
+
+ if (!cfg->compile_aot && !mono_gc_precise_stack_mark_enabled ())
+ return;
+
#if 1
/* Debugging support */
{
* need to call this function for those slots.
*/
void
-mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, StackSlotType type)
+mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
{
MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
* call this function for those slots.
*/
void
-mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, StackSlotType type)
+mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
{
MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
int slot = - (slot_offset / sizeof (mgreg_t));
gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
}
-static inline void
-set_slot (MonoCompileGC *gcfg, int pos, StackSlotType val)
+static inline int
+fp_offset_to_slot (MonoCompile *cfg, int offset)
{
- g_assert (pos >= 0 && pos < gcfg->nslots);
- gcfg->slots [pos] = val;
+ MonoCompileGC *gcfg = cfg->gc_info;
+
+ return (offset - gcfg->min_offset) / sizeof (mgreg_t);
}
static inline int
-fp_offset_to_slot (MonoCompile *cfg, int offset)
+slot_to_fp_offset (MonoCompile *cfg, int slot)
{
MonoCompileGC *gcfg = cfg->gc_info;
- return (offset - gcfg->min_offset) / sizeof (mgreg_t);
+ return (slot * sizeof (mgreg_t)) + gcfg->min_offset;
+}
+
+static inline void
+set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
+{
+ g_assert (slot >= 0 && slot < gcfg->nslots);
+
+ if (type == SLOT_PIN) {
+ clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
+ set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
+ } else if (type == SLOT_REF) {
+ set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
+ clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
+ } else if (type == SLOT_NOREF) {
+ clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
+ clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
+ }
+}
+
+static inline void
+set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
+{
+ int cindex;
+
+ for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
+ set_slot (gcfg, slot, cindex, type);
+}
+
+static inline void
+set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
+{
+ int cindex;
+
+ for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
+ int callsite_offset = gcfg->callsites [cindex]->pc_offset;
+ if (callsite_offset >= from && callsite_offset < to)
+ set_slot (gcfg, slot, cindex, type);
+ }
+}
+
+static inline void
+set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
+{
+ g_assert (slot >= 0 && slot < gcfg->nregs);
+
+ if (type == SLOT_PIN) {
+ clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
+ set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
+ } else if (type == SLOT_REF) {
+ set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
+ clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
+ } else if (type == SLOT_NOREF) {
+ clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
+ clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
+ }
+}
+
+static inline void
+set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
+{
+ int cindex;
+
+ for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
+ set_reg_slot (gcfg, slot, cindex, type);
+}
+
+static inline void
+set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
+{
+ int cindex;
+
+ for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
+ int callsite_offset = gcfg->callsites [cindex]->pc_offset;
+ if (callsite_offset >= from && callsite_offset < to)
+ set_reg_slot (gcfg, slot, cindex, type);
+ }
}
static void
GSList *l;
int i;
+ /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
- /*
- * Extend the live interval for the GC tracked spill slots
- * defined in this bblock.
- */
for (l = bb->spill_slot_defs; l; l = l->next) {
MonoInst *def = l->data;
int spill_slot = def->inst_c0;
- int offset = cfg->spill_info [MONO_REG_INT_REF][spill_slot].offset;
+ int bank = def->inst_c1;
+ int offset = cfg->spill_info [bank][spill_slot].offset;
int slot = fp_offset_to_slot (cfg, offset);
- MonoLiveInterval *interval;
- set_slot (gcfg, slot, SLOT_REF);
+ if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
+ set_slot_everywhere (gcfg, slot, SLOT_NOREF);
+ }
+ }
- interval = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoLiveInterval));
- mono_linterval_add_range (cfg, interval, def->backend.pc_offset, bb->native_offset + bb->native_length);
- gcfg->live_intervals [slot] = g_slist_prepend_mempool (cfg->mempool, gcfg->live_intervals [slot], interval);
+ for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
+ for (l = bb->spill_slot_defs; l; l = l->next) {
+ MonoInst *def = l->data;
+ int spill_slot = def->inst_c0;
+ int bank = def->inst_c1;
+ int offset = cfg->spill_info [bank][spill_slot].offset;
+ int slot = fp_offset_to_slot (cfg, offset);
+ GCSlotType type;
+
+ if (bank == MONO_REG_INT_MP)
+ type = SLOT_PIN;
+ else
+ type = SLOT_REF;
+
+ /*
+ * Extend the live interval for the GC tracked spill slots
+ * defined in this bblock.
+ * FIXME: This is not needed.
+ */
+ set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
if (cfg->verbose_level > 1)
- printf ("\tref spill slot at fp+0x%x (slot = %d)\n", offset, slot);
+ printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
}
}
slot = fp_offset_to_slot (cfg, offset);
- set_slot (gcfg, slot, SLOT_NOREF);
+ set_slot_everywhere (gcfg, slot, SLOT_NOREF);
/* FIXME: 32 bit */
if (cfg->verbose_level > 1)
- printf ("\tfp spill slot at fp+0x%x (slot = %d)\n", offset, slot);
+ printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
}
/* Set int spill slots to NOREF */
slot = fp_offset_to_slot (cfg, offset);
- set_slot (gcfg, slot, SLOT_NOREF);
+ set_slot_everywhere (gcfg, slot, SLOT_NOREF);
if (cfg->verbose_level > 1)
- printf ("\tint spill slot at fp+0x%x (slot = %d)\n", offset, slot);
+ printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
}
}
/* Relative to the CFA */
for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
guint data = GPOINTER_TO_UINT (l->data);
- int slot = data >> 16;
- StackSlotType type = data & 0xff;
- int fp_slot;
+ int cfa_slot = data >> 16;
+ GCSlotType type = data & 0xff;
+ int slot;
/*
* Map the cfa relative slot to an fp relative slot.
- * fp_slot_addr == cfa - <slot>*4/8
+ * slot_addr == cfa - <cfa_slot>*4/8
* fp + cfa_offset == cfa
- * -> fp_slot_addr == fp + (cfa_offset - <slot>*4/8)
+ * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
*/
- fp_slot = (cfg->cfa_offset / sizeof (mgreg_t)) - slot - (gcfg->min_offset / sizeof (mgreg_t));
+ slot = (cfg->cfa_offset / sizeof (mgreg_t)) - cfa_slot - (gcfg->min_offset / sizeof (mgreg_t));
- set_slot (gcfg, fp_slot, type);
+ set_slot_everywhere (gcfg, slot, type);
if (cfg->verbose_level > 1) {
+ int fp_offset = slot_to_fp_offset (cfg, slot);
if (type == SLOT_NOREF)
- printf ("\tnoref slot at fp+0x%x (slot = %d) (cfa - 0x%x)\n", (int)(fp_slot * sizeof (mgreg_t)), fp_slot, (int)(slot * sizeof (mgreg_t)));
+ printf ("\tnoref slot at %s0x%x(fp) (slot = %d) (cfa - 0x%x)\n", fp_offset >= 0 ? "" : "-", ABS (fp_offset), slot, (int)(cfa_slot * sizeof (mgreg_t)));
}
}
for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
gint data = GPOINTER_TO_INT (l->data);
int offset = data >> 16;
- StackSlotType type = data & 0xff;
+ GCSlotType type = data & 0xff;
int slot;
slot = fp_offset_to_slot (cfg, offset);
- set_slot (gcfg, slot, type);
+ set_slot_everywhere (gcfg, slot, type);
/* Liveness for these slots is handled by process_spill_slots () */
}
static void
-process_locals (MonoCompile *cfg)
+process_variables (MonoCompile *cfg)
{
- int i, locals_min_slot, locals_max_slot;
+ MonoCompileGC *gcfg = cfg->gc_info;
+ MonoMethodSignature *sig = mono_method_signature (cfg->method);
+ int i, locals_min_slot, locals_max_slot, cindex;
MonoBasicBlock *bb;
MonoInst *tmp;
int *pc_offsets;
- MonoCompileGC *gcfg = cfg->gc_info;
- GSList **live_intervals = gcfg->live_intervals;
- gboolean *starts_pinned = gcfg->starts_pinned;
int locals_min_offset = gcfg->locals_min_offset;
int locals_max_offset = gcfg->locals_max_offset;
/* Slots for locals are NOREF by default */
locals_min_slot = (locals_min_offset - gcfg->min_offset) / sizeof (mgreg_t);
locals_max_slot = (locals_max_offset - gcfg->min_offset) / sizeof (mgreg_t);
- for (i = locals_min_slot; i < locals_max_slot; ++i)
- set_slot (gcfg, i, SLOT_NOREF);
+ for (i = locals_min_slot; i < locals_max_slot; ++i) {
+ set_slot_everywhere (gcfg, i, SLOT_NOREF);
+ }
/*
* Compute the offset where variables are initialized in the first bblock, if any.
g_assert (tmp->backend.pc_offset > 0);
pc_offsets [vreg] = tmp->backend.pc_offset;
}
- break;
}
}
- for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
+ /*
+ * Stack slots holding arguments are initialized in the prolog.
+ * This means we can treat them alive for the whole method.
+ */
+ for (i = 0; i < cfg->num_varinfo; i++) {
MonoInst *ins = cfg->varinfo [i];
MonoType *t = ins->inst_vtype;
MonoMethodVar *vmv;
guint32 pos;
- gboolean byref = t->byref;
- MonoLiveInterval *li;
+ gboolean byref, is_this = FALSE;
+ gboolean is_arg = i < cfg->locals_start;
+
+ if (ins == cfg->ret)
+ continue;
vmv = MONO_VARINFO (cfg, i);
+ /* For some reason, 'this' is byref */
+ if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
+ t = &cfg->method->klass->byval_arg;
+ is_this = TRUE;
+ }
+
+ byref = t->byref;
+
if (ins->opcode == OP_REGVAR) {
int hreg;
- StackSlotType slot_type;
+ GCSlotType slot_type;
t = mini_type_get_underlying_type (NULL, t);
hreg = ins->dreg;
g_assert (hreg < MONO_MAX_IREGS);
- if (byref) {
- /* These have no live interval, be conservative */
+ if (byref)
slot_type = SLOT_PIN;
- li = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoLiveInterval));
- mono_linterval_add_range (cfg, li, 0, cfg->code_size);
- } else if (!MONO_TYPE_IS_REFERENCE (t)) {
- // FIXME: These have no gc live interval
- continue;
+ else
+ slot_type = MONO_TYPE_IS_REFERENCE (t) ? SLOT_REF : SLOT_NOREF;
+
+ if (slot_type == SLOT_PIN) {
+ /* These have no live interval, be conservative */
+ set_reg_slot_everywhere (gcfg, hreg, slot_type);
} else {
- slot_type = SLOT_REF;
- li = vmv->gc_interval;
+ /*
+ * Unlike variables allocated to the stack, we generate liveness info
+ * for noref vars in registers in mono_spill_global_vars (), because
+ * knowing that a register doesn't contain a ref allows us to mark its save
+ * locations precisely.
+ */
+ for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
+ if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
+ set_reg_slot (gcfg, hreg, cindex, slot_type);
}
- gcfg->reg_types [hreg] = g_slist_prepend_mempool (cfg->mempool, gcfg->reg_types [hreg], GINT_TO_POINTER (slot_type));
- gcfg->reg_live_intervals [hreg] = g_slist_prepend_mempool (cfg->mempool, gcfg->reg_live_intervals [hreg], li);
-
if (cfg->verbose_level > 1) {
- printf ("\t%s reg %s: ", slot_type_to_string (slot_type), mono_arch_regname (hreg));
- mono_linterval_print (li);
- printf ("\n");
+ printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
}
continue;
if (ins->inst_offset % sizeof (mgreg_t) != 0)
continue;
+ if (is_arg && ins->inst_offset >= gcfg->max_offset)
+ /* In parent frame */
+ continue;
+
pos = fp_offset_to_slot (cfg, ins->inst_offset);
+ if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
+ /* These do not get stored in the prolog */
+ set_slot_everywhere (gcfg, pos, SLOT_NOREF);
+
+ if (cfg->verbose_level > 1) {
+ printf ("\tdead arg at fp%s0x%x (slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "+", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, pos, mono_type_full_name (ins->inst_vtype));
+ }
+ continue;
+ }
+
if (MONO_TYPE_ISSTRUCT (t)) {
int numbits = 0, j;
gsize *bitmap = NULL;
gboolean pin = FALSE;
- MonoLiveInterval *interval = NULL;
int size;
+ int size_in_slots;
+
+ if (ins->backend.is_pinvoke)
+ size = mono_class_native_size (ins->klass, NULL);
+ else
+ size = mono_class_value_size (ins->klass, NULL);
+ size_in_slots = ALIGN_TO (size, sizeof (mgreg_t)) / sizeof (mgreg_t);
- if (!ins->klass->has_references)
+ if (!ins->klass->has_references) {
+ if (is_arg) {
+ for (j = 0; j < size_in_slots; ++j)
+ set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
+ }
continue;
+ }
if (ins->klass->generic_container || mono_class_is_open_constructed_type (t)) {
/* FIXME: Generic sharing */
mono_class_compute_gc_descriptor (ins->klass);
bitmap = mono_gc_get_bitmap_for_descr (ins->klass->gc_descr, &numbits);
-
if (!bitmap)
- // FIXME:
pin = TRUE;
/*
* before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
* them during VZERO decomposition.
*/
- if (pc_offsets [vmv->vreg]) {
- interval = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoLiveInterval));
- mono_linterval_add_range (cfg, interval, pc_offsets [vmv->vreg], cfg->code_size);
- } else {
+ if (!pc_offsets [vmv->vreg])
pin = TRUE;
- }
}
if (ins->backend.is_pinvoke)
pin = TRUE;
- if (ins->backend.is_pinvoke)
- size = mono_class_native_size (ins->klass, NULL);
- else
- size = mono_class_value_size (ins->klass, NULL);
+ if (cfg->verbose_level > 1)
+ printf ("\tvtype R%d at fp+0x%x-0x%x: %s\n", vmv->vreg, (int)ins->inst_offset, (int)(ins->inst_offset + (size / sizeof (mgreg_t))), mono_type_full_name (ins->inst_vtype));
if (bitmap) {
- for (j = 0; j < numbits; ++j) {
- if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
- /* The descriptor is for the boxed object */
- set_slot (gcfg, (pos + j - (sizeof (MonoObject) / sizeof (gpointer))), pin ? SLOT_PIN : SLOT_REF);
+ for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
+ if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
+ for (j = 0; j < numbits; ++j) {
+ if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
+ /* The descriptor is for the boxed object */
+ set_slot (gcfg, (pos + j - (sizeof (MonoObject) / sizeof (mgreg_t))), cindex, pin ? SLOT_PIN : SLOT_REF);
+ }
+ }
}
}
- } else if (pin) {
- for (j = 0; j < size / sizeof (mgreg_t); ++j)
- set_slot (gcfg, pos + j, SLOT_PIN);
- }
- if (!pin) {
- for (j = 0; j < size / sizeof (mgreg_t); ++j) {
- live_intervals [pos + j] = g_slist_prepend_mempool (cfg->mempool, live_intervals [pos + j], interval);
- starts_pinned [pos + j] = TRUE;
+ if (cfg->verbose_level > 1) {
+ for (j = 0; j < numbits; ++j) {
+ if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
+ printf ("\t\t%s slot at 0x%x(fp) (slot = %d)\n", pin ? "pin" : "ref", (int)(ins->inst_offset + (j * sizeof (mgreg_t))), (int)(pos + j - (sizeof (MonoObject) / sizeof (mgreg_t))));
+ }
+ }
+ } else {
+ if (cfg->verbose_level > 1)
+ printf ("\t\tpinned\n");
+ for (j = 0; j < size_in_slots; ++j) {
+ set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
}
}
g_free (bitmap);
- if (cfg->verbose_level > 1) {
- printf ("\tvtype R%d at fp+0x%x-0x%x: %s ", vmv->vreg, (int)ins->inst_offset, (int)(ins->inst_offset + (size / sizeof (mgreg_t))), mono_type_full_name (ins->inst_vtype));
- if (interval)
- mono_linterval_print (interval);
- else
- printf ("(pinned)");
- printf ("\n");
- }
-
continue;
}
- if (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset)
+ if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
/* Vret addr etc. */
continue;
if (t->byref) {
- set_slot (gcfg, pos, SLOT_PIN);
+ if (is_arg) {
+ set_slot_everywhere (gcfg, pos, SLOT_PIN);
+ } else {
+ for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
+ if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
+ set_slot (gcfg, pos, cindex, SLOT_PIN);
+ }
+ if (cfg->verbose_level > 1)
+ printf ("\tbyref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
continue;
}
* could hold GC refs since the vregs they were created from might not been
* marked as holding a GC ref. So be conservative.
*/
- set_slot (gcfg, pos, SLOT_PIN);
+ set_slot_everywhere (gcfg, pos, SLOT_PIN);
continue;
}
#endif
t = mini_type_get_underlying_type (NULL, t);
- if (MONO_TYPE_IS_REFERENCE (ins->inst_vtype)) {
- if (vmv && !vmv->gc_interval) {
- set_slot (gcfg, pos, SLOT_PIN);
- continue;
- }
+ if (!MONO_TYPE_IS_REFERENCE (t)) {
+ set_slot_everywhere (gcfg, pos, SLOT_NOREF);
+ if (cfg->verbose_level > 1)
+ printf ("\tnoref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
+ continue;
+ }
- if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT)) {
- /*
- * For volatile variables, treat them alive from the point they are
- * initialized in the first bblock until the end of the method.
- */
- if (pc_offsets [vmv->vreg]) {
- vmv->gc_interval = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoLiveInterval));
- mono_linterval_add_range (cfg, vmv->gc_interval, pc_offsets [vmv->vreg], cfg->code_size);
- starts_pinned [pos] = TRUE;
- } else {
- set_slot (gcfg, pos, SLOT_PIN);
- continue;
- }
+ /* 'this' is marked INDIRECT for gshared methods */
+ if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
+ /*
+ * For volatile variables, treat them alive from the point they are
+ * initialized in the first bblock until the end of the method.
+ */
+ if (is_arg) {
+ set_slot_everywhere (gcfg, pos, SLOT_REF);
+ } else if (pc_offsets [vmv->vreg]) {
+ set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
+ set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
+ } else {
+ set_slot_everywhere (gcfg, pos, SLOT_PIN);
}
+ if (cfg->verbose_level > 1)
+ printf ("\tvolatile ref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
+ continue;
+ }
- set_slot (gcfg, pos, SLOT_REF);
-
- live_intervals [pos] = g_slist_prepend_mempool (cfg->mempool, live_intervals [pos], vmv->gc_interval);
+ if (is_arg) {
+ /* Live for the whole method */
+ set_slot_everywhere (gcfg, pos, SLOT_REF);
+ } else {
+ for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
+ if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
+ set_slot (gcfg, pos, cindex, SLOT_REF);
+ }
- if (cfg->verbose_level > 1) {
- printf ("\tref at %s0x%x(fp) (slot=%d): %s ", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, pos, mono_type_full_name (ins->inst_vtype));
- mono_linterval_print (vmv->gc_interval);
- printf ("\n");
- }
+ if (cfg->verbose_level > 1) {
+ printf ("\tref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
}
}
g_free (pc_offsets);
}
-// FIXME: Merge this with process_locals ()
-static void
-process_arguments (MonoCompile *cfg)
+static int
+sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
{
- int i, j;
- MonoCompileGC *gcfg = cfg->gc_info;
- MonoMethodSignature *sig = mono_method_signature (cfg->method);
- MonoLiveInterval *li;
-
- /*
- * Stack slots holding arguments are initialized in the prolog.
- * This means we can treat them alive for the whole method.
+ /*
+ * Convert a sp relative offset to a slot index. This is
+ * platform specific.
*/
- for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
- MonoInst *ins = cfg->args [i];
- MonoType *t = ins->inst_vtype;
- int pos;
- gboolean byref = t->byref;
-
- /* For some reason, 'this' is byref */
- if (sig->hasthis && i == 0 && !cfg->method->klass->valuetype)
- t = &cfg->method->klass->byval_arg;
-
- if (ins->opcode == OP_REGVAR) {
- int hreg;
- StackSlotType slot_type;
-
- t = mini_type_get_underlying_type (NULL, t);
-
- if (byref)
- slot_type = SLOT_PIN;
- else
- slot_type = MONO_TYPE_IS_REFERENCE (t) ? SLOT_REF : SLOT_NOREF;
-
- hreg = ins->dreg;
- g_assert (hreg < MONO_MAX_IREGS);
-
- if (gcfg->reg_live_intervals [hreg]) {
- /*
- * FIXME: This argument shares a hreg with a local, we can't add the whole
- * method as a live interval, since it would overlap with the locals
- * live interval.
- */
- continue;
- }
-
- gcfg->reg_types [hreg] = g_slist_prepend_mempool (cfg->mempool, gcfg->reg_types [hreg], GINT_TO_POINTER (slot_type));
-
- /* Live for the whole method */
- li = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoLiveInterval));
- mono_linterval_add_range (cfg, li, 0, cfg->code_size);
- gcfg->reg_live_intervals [hreg] = g_slist_prepend_mempool (cfg->mempool, gcfg->reg_live_intervals [hreg], li);
+#ifdef TARGET_AMD64
+ /* fp = sp + offset */
+ g_assert (cfg->frame_reg == AMD64_RBP);
+ return (- cfg->arch.sp_fp_offset + sp_offset);
+#else
+ NOT_IMPLEMENTED;
+ return -1;
+#endif
+}
- if (cfg->verbose_level > 1) {
- printf ("\t%s arg reg %s: ", slot_type_to_string (slot_type), mono_arch_regname (hreg));
- mono_linterval_print (li);
- printf ("\n");
+static GCSlotType
+type_to_gc_slot_type (MonoType *t)
+{
+ if (t->byref)
+ return SLOT_PIN;
+ t = mini_type_get_underlying_type (NULL, t);
+ if (MONO_TYPE_IS_REFERENCE (t))
+ return SLOT_REF;
+ else {
+ if (MONO_TYPE_ISSTRUCT (t)) {
+ MonoClass *klass = mono_class_from_mono_type (t);
+ if (!klass->has_references) {
+ return SLOT_NOREF;
+ } else {
+ // FIXME:
+ return SLOT_PIN;
}
-
- continue;
}
+ return SLOT_NOREF;
+ }
+}
- if (ins->opcode != OP_REGOFFSET)
- continue;
+static void
+process_param_area_slots (MonoCompile *cfg)
+{
+ MonoCompileGC *gcfg = cfg->gc_info;
+ int i;
+ gboolean *is_param;
- g_assert (ins->inst_offset % sizeof (mgreg_t) == 0);
+ /*
+ * These slots are used for passing parameters during calls. They are sp relative, not
+ * fp relative, so they are harder to handle.
+ */
+ if (cfg->flags & MONO_CFG_HAS_ALLOCA)
+ /* The distance between fp and sp is not constant */
+ return;
- pos = (ins->inst_offset - gcfg->min_offset) / sizeof (mgreg_t);
+ is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
- if (ins->inst_offset >= gcfg->max_offset)
- /* In parent frame */
- continue;
+ for (i = 0; i < gcfg->ncallsites; ++i) {
+ GCCallSite *callsite = gcfg->callsites [i];
+ GSList *l;
- if (ins->flags & MONO_INST_IS_DEAD) {
- /* These do not get stored in the prolog */
- set_slot (gcfg, pos, SLOT_NOREF);
+ for (l = callsite->param_slots; l; l = l->next) {
+ MonoInst *def = l->data;
+ int sp_offset = def->inst_offset;
+ int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
+ int slot = fp_offset_to_slot (cfg, fp_offset);
- if (cfg->verbose_level > 1) {
- printf ("\tdead arg at fp%s0x%x (slot=%d): %s\n", ins->inst_offset < 0 ? "-" : "+", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, pos, mono_type_full_name (ins->inst_vtype));
- }
- continue;
+ g_assert (slot >= 0 && slot < gcfg->nslots);
+ is_param [slot] = TRUE;
}
+ }
- if (MONO_TYPE_ISSTRUCT (t)) {
- int size;
-
- if (ins->backend.is_pinvoke)
- size = mono_class_native_size (ins->klass, NULL);
- else
- size = mono_class_value_size (ins->klass, NULL);
+ /* All param area slots are noref by default */
+ for (i = 0; i < gcfg->nslots; ++i) {
+ if (is_param [i])
+ set_slot_everywhere (gcfg, i, SLOT_NOREF);
+ }
- if (!ins->klass->has_references) {
- for (j = 0; j < size / sizeof (mgreg_t); ++j)
- set_slot (gcfg, pos + j, SLOT_NOREF);
- continue;
- }
+ for (i = 0; i < gcfg->ncallsites; ++i) {
+ GCCallSite *callsite = gcfg->callsites [i];
+ GSList *l;
- // FIXME:
- continue;
+ for (l = callsite->param_slots; l; l = l->next) {
+ MonoInst *def = l->data;
+ MonoType *t = def->inst_vtype;
+ int sp_offset = def->inst_offset;
+ int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
+ int slot = fp_offset_to_slot (cfg, fp_offset);
+ GCSlotType type = type_to_gc_slot_type (t);
+
+ /* The slot is live between the def instruction and the call */
+ set_slot_in_range (gcfg, slot, def->backend.pc_offset, callsite->pc_offset + 1, type);
+ if (cfg->verbose_level > 1)
+ printf ("\t%s param area slot at %s0x%x(fp)=0x%x(sp) (slot = %d) [0x%x-0x%x]\n", slot_type_to_string (type), fp_offset >= 0 ? "+" : "-", ABS (fp_offset), sp_offset, slot, def->backend.pc_offset, callsite->pc_offset + 1);
}
+ }
+}
- if (t->byref)
- continue;
-
- t = mini_type_get_underlying_type (NULL, t);
-
- if (!MONO_TYPE_IS_REFERENCE (t)) {
- set_slot (gcfg, pos, SLOT_NOREF);
- continue;
- }
+static void
+process_finally_clauses (MonoCompile *cfg)
+{
+ MonoCompileGC *gcfg = cfg->gc_info;
+ GCCallSite **callsites;
+ int ncallsites;
+ gboolean has_finally;
+ int i, j, nslots, nregs;
- set_slot (gcfg, pos, SLOT_REF);
+ ncallsites = gcfg->ncallsites;
+ nslots = gcfg->nslots;
+ nregs = gcfg->nregs;
+ callsites = gcfg->callsites;
- /* Live for the whole method */
- li = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoLiveInterval));
- mono_linterval_add_range (cfg, li, 0, cfg->code_size);
+ /*
+ * The calls to the finally clauses don't show up in the cfg. See
+ * test_0_liveness_8 ().
+ * Variables accessed inside the finally clause are already marked VOLATILE by
+ * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
+ * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
+ */
+ has_finally = FALSE;
+ for (i = 0; i < cfg->header->num_clauses; ++i) {
+ MonoExceptionClause *clause = &cfg->header->clauses [i];
- gcfg->live_intervals [pos] = g_slist_prepend_mempool (cfg->mempool, gcfg->live_intervals [pos], li);
+ if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
+ has_finally = TRUE;
+ }
+ }
+ if (has_finally) {
+ if (cfg->verbose_level > 1)
+ printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
+ for (j = 0; j < ncallsites; ++j) {
+ MonoBasicBlock *bb = callsites [j]->bb;
+ MonoExceptionClause *clause;
+ gboolean is_in_finally = FALSE;
+
+ for (i = 0; i < cfg->header->num_clauses; ++i) {
+ clause = &cfg->header->clauses [i];
+
+ if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
+ if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
+ is_in_finally = TRUE;
+ break;
+ }
+ }
+ }
- if (cfg->verbose_level > 1) {
- printf ("\tref at %s0x%x(fp) (slot=%d): %s ", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, pos, mono_type_full_name (ins->inst_vtype));
- mono_linterval_print (li);
- printf ("\n");
+ if (is_in_finally) {
+ for (i = 0; i < nslots; ++i)
+ set_slot (gcfg, i, j, SLOT_PIN);
+ for (i = 0; i < nregs; ++i)
+ set_reg_slot (gcfg, i, j, SLOT_PIN);
+ }
}
}
}
int min_offset, max_offset;
MonoCompileGC *gcfg = cfg->gc_info;
MonoMethodSignature *sig = mono_method_signature (cfg->method);
+ GSList *l;
/* Compute min/max offsets from the fp */
/* Locals */
-#ifdef TARGET_AMD64
+#if defined(TARGET_AMD64) || defined(TARGET_X86)
locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, sizeof (mgreg_t));
locals_max_offset = cfg->locals_max_stack_offset;
#else
min_offset = MIN (min_offset, cfa_min_offset);
max_offset = MAX (max_offset, cfa_max_offset);
+ /* Fp relative slots */
+ for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
+ gint data = GPOINTER_TO_INT (l->data);
+ int offset = data >> 16;
+
+ min_offset = MIN (min_offset, offset);
+ }
+
/* Spill slots */
if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
int stack_offset = ALIGN_TO (cfg->stack_offset, sizeof (mgreg_t));
min_offset = MIN (min_offset, (-stack_offset));
}
+ /* Param area slots */
+#ifdef TARGET_AMD64
+ min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
+#endif
+
gcfg->min_offset = min_offset;
gcfg->max_offset = max_offset;
gcfg->locals_min_offset = locals_min_offset;
gcfg->locals_max_offset = locals_max_offset;
}
-void
-mini_gc_create_gc_map (MonoCompile *cfg)
+static void
+init_gcfg (MonoCompile *cfg)
{
- GCMap *map;
- int i, nregs, nslots, alloc_size;
- int ntypes [16];
- StackSlotType *slots = NULL;
- GSList **live_intervals;
- int bitmap_width, bitmap_size, reg_bitmap_width, reg_bitmap_size;
- gboolean *starts_pinned;
- gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
+ int i, nregs, nslots;
MonoCompileGC *gcfg = cfg->gc_info;
-
- if (!cfg->compute_gc_maps)
- return;
+ GCCallSite **callsites;
+ int ncallsites;
+ MonoBasicBlock *bb;
+ GSList *l;
/*
- * Since we currently don't use GC safe points, we need to create GC maps which
- * are precise at every instruction within a method. The live ranges calculated by
- * the liveness pass are not usable for this, since they contain abstract positions, not
- * pc offsets. The live ranges calculated by mono_spill_global_vars () are not usable
- * either, since they can't model holes. Instead of these, we implement our own
- * liveness analysis which is precise, and works with PC offsets. It calculates live
- * intervals, which are unions of live ranges.
- * FIXME:
- * - it would simplify things if we extended live ranges to the end of basic blocks
- * instead of computing them precisely.
- * - maybe mark loads+stores as needing GC tracking, instead of using DEF/USE
- * instructions ?
- * - group ref and no-ref slots together on the stack, to speed up marking, and
- * to make the gc bitmaps better compressable.
- * During marking, all frames except the top frame are at a call site, and we mark the
- * top frame conservatively. This means that stack slots initialized in the prolog can
- * be assumed to be valid/live through the whole method.
+ * Collect callsites
*/
+ ncallsites = 0;
+ for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
+ ncallsites += g_slist_length (bb->gc_callsites);
+ }
+ callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
+ i = 0;
+ for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
+ for (l = bb->gc_callsites; l; l = l->next)
+ callsites [i++] = l->data;
+ }
- if (!(cfg->comp_done & MONO_COMP_LIVENESS))
- /* Without liveness info, the live ranges are not precise enough */
- return;
-
- if (cfg->header->num_clauses)
- /*
- * The calls to the finally clauses don't show up in the cfg. See
- * test_0_liveness_8 ().
- */
- return;
-
- mono_analyze_liveness_gc (cfg);
-
- compute_frame_size (cfg);
+ /* The callsites should already be ordered by pc offset */
+ for (i = 1; i < ncallsites; ++i)
+ g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
/*
* The stack frame looks like this:
printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
nslots = (gcfg->max_offset - gcfg->min_offset) / sizeof (mgreg_t);
- nregs = MONO_MAX_IREGS;
- /* slot [i] == type of slot at fp - <min_offset> + i*4/8 */
- slots = g_new0 (StackSlotType, nslots);
- live_intervals = g_new0 (GSList*, nslots);
- starts_pinned = g_new0 (gboolean, nslots);
+ nregs = NREGS;
- gcfg->slots = slots;
gcfg->nslots = nslots;
- gcfg->live_intervals = live_intervals;
- gcfg->starts_pinned = starts_pinned;
- gcfg->reg_types = mono_mempool_alloc0 (cfg->mempool, sizeof (GSList*) * nregs);
- gcfg->reg_live_intervals = mono_mempool_alloc0 (cfg->mempool, sizeof (GSList*) * nregs);
+ gcfg->nregs = nregs;
+ gcfg->callsites = callsites;
+ gcfg->ncallsites = ncallsites;
+ gcfg->stack_bitmap_width = ALIGN_TO (nslots, 8) / 8;
+ gcfg->reg_bitmap_width = ALIGN_TO (nregs, 8) / 8;
+ gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
+ gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
+ gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
+ gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
/* All slots start out as PIN */
- for (i = 0; i < nslots; ++i)
- set_slot (gcfg, i, SLOT_PIN);
+ memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * ncallsites);
+ for (i = 0; i < nregs; ++i) {
+ /*
+ * By default, registers are NOREF.
+ * It is possible for a callee to save them before being defined in this method,
+ * but the saved value is dead too, so it doesn't need to be marked.
+ */
+ if ((cfg->used_int_regs & (1 << i)))
+ set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
+ }
+}
- process_spill_slots (cfg);
- process_other_slots (cfg);
- process_locals (cfg);
- process_arguments (cfg);
+static void
+create_map (MonoCompile *cfg)
+{
+ GCMap *map;
+ int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
+ int ntypes [16];
+ int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
+ int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
+ int start, end;
+ gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
+ MonoCompileGC *gcfg = cfg->gc_info;
+ GCCallSite **callsites;
+ int ncallsites;
+ guint8 *bitmap, *bitmaps;
+ guint32 reg_ref_mask, reg_pin_mask;
- /* Create the GC Map */
+ ncallsites = gcfg->ncallsites;
+ nslots = gcfg->nslots;
+ nregs = gcfg->nregs;
+ callsites = gcfg->callsites;
+ /*
+ * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
+ * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
+ */
has_ref_slots = FALSE;
has_pin_slots = FALSE;
+ start = -1;
+ end = -1;
memset (ntypes, 0, sizeof (ntypes));
for (i = 0; i < nslots; ++i) {
- ntypes [slots [i]] ++;
- if (slots [i] == SLOT_REF)
+ gboolean has_ref = FALSE;
+ gboolean has_pin = FALSE;
+
+ for (j = 0; j < ncallsites; ++j) {
+ if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
+ has_pin = TRUE;
+ if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
+ has_ref = TRUE;
+ }
+
+ if (has_ref)
has_ref_slots = TRUE;
- if (slots [i] == SLOT_PIN || (slots [i] == SLOT_REF && starts_pinned [i]))
+ if (has_pin)
has_pin_slots = TRUE;
+
+ if (has_ref)
+ ntypes [SLOT_REF] ++;
+ else if (has_pin)
+ ntypes [SLOT_PIN] ++;
+ else
+ ntypes [SLOT_NOREF] ++;
+
+ if (has_ref || has_pin) {
+ if (start == -1)
+ start = i;
+ end = i + 1;
+ }
+ }
+ if (start == -1) {
+ start = end = nslots;
+ } else {
+ g_assert (start != -1);
+ g_assert (start < end);
}
+
has_ref_regs = FALSE;
- has_pin_regs = TRUE;
+ has_pin_regs = FALSE;
+ reg_ref_mask = 0;
+ reg_pin_mask = 0;
+ nref_regs = 0;
+ npin_regs = 0;
for (i = 0; i < nregs; ++i) {
- GSList *l;
- for (l = gcfg->reg_types [i]; l; l = l->next)
- if (GPOINTER_TO_UINT (l->data) == SLOT_REF)
- has_ref_regs = TRUE;
+ gboolean has_ref = FALSE;
+ gboolean has_pin = FALSE;
+
+ if (!(cfg->used_int_regs & (1 << i)))
+ continue;
+
+ for (j = 0; j < ncallsites; ++j) {
+ if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i)) {
+ has_ref = TRUE;
+ break;
+ }
+ }
+ for (j = 0; j < ncallsites; ++j) {
+ if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i)) {
+ has_pin = TRUE;
+ break;
+ }
+ }
+
+ if (has_ref) {
+ reg_ref_mask |= (1 << i);
+ has_ref_regs = TRUE;
+ nref_regs ++;
+ }
+ if (has_pin) {
+ reg_pin_mask |= (1 << i);
+ has_pin_regs = TRUE;
+ npin_regs ++;
+ }
}
if (cfg->verbose_level > 1)
- printf ("Slots: %d Refs: %d NoRefs: %d Pin: %d\n", nslots, ntypes [SLOT_REF], ntypes [SLOT_NOREF], ntypes [SLOT_PIN]);
+ printf ("Slots: %d Start: %d End: %d Refs: %d NoRefs: %d Pin: %d Callsites: %d\n", nslots, start, end, ntypes [SLOT_REF], ntypes [SLOT_NOREF], ntypes [SLOT_PIN], ncallsites);
+
+ /* Create the GC Map */
- bitmap_width = ALIGN_TO (nslots, 8) / 8;
- bitmap_size = bitmap_width * cfg->code_len;
- reg_bitmap_width = ALIGN_TO (nregs, 8) / 8;
- reg_bitmap_size = reg_bitmap_width * cfg->code_len;
- alloc_size = sizeof (GCMap);
- map = mono_domain_alloc0 (cfg->domain, alloc_size);
- gc_maps_size += alloc_size;
+ stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
+ stack_bitmap_size = stack_bitmap_width * ncallsites;
+ reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
+ reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
+ reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
+ reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
+ bitmaps_size = (has_ref_slots ? stack_bitmap_size : 0) + (has_pin_slots ? stack_bitmap_size : 0) + (has_ref_regs ? reg_ref_bitmap_size : 0) + (has_pin_regs ? reg_pin_bitmap_size : 0);
+
+ map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
map->frame_reg = cfg->frame_reg;
- map->frame_offset = gcfg->min_offset;
- map->nslots = nslots;
- map->nregs = nregs;
- map->bitmap_width = bitmap_width;
- map->reg_bitmap_width = reg_bitmap_width;
+ map->start_offset = gcfg->min_offset;
+ map->end_offset = gcfg->min_offset + (nslots * sizeof (mgreg_t));
+ map->map_offset = start * sizeof (mgreg_t);
+ map->nslots = end - start;
map->has_ref_slots = has_ref_slots;
map->has_pin_slots = has_pin_slots;
map->has_ref_regs = has_ref_regs;
map->has_pin_regs = has_pin_regs;
- map->ref_slots = mono_domain_alloc0 (cfg->domain, bitmap_width);
- gc_maps_size += bitmap_width;
+ g_assert (nregs < 32);
+ map->used_int_regs = cfg->used_int_regs;
+ map->reg_ref_mask = reg_ref_mask;
+ map->reg_pin_mask = reg_pin_mask;
+ map->nref_regs = nref_regs;
+ map->npin_regs = npin_regs;
+
+ bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
+ bitmaps_offset = 0;
if (has_ref_slots) {
- map->ref_bitmap = mono_domain_alloc0 (cfg->domain, bitmap_size);
- gc_maps_size += bitmap_size;
+ map->stack_ref_bitmap_offset = bitmaps_offset;
+ bitmaps_offset += stack_bitmap_size;
+
+ bitmap = &bitmaps [map->stack_ref_bitmap_offset];
+ for (i = 0; i < nslots; ++i) {
+ for (j = 0; j < ncallsites; ++j) {
+ if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
+ set_bit (bitmap, stack_bitmap_width, j, i - start);
+ }
+ }
}
if (has_pin_slots) {
- map->pin_bitmap = mono_domain_alloc0 (cfg->domain, bitmap_size);
- gc_maps_size += bitmap_size;
+ map->stack_pin_bitmap_offset = bitmaps_offset;
+ bitmaps_offset += stack_bitmap_size;
+
+ bitmap = &bitmaps [map->stack_pin_bitmap_offset];
+ for (i = 0; i < nslots; ++i) {
+ for (j = 0; j < ncallsites; ++j) {
+ if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
+ set_bit (bitmap, stack_bitmap_width, j, i - start);
+ }
+ }
}
if (has_ref_regs) {
- map->reg_ref_bitmap = mono_domain_alloc0 (cfg->domain, reg_bitmap_size);
- gc_maps_size += reg_bitmap_size;
+ map->reg_ref_bitmap_offset = bitmaps_offset;
+ bitmaps_offset += reg_ref_bitmap_size;
+
+ bitmap = &bitmaps [map->reg_ref_bitmap_offset];
+ bindex = 0;
+ for (i = 0; i < nregs; ++i) {
+ if (reg_ref_mask & (1 << i)) {
+ for (j = 0; j < ncallsites; ++j) {
+ if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i))
+ set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
+ }
+ bindex ++;
+ }
+ }
}
if (has_pin_regs) {
- map->reg_pin_bitmap = mono_domain_alloc0 (cfg->domain, reg_bitmap_size);
- gc_maps_size += reg_bitmap_size;
+ map->reg_pin_bitmap_offset = bitmaps_offset;
+ bitmaps_offset += reg_pin_bitmap_size;
+
+ bitmap = &bitmaps [map->reg_pin_bitmap_offset];
+ bindex = 0;
+ for (i = 0; i < nregs; ++i) {
+ if (reg_pin_mask & (1 << i)) {
+ for (j = 0; j < ncallsites; ++j) {
+ if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i))
+ set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
+ }
+ bindex ++;
+ }
+ }
}
- /* Create liveness bitmaps */
+ /* Call sites */
+ map->ncallsites = ncallsites;
+ if (cfg->code_len < 256)
+ map->callsite_entry_size = 1;
+ else if (cfg->code_len < 65536)
+ map->callsite_entry_size = 2;
+ else
+ map->callsite_entry_size = 4;
- /* Stack slots */
- for (i = 0; i < nslots; ++i) {
- int pc_offset;
+ /* Encode the GC Map */
+ {
+ guint8 buf [256];
+ guint8 *endbuf;
+ GCEncodedMap *emap;
+ int encoded_size;
+ guint8 *p;
+
+ encode_gc_map (map, buf, &endbuf);
+ g_assert (endbuf - buf < 256);
+
+ encoded_size = endbuf - buf;
+ alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
+
+ emap = mono_domain_alloc0 (cfg->domain, alloc_size);
+ //emap->ref_slots = map->ref_slots;
+
+ /* Encoded fixed fields */
+ p = &emap->encoded [0];
+ //emap->encoded_size = encoded_size;
+ memcpy (p, buf, encoded_size);
+ p += encoded_size;
+
+ /* Callsite table */
+ p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
+ if (map->callsite_entry_size == 1) {
+ guint8 *offsets = p;
+ for (i = 0; i < ncallsites; ++i)
+ offsets [i] = callsites [i]->pc_offset;
+ stats.gc_callsites8_size += ncallsites * sizeof (guint8);
+ } else if (map->callsite_entry_size == 2) {
+ guint16 *offsets = (guint16*)p;
+ for (i = 0; i < ncallsites; ++i)
+ offsets [i] = callsites [i]->pc_offset;
+ stats.gc_callsites16_size += ncallsites * sizeof (guint16);
+ } else {
+ guint32 *offsets = (guint32*)p;
+ for (i = 0; i < ncallsites; ++i)
+ offsets [i] = callsites [i]->pc_offset;
+ stats.gc_callsites32_size += ncallsites * sizeof (guint32);
+ }
+ p += ncallsites * map->callsite_entry_size;
- if (slots [i] == SLOT_REF) {
- MonoLiveInterval *iv;
- GSList *l;
- MonoLiveRange2 *r;
+ /* Bitmaps */
+ memcpy (p, bitmaps, bitmaps_size);
+ p += bitmaps_size;
- map->ref_slots [i / 8] |= (1 << i);
+ g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
- if (starts_pinned [i]) {
- /* The slots start out as pinned until they are first defined */
- g_assert (live_intervals [i]);
- g_assert (!live_intervals [i]->next);
+ stats.gc_maps_size += alloc_size;
+ stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
+ stats.gc_bitmaps_size += bitmaps_size;
+ stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
- iv = live_intervals [i]->data;
- for (pc_offset = 0; pc_offset < iv->range->from; ++pc_offset)
- set_bit (map->pin_bitmap, bitmap_width, pc_offset, i);
- }
+ cfg->jit_info->gc_info = emap;
- for (l = live_intervals [i]; l; l = l->next) {
- iv = l->data;
- for (r = iv->range; r; r = r->next) {
- for (pc_offset = r->from; pc_offset < r->to; ++pc_offset)
- set_bit (map->ref_bitmap, bitmap_width, pc_offset, i);
- }
- }
- } else if (slots [i] == SLOT_PIN) {
- for (pc_offset = 0; pc_offset < cfg->code_len; ++pc_offset)
- set_bit (map->pin_bitmap, bitmap_width, pc_offset, i);
- }
+ cfg->gc_map = (guint8*)emap;
+ cfg->gc_map_size = alloc_size;
}
- /* Registers */
- for (i = 0; i < nregs; ++i) {
- MonoLiveInterval *iv;
- GSList *l, *l2;
- MonoLiveRange2 *r;
- int pc_offset;
- StackSlotType type;
-
- if (!(cfg->used_int_regs & (1 << i))) {
- g_assert (!gcfg->reg_types [i]);
- continue;
- }
+ stats.all_slots += nslots;
+ stats.ref_slots += ntypes [SLOT_REF];
+ stats.noref_slots += ntypes [SLOT_NOREF];
+ stats.pin_slots += ntypes [SLOT_PIN];
+}
- g_assert (i < 64);
+void
+mini_gc_create_gc_map (MonoCompile *cfg)
+{
+ if (!cfg->compute_gc_maps)
+ return;
- map->used_regs |= (1 << i);
+ /*
+ * During marking, all frames except the top frame are at a call site, and we mark the
+ * top frame conservatively. This means that we only need to compute and record
+ * GC maps for call sites.
+ */
- /*
- * By default, registers are PIN.
- * This is because we don't know their type outside their live range, since
- * they could have the same value as in the caller, or a value set by the
- * current method etc.
- */
- for (pc_offset = 0; pc_offset < cfg->code_len; ++pc_offset)
- set_bit (map->reg_pin_bitmap, reg_bitmap_width, pc_offset, i);
-
- l2 = gcfg->reg_types [i];
- for (l = gcfg->reg_live_intervals [i]; l; l = l->next) {
- iv = l->data;
- type = GPOINTER_TO_UINT (l2->data);
-
- if (type == SLOT_REF) {
- for (r = iv->range; r; r = r->next) {
- for (pc_offset = r->from; pc_offset < r->to; ++pc_offset) {
- set_bit (map->reg_ref_bitmap, reg_bitmap_width, pc_offset, i);
- }
- }
- } else if (type == SLOT_PIN) {
- for (r = iv->range; r; r = r->next) {
- for (pc_offset = r->from; pc_offset < r->to; ++pc_offset) {
- set_bit (map->reg_pin_bitmap, reg_bitmap_width, pc_offset, i);
- }
- }
- } else if (type == SLOT_NOREF) {
- for (r = iv->range; r; r = r->next) {
- for (pc_offset = r->from; pc_offset < r->to; ++pc_offset) {
- clear_bit (map->reg_pin_bitmap, reg_bitmap_width, pc_offset, i);
- }
- }
- }
+ if (!(cfg->comp_done & MONO_COMP_LIVENESS))
+ /* Without liveness info, the live ranges are not precise enough */
+ return;
- l2 = l2->next;
- }
- }
+ mono_analyze_liveness_gc (cfg);
- stats.all_slots += nslots;
- for (i = 0; i < nslots; ++i) {
- if (slots [i] == SLOT_REF)
- stats.ref_slots ++;
- else if (slots [i] == SLOT_NOREF)
- stats.noref_slots ++;
- else
- stats.pin_slots ++;
- }
+ compute_frame_size (cfg);
+
+ init_gcfg (cfg);
+
+ process_spill_slots (cfg);
+ process_other_slots (cfg);
+ process_param_area_slots (cfg);
+ process_variables (cfg);
+ process_finally_clauses (cfg);
+
+ create_map (cfg);
+}
- cfg->jit_info->gc_info = map;
+static void
+parse_debug_options (void)
+{
+ char **opts, **ptr;
+ char *env;
+
+ env = getenv ("MONO_GCMAP_DEBUG");
+ if (!env)
+ return;
- g_free (live_intervals);
- g_free (slots);
- g_free (starts_pinned);
+ opts = g_strsplit (env, ",", -1);
+ for (ptr = opts; ptr && *ptr; ptr ++) {
+ /* No options yet */
+ fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
+ exit (1);
+ }
+ g_strfreev (opts);
}
void
memset (&cb, 0, sizeof (cb));
cb.thread_attach_func = thread_attach_func;
+ cb.thread_detach_func = thread_detach_func;
cb.thread_suspend_func = thread_suspend_func;
/* Comment this out to disable precise stack marking */
cb.thread_mark_func = thread_mark_func;
mono_gc_set_gc_callbacks (&cb);
+ logfile = mono_gc_get_logfile ();
+
+ parse_debug_options ();
+
mono_counters_register ("GC Maps size",
- MONO_COUNTER_GC | MONO_COUNTER_INT, &gc_maps_size);
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
+ mono_counters_register ("GC Call Sites size",
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
+ mono_counters_register ("GC Bitmaps size",
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
+ mono_counters_register ("GC Map struct size",
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
+ mono_counters_register ("GC Call Sites encoded using 8 bits",
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
+ mono_counters_register ("GC Call Sites encoded using 16 bits",
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
+ mono_counters_register ("GC Call Sites encoded using 32 bits",
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
mono_counters_register ("GC Map slots (all)",
MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
mono_counters_register ("GC Map slots (pin)",
MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
+ mono_counters_register ("GC TLS Data size",
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
+
mono_counters_register ("Stack space scanned (all)",
MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
+ mono_counters_register ("Stack space scanned (native)",
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
+ mono_counters_register ("Stack space scanned (other)",
+ MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
mono_counters_register ("Stack space scanned (using GC Maps)",
MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
mono_counters_register ("Stack space scanned (precise)",
}
void
-mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, StackSlotType type)
+mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
{
}
void
-mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, StackSlotType type)
+mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
{
}
/*
* Problems with the current code:
- * - it makes two passes over the stack
* - the stack walk is slow
- * - only the locals are scanned precisely
* - vtypes/refs used in EH regions are treated conservatively
- * - the computation of the GC maps is slow since it involves a liveness analysis pass
- * - the GC maps are uncompressed and take up a lot of memory.
* - if the code is finished, less pinning will be done, causing problems because
* we promote all surviving objects to old-gen.
+ * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
+ * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
+ * This manifests in "Unable to find ip offset x in callsite list" assertions.
+ * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
+ * precise there.
+ */
+
+/*
+ * Ideas for creating smaller GC maps:
+ * - remove empty columns from the bitmaps. This requires adding a mask bit array for
+ * each bitmap.
+ * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
+ * not wasted.
+ * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
+ * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
+ * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
+ * from an index entry.
*/
projects / mono.git / blobdiff