#include <mono/metadata/threads.h>
#include <mono/metadata/profiler-private.h>
#include <mono/metadata/mono-debug.h>
+#include <mono/metadata/gc-internal.h>
#include <mono/utils/mono-math.h>
#include <mono/utils/mono-counters.h>
+#include <mono/utils/mono-mmap.h>
#include "trace.h"
#include "mini-x86.h"
static gint lmf_tls_offset = -1;
static gint lmf_addr_tls_offset = -1;
static gint appdomain_tls_offset = -1;
-static gint thread_tls_offset = -1;
#ifdef MONO_XEN_OPT
static gboolean optimize_for_xen = TRUE;
#define optimize_for_xen 0
#endif
-#ifdef PLATFORM_WIN32
+#ifdef TARGET_WIN32
static gboolean is_win32 = TRUE;
#else
static gboolean is_win32 = FALSE;
#define ARGS_OFFSET 8
-#ifdef PLATFORM_WIN32
+#ifdef TARGET_WIN32
/* Under windows, the default pinvoke calling convention is stdcall */
#define CALLCONV_IS_STDCALL(sig) ((((sig)->call_convention) == MONO_CALL_STDCALL) || ((sig)->pinvoke && ((sig)->call_convention) == MONO_CALL_DEFAULT))
#else
MonoBreakpointInfo
mono_breakpoint_info [MONO_BREAKPOINT_ARRAY_SIZE];
+static gpointer
+mono_realloc_native_code (MonoCompile *cfg)
+{
+#ifdef __native_client_codegen__
+ guint old_padding;
+ gpointer native_code;
+ guint alignment_check;
+
+ /* Save the old alignment offset so we can re-align after the realloc. */
+ old_padding = (guint)(cfg->native_code - cfg->native_code_alloc);
+
+ cfg->native_code_alloc = g_realloc (cfg->native_code_alloc,
+ cfg->code_size + kNaClAlignment);
+
+ /* Align native_code to next nearest kNaClAlignment byte. */
+ native_code = (guint)cfg->native_code_alloc + kNaClAlignment;
+ native_code = (guint)native_code & ~kNaClAlignmentMask;
+
+ /* Shift the data to be 32-byte aligned again. */
+ memmove (native_code, cfg->native_code_alloc + old_padding, cfg->code_size);
+
+ alignment_check = (guint)native_code & kNaClAlignmentMask;
+ g_assert (alignment_check == 0);
+ return native_code;
+#else
+ return g_realloc (cfg->native_code, cfg->code_size);
+#endif
+}
+
+#ifdef __native_client_codegen__
+
+/* mono_arch_nacl_pad: Add pad bytes of alignment instructions at code, */
+/* Check that alignment doesn't cross an alignment boundary. */
+guint8 *
+mono_arch_nacl_pad (guint8 *code, int pad)
+{
+ const int kMaxPadding = 7; /* see x86-codegen.h: x86_padding() */
+
+ if (pad == 0) return code;
+ /* assertion: alignment cannot cross a block boundary */
+ g_assert(((uintptr_t)code & (~kNaClAlignmentMask)) ==
+ (((uintptr_t)code + pad - 1) & (~kNaClAlignmentMask)));
+ while (pad >= kMaxPadding) {
+ x86_padding (code, kMaxPadding);
+ pad -= kMaxPadding;
+ }
+ if (pad != 0) x86_padding (code, pad);
+ return code;
+}
+
+guint8 *
+mono_arch_nacl_skip_nops (guint8 *code)
+{
+ x86_skip_nops (code);
+ return code;
+}
+
+#endif /* __native_client_codegen__ */
+
+/*
+ * The code generated for sequence points reads from this location, which is
+ * made read-only when single stepping is enabled.
+ */
+static gpointer ss_trigger_page;
+
+/* Enabled breakpoints read from this trigger page */
+static gpointer bp_trigger_page;
+
const char*
mono_arch_regname (int reg)
{
}
}
+void
+mono_x86_patch (unsigned char* code, gpointer target)
+{
+ x86_patch (code, (unsigned char*)target);
+}
typedef enum {
ArgInIReg,
guint32 freg_usage;
gboolean need_stack_align;
guint32 stack_align_amount;
+ gboolean vtype_retaddr;
+ /* The index of the vret arg in the argument list */
+ int vret_arg_index;
ArgInfo ret;
ArgInfo sig_cookie;
ArgInfo args [1];
static X86_Reg_No param_regs [] = { 0 };
-#if defined(PLATFORM_WIN32) || defined(__APPLE__) || defined(__FreeBSD__)
+#if defined(TARGET_WIN32) || defined(__APPLE__) || defined(__FreeBSD__)
#define SMALL_STRUCTS_IN_REGS
static X86_Reg_No return_regs [] = { X86_EAX, X86_EDX };
#endif
* For x86 win32, see ???.
*/
static CallInfo*
-get_call_info (MonoGenericSharingContext *gsctx, MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
+get_call_info_internal (MonoGenericSharingContext *gsctx, CallInfo *cinfo, MonoMethodSignature *sig, gboolean is_pinvoke)
{
- guint32 i, gr, fr;
+ guint32 i, gr, fr, pstart;
MonoType *ret_type;
int n = sig->hasthis + sig->param_count;
guint32 stack_size = 0;
- CallInfo *cinfo;
-
- if (mp)
- cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
- else
- cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
gr = 0;
fr = 0;
cinfo->ret.storage = ArgOnDoubleFpStack;
break;
case MONO_TYPE_GENERICINST:
- if (!mono_type_generic_inst_is_valuetype (sig->ret)) {
+ if (!mono_type_generic_inst_is_valuetype (ret_type)) {
cinfo->ret.storage = ArgInIReg;
cinfo->ret.reg = X86_EAX;
break;
guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
add_valuetype (gsctx, sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
- if (cinfo->ret.storage == ArgOnStack)
+ if (cinfo->ret.storage == ArgOnStack) {
+ cinfo->vtype_retaddr = TRUE;
/* The caller passes the address where the value is stored */
- add_general (&gr, &stack_size, &cinfo->ret);
+ }
break;
}
case MONO_TYPE_TYPEDBYREF:
- /* Same as a valuetype with size 24 */
- add_general (&gr, &stack_size, &cinfo->ret);
- ;
+ /* Same as a valuetype with size 12 */
+ cinfo->vtype_retaddr = TRUE;
break;
case MONO_TYPE_VOID:
cinfo->ret.storage = ArgNone;
}
}
- /* this */
- if (sig->hasthis)
- add_general (&gr, &stack_size, cinfo->args + 0);
+ pstart = 0;
+ /*
+ * To simplify get_this_arg_reg () and LLVM integration, emit the vret arg after
+ * the first argument, allowing 'this' to be always passed in the first arg reg.
+ * Also do this if the first argument is a reference type, since virtual calls
+ * are sometimes made using calli without sig->hasthis set, like in the delegate
+ * invoke wrappers.
+ */
+ if (cinfo->vtype_retaddr && !is_pinvoke && (sig->hasthis || (sig->param_count > 0 && MONO_TYPE_IS_REFERENCE (mini_type_get_underlying_type (gsctx, sig->params [0]))))) {
+ if (sig->hasthis) {
+ add_general (&gr, &stack_size, cinfo->args + 0);
+ } else {
+ add_general (&gr, &stack_size, &cinfo->args [sig->hasthis + 0]);
+ pstart = 1;
+ }
+ add_general (&gr, &stack_size, &cinfo->ret);
+ cinfo->vret_arg_index = 1;
+ } else {
+ /* this */
+ if (sig->hasthis)
+ add_general (&gr, &stack_size, cinfo->args + 0);
+
+ if (cinfo->vtype_retaddr)
+ add_general (&gr, &stack_size, &cinfo->ret);
+ }
if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
gr = PARAM_REGS;
add_general (&gr, &stack_size, &cinfo->sig_cookie);
}
- for (i = 0; i < sig->param_count; ++i) {
+ for (i = pstart; i < sig->param_count; ++i) {
ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
MonoType *ptype;
add_general (&gr, &stack_size, ainfo);
break;
case MONO_TYPE_GENERICINST:
- if (!mono_type_generic_inst_is_valuetype (sig->params [i])) {
+ if (!mono_type_generic_inst_is_valuetype (ptype)) {
add_general (&gr, &stack_size, ainfo);
break;
}
return cinfo;
}
+static CallInfo*
+get_call_info (MonoGenericSharingContext *gsctx, MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
+{
+ int n = sig->hasthis + sig->param_count;
+ CallInfo *cinfo;
+
+ if (mp)
+ cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
+ else
+ cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
+
+ return get_call_info_internal (gsctx, cinfo, sig, is_pinvoke);
+}
+
/*
* mono_arch_get_argument_info:
* @csig: a method signature
* padding. arg_info should be large enought to hold param_count + 1 entries.
*
* Returns the size of the argument area on the stack.
+ * This should be signal safe, since it is called from
+ * mono_arch_find_jit_info_ext ().
+ * FIXME: The metadata calls might not be signal safe.
*/
int
mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
{
- int k, args_size = 0;
+ int len, k, args_size = 0;
int size, pad;
guint32 align;
int offset = 8;
CallInfo *cinfo;
- cinfo = get_call_info (NULL, NULL, csig, FALSE);
+ /* Avoid g_malloc as it is not signal safe */
+ len = sizeof (CallInfo) + (sizeof (ArgInfo) * (csig->param_count + 1));
+ cinfo = (CallInfo*)g_newa (guint8*, len);
+ memset (cinfo, 0, len);
- if (MONO_TYPE_ISSTRUCT (csig->ret) && (cinfo->ret.storage == ArgOnStack)) {
- args_size += sizeof (gpointer);
- offset += 4;
- }
+ cinfo = get_call_info_internal (NULL, cinfo, csig, FALSE);
arg_info [0].offset = offset;
offset += 4;
}
+ if (MONO_TYPE_ISSTRUCT (csig->ret) && (cinfo->ret.storage == ArgOnStack)) {
+ args_size += sizeof (gpointer);
+ offset += 4;
+ }
+
arg_info [0].size = args_size;
for (k = 0; k < param_count; k++) {
args_size += pad = (align - (args_size & (align - 1))) & (align - 1);
arg_info [k].pad = pad;
- g_free (cinfo);
-
return args_size;
}
static int
cpuid (int id, int* p_eax, int* p_ebx, int* p_ecx, int* p_edx)
{
+#if defined(__native_client__)
+ /* Taken from below, the bug listed in the comment is */
+ /* only valid for non-static cases. */
+ __asm__ __volatile__ ("cpuid"
+ : "=a" (*p_eax), "=b" (*p_ebx), "=c" (*p_ecx), "=d" (*p_edx)
+ : "a" (id));
+ return 1;
+#else
int have_cpuid = 0;
#ifndef _MSC_VER
__asm__ __volatile__ (
return 1;
}
return 0;
+#endif
}
/*
mono_arch_init (void)
{
InitializeCriticalSection (&mini_arch_mutex);
+
+ ss_trigger_page = mono_valloc (NULL, mono_pagesize (), MONO_MMAP_READ);
+ bp_trigger_page = mono_valloc (NULL, mono_pagesize (), MONO_MMAP_READ|MONO_MMAP_32BIT);
+ mono_mprotect (bp_trigger_page, mono_pagesize (), 0);
+
+ mono_aot_register_jit_icall ("mono_x86_throw_exception", mono_x86_throw_exception);
+ mono_aot_register_jit_icall ("mono_x86_throw_corlib_exception", mono_x86_throw_corlib_exception);
}
/*
guint32
mono_arch_cpu_optimizazions (guint32 *exclude_mask)
{
+#if !defined(__native_client__)
int eax, ebx, ecx, edx;
guint32 opts = 0;
*exclude_mask = 0;
+
+ if (mono_aot_only)
+ /* The cpuid function allocates from the global codeman */
+ return opts;
+
/* Feature Flags function, flags returned in EDX. */
if (cpuid (1, &eax, &ebx, &ecx, &edx)) {
if (edx & (1 << 15)) {
#endif
}
return opts;
+#else
+ return MONO_OPT_CMOV | MONO_OPT_FCMOV | MONO_OPT_SSE2;
+#endif
}
/*
*
* Returns a bitmask corresponding to all supported versions.
*
- * TODO detect other versions like SSE4a.
*/
guint32
mono_arch_cpu_enumerate_simd_versions (void)
int eax, ebx, ecx, edx;
guint32 sse_opts = 0;
+ if (mono_aot_only)
+ /* The cpuid function allocates from the global codeman */
+ return sse_opts;
+
if (cpuid (1, &eax, &ebx, &ecx, &edx)) {
if (edx & (1 << 25))
- sse_opts |= 1 << SIMD_VERSION_SSE1;
+ sse_opts |= SIMD_VERSION_SSE1;
if (edx & (1 << 26))
- sse_opts |= 1 << SIMD_VERSION_SSE2;
+ sse_opts |= SIMD_VERSION_SSE2;
if (ecx & (1 << 0))
- sse_opts |= 1 << SIMD_VERSION_SSE3;
+ sse_opts |= SIMD_VERSION_SSE3;
if (ecx & (1 << 9))
- sse_opts |= 1 << SIMD_VERSION_SSSE3;
+ sse_opts |= SIMD_VERSION_SSSE3;
if (ecx & (1 << 19))
- sse_opts |= 1 << SIMD_VERSION_SSE41;
+ sse_opts |= SIMD_VERSION_SSE41;
if (ecx & (1 << 20))
- sse_opts |= 1 << SIMD_VERSION_SSE42;
+ sse_opts |= SIMD_VERSION_SSE42;
+ }
+
+ /* Yes, all this needs to be done to check for sse4a.
+ See: "Amd: CPUID Specification"
+ */
+ if (cpuid (0x80000000, &eax, &ebx, &ecx, &edx)) {
+ /* eax greater or equal than 0x80000001, ebx = 'htuA', ecx = DMAc', edx = 'itne'*/
+ if ((((unsigned int) eax) >= 0x80000001) && (ebx == 0x68747541) && (ecx == 0x444D4163) && (edx == 0x69746E65)) {
+ cpuid (0x80000001, &eax, &ebx, &ecx, &edx);
+ if (ecx & (1 << 6))
+ sse_opts |= SIMD_VERSION_SSE4a;
+ }
}
+
+
return sse_opts;
}
MonoMethodHeader *header;
gboolean result = FALSE;
+#if defined(__APPLE__)
+ /*OSX requires stack frame code to have the correct alignment. */
+ return TRUE;
+#endif
+
if (cfg->arch.need_stack_frame_inited)
return cfg->arch.need_stack_frame;
- header = mono_method_get_header (cfg->method);
+ header = cfg->header;
sig = mono_method_signature (cfg->method);
if (cfg->disable_omit_fp)
gint32 *offsets;
CallInfo *cinfo;
- header = mono_method_get_header (cfg->method);
+ header = cfg->header;
sig = mono_method_signature (cfg->method);
cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig, FALSE);
/* Allocate locals */
offsets = mono_allocate_stack_slots (cfg, &locals_stack_size, &locals_stack_align);
+ if (locals_stack_size > MONO_ARCH_MAX_FRAME_SIZE) {
+ char *mname = mono_method_full_name (cfg->method, TRUE);
+ cfg->exception_type = MONO_EXCEPTION_INVALID_PROGRAM;
+ cfg->exception_message = g_strdup_printf ("Method %s stack is too big.", mname);
+ g_free (mname);
+ return;
+ }
if (locals_stack_align) {
offset += (locals_stack_align - 1);
offset &= ~(locals_stack_align - 1);
MONO_EMIT_NEW_BIALU_IMM (cfg, OP_X86_PUSH_IMM, -1, -1, tmp_sig);
}
+#ifdef ENABLE_LLVM
+LLVMCallInfo*
+mono_arch_get_llvm_call_info (MonoCompile *cfg, MonoMethodSignature *sig)
+{
+ int i, n;
+ CallInfo *cinfo;
+ ArgInfo *ainfo;
+ LLVMCallInfo *linfo;
+ MonoType *t;
+
+ n = sig->param_count + sig->hasthis;
+
+ cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig, sig->pinvoke);
+
+ linfo = mono_mempool_alloc0 (cfg->mempool, sizeof (LLVMCallInfo) + (sizeof (LLVMArgInfo) * n));
+
+ /*
+ * LLVM always uses the native ABI while we use our own ABI, the
+ * only difference is the handling of vtypes:
+ * - we only pass/receive them in registers in some cases, and only
+ * in 1 or 2 integer registers.
+ */
+ if (cinfo->ret.storage == ArgValuetypeInReg) {
+ if (sig->pinvoke) {
+ cfg->exception_message = g_strdup ("pinvoke + vtypes");
+ cfg->disable_llvm = TRUE;
+ return linfo;
+ }
+
+ cfg->exception_message = g_strdup ("vtype ret in call");
+ cfg->disable_llvm = TRUE;
+ /*
+ linfo->ret.storage = LLVMArgVtypeInReg;
+ for (j = 0; j < 2; ++j)
+ linfo->ret.pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, cinfo->ret.pair_storage [j]);
+ */
+ }
+
+ if (MONO_TYPE_ISSTRUCT (sig->ret) && cinfo->ret.storage == ArgInIReg) {
+ /* Vtype returned using a hidden argument */
+ linfo->ret.storage = LLVMArgVtypeRetAddr;
+ linfo->vret_arg_index = cinfo->vret_arg_index;
+ }
+
+ if (MONO_TYPE_ISSTRUCT (sig->ret) && cinfo->ret.storage != ArgInIReg) {
+ // FIXME:
+ cfg->exception_message = g_strdup ("vtype ret in call");
+ cfg->disable_llvm = TRUE;
+ }
+
+ for (i = 0; i < n; ++i) {
+ ainfo = cinfo->args + i;
+
+ if (i >= sig->hasthis)
+ t = sig->params [i - sig->hasthis];
+ else
+ t = &mono_defaults.int_class->byval_arg;
+
+ linfo->args [i].storage = LLVMArgNone;
+
+ switch (ainfo->storage) {
+ case ArgInIReg:
+ linfo->args [i].storage = LLVMArgInIReg;
+ break;
+ case ArgInDoubleSSEReg:
+ case ArgInFloatSSEReg:
+ linfo->args [i].storage = LLVMArgInFPReg;
+ break;
+ case ArgOnStack:
+ if (MONO_TYPE_ISSTRUCT (t)) {
+ if (mono_class_value_size (mono_class_from_mono_type (t), NULL) == 0)
+ /* LLVM seems to allocate argument space for empty structures too */
+ linfo->args [i].storage = LLVMArgNone;
+ else
+ linfo->args [i].storage = LLVMArgVtypeByVal;
+ } else {
+ linfo->args [i].storage = LLVMArgInIReg;
+ if (t->byref) {
+ if (t->type == MONO_TYPE_R4)
+ linfo->args [i].storage = LLVMArgInFPReg;
+ else if (t->type == MONO_TYPE_R8)
+ linfo->args [i].storage = LLVMArgInFPReg;
+ }
+ }
+ break;
+ case ArgValuetypeInReg:
+ if (sig->pinvoke) {
+ cfg->exception_message = g_strdup ("pinvoke + vtypes");
+ cfg->disable_llvm = TRUE;
+ return linfo;
+ }
+
+ cfg->exception_message = g_strdup ("vtype arg");
+ cfg->disable_llvm = TRUE;
+ /*
+ linfo->args [i].storage = LLVMArgVtypeInReg;
+ for (j = 0; j < 2; ++j)
+ linfo->args [i].pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, ainfo->pair_storage [j]);
+ */
+ break;
+ default:
+ cfg->exception_message = g_strdup ("ainfo->storage");
+ cfg->disable_llvm = TRUE;
+ break;
+ }
+ }
+
+ return linfo;
+}
+#endif
+
void
mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
{
}
if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
- MonoInst *vtarg;
-
if (cinfo->ret.storage == ArgValuetypeInReg) {
- if (cinfo->ret.pair_storage [0] == ArgInIReg && cinfo->ret.pair_storage [1] == ArgNone) {
- /*
- * Tell the JIT to use a more efficient calling convention: call using
- * OP_CALL, compute the result location after the call, and save the
- * result there.
- */
- call->vret_in_reg = TRUE;
- } else {
- /*
- * The valuetype is in EAX:EDX after the call, needs to be copied to
- * the stack. Save the address here, so the call instruction can
- * access it.
- */
- MONO_INST_NEW (cfg, vtarg, OP_X86_PUSH);
- vtarg->sreg1 = call->vret_var->dreg;
- MONO_ADD_INS (cfg->cbb, vtarg);
- }
+ /*
+ * Tell the JIT to use a more efficient calling convention: call using
+ * OP_CALL, compute the result location after the call, and save the
+ * result there.
+ */
+ call->vret_in_reg = TRUE;
+ if (call->vret_var)
+ NULLIFY_INS (call->vret_var);
}
}
ArgInfo *ainfo = cinfo->args + i;
MonoType *t;
+ if (cinfo->vtype_retaddr && cinfo->vret_arg_index == 1 && i == 0) {
+ /* Push the vret arg before the first argument */
+ MonoInst *vtarg;
+ MONO_INST_NEW (cfg, vtarg, OP_X86_PUSH);
+ vtarg->type = STACK_MP;
+ vtarg->sreg1 = call->vret_var->dreg;
+ MONO_ADD_INS (cfg->cbb, vtarg);
+ }
+
if (i >= sig->hasthis)
t = sig->params [i - sig->hasthis];
else
MONO_ADD_INS (cfg->cbb, vtarg);
mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, cinfo->ret.reg, FALSE);
- } else {
+ } else if (cinfo->vtype_retaddr && cinfo->vret_arg_index == 0) {
MonoInst *vtarg;
MONO_INST_NEW (cfg, vtarg, OP_X86_PUSH);
vtarg->type = STACK_MP;
MONO_ADD_INS (cfg->cbb, vtarg);
}
- /* if the function returns a struct, the called method already does a ret $0x4 */
- cinfo->stack_usage -= 4;
+ /* if the function returns a struct on stack, the called method already does a ret $0x4 */
+ if (cinfo->ret.storage != ArgValuetypeInReg)
+ cinfo->stack_usage -= 4;
}
call->stack_usage = cinfo->stack_usage;
if (!ret->byref) {
if (ret->type == MONO_TYPE_R4) {
+ if (COMPILE_LLVM (cfg))
+ MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
/* Nothing to do */
return;
} else if (ret->type == MONO_TYPE_R8) {
+ if (COMPILE_LLVM (cfg))
+ MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
/* Nothing to do */
return;
} else if (ret->type == MONO_TYPE_I8 || ret->type == MONO_TYPE_U8) {
- MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, X86_EAX, val->dreg + 1);
- MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, X86_EDX, val->dreg + 2);
+ if (COMPILE_LLVM (cfg))
+ MONO_EMIT_NEW_UNALU (cfg, OP_LMOVE, cfg->ret->dreg, val->dreg);
+ else {
+ MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, X86_EAX, val->dreg + 1);
+ MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, X86_EDX, val->dreg + 2);
+ }
return;
}
}
};
void*
-mono_arch_instrument_epilog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
+mono_arch_instrument_epilog_full (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments, gboolean preserve_argument_registers)
{
guchar *code = p;
- int arg_size = 0, save_mode = SAVE_NONE;
+ int arg_size = 0, stack_usage = 0, save_mode = SAVE_NONE;
MonoMethod *method = cfg->method;
-
- switch (mini_type_get_underlying_type (cfg->generic_sharing_context, mono_method_signature (method)->ret)->type) {
+ MonoType *ret_type = mini_type_get_underlying_type (cfg->generic_sharing_context, mono_method_signature (method)->ret);
+
+ switch (ret_type->type) {
case MONO_TYPE_VOID:
/* special case string .ctor icall */
- if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
+ if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class) {
save_mode = SAVE_EAX;
- else
+ stack_usage = enable_arguments ? 8 : 4;
+ } else
save_mode = SAVE_NONE;
break;
case MONO_TYPE_I8:
case MONO_TYPE_U8:
save_mode = SAVE_EAX_EDX;
+ stack_usage = enable_arguments ? 16 : 8;
break;
case MONO_TYPE_R4:
case MONO_TYPE_R8:
save_mode = SAVE_FP;
+ stack_usage = enable_arguments ? 16 : 8;
break;
case MONO_TYPE_GENERICINST:
- if (!mono_type_generic_inst_is_valuetype (mono_method_signature (method)->ret)) {
+ if (!mono_type_generic_inst_is_valuetype (ret_type)) {
save_mode = SAVE_EAX;
+ stack_usage = enable_arguments ? 8 : 4;
break;
}
/* Fall through */
case MONO_TYPE_VALUETYPE:
+ // FIXME: Handle SMALL_STRUCT_IN_REG here for proper alignment on darwin-x86
save_mode = SAVE_STRUCT;
+ stack_usage = enable_arguments ? 4 : 0;
break;
default:
save_mode = SAVE_EAX;
+ stack_usage = enable_arguments ? 8 : 4;
break;
}
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, MONO_ARCH_FRAME_ALIGNMENT - stack_usage - 4);
+
switch (save_mode) {
case SAVE_EAX_EDX:
x86_push_reg (code, X86_EDX);
mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
x86_call_code (code, 0);
}
+
x86_alu_reg_imm (code, X86_ADD, X86_ESP, arg_size + 4);
switch (save_mode) {
default:
break;
}
+
+ x86_alu_reg_imm (code, X86_ADD, X86_ESP, MONO_ARCH_FRAME_ALIGNMENT - stack_usage);
return code;
}
#define EMIT_COND_BRANCH(ins,cond,sign) \
-if (ins->flags & MONO_INST_BRLABEL) { \
- if (ins->inst_i0->inst_c0) { \
- x86_branch (code, cond, cfg->native_code + ins->inst_i0->inst_c0, sign); \
- } else { \
- mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_LABEL, ins->inst_i0); \
- if ((cfg->opt & MONO_OPT_BRANCH) && \
- x86_is_imm8 (ins->inst_i0->inst_c1 - cpos)) \
- x86_branch8 (code, cond, 0, sign); \
- else \
- x86_branch32 (code, cond, 0, sign); \
- } \
+if (ins->inst_true_bb->native_offset) { \
+ x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
} else { \
- if (ins->inst_true_bb->native_offset) { \
- x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
- } else { \
- mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
- if ((cfg->opt & MONO_OPT_BRANCH) && \
- x86_is_imm8 (ins->inst_true_bb->max_offset - cpos)) \
- x86_branch8 (code, cond, 0, sign); \
- else \
- x86_branch32 (code, cond, 0, sign); \
- } \
+ mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
+ if ((cfg->opt & MONO_OPT_BRANCH) && \
+ x86_is_imm8 (ins->inst_true_bb->max_offset - cpos)) \
+ x86_branch8 (code, cond, 0, sign); \
+ else \
+ x86_branch32 (code, cond, 0, sign); \
}
/*
int sreg = tree->sreg1;
int need_touch = FALSE;
-#if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
+#if defined(TARGET_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
need_touch = TRUE;
#endif
static guint8*
emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
{
- CallInfo *cinfo;
- int quad;
-
/* Move return value to the target register */
switch (ins->opcode) {
case OP_CALL:
if (ins->dreg != X86_EAX)
x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
break;
- case OP_VCALL:
- case OP_VCALL_REG:
- case OP_VCALL_MEMBASE:
- case OP_VCALL2:
- case OP_VCALL2_REG:
- case OP_VCALL2_MEMBASE:
- cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, ((MonoCallInst*)ins)->signature, FALSE);
- if (cinfo->ret.storage == ArgValuetypeInReg) {
- /* Pop the destination address from the stack */
- x86_pop_reg (code, X86_ECX);
-
- for (quad = 0; quad < 2; quad ++) {
- switch (cinfo->ret.pair_storage [quad]) {
- case ArgInIReg:
- g_assert (cinfo->ret.pair_regs [quad] != X86_ECX);
- x86_mov_membase_reg (code, X86_ECX, (quad * sizeof (gpointer)), cinfo->ret.pair_regs [quad], sizeof (gpointer));
- break;
- case ArgNone:
- break;
- default:
- g_assert_not_reached ();
- }
- }
- }
- break;
- case OP_FCALL: {
- MonoCallInst *call = (MonoCallInst*)ins;
- if (call->method && !mono_method_signature (call->method)->ret->byref && mono_method_signature (call->method)->ret->type == MONO_TYPE_R4) {
- /* Avoid some precision issues by saving/reloading the return value */
- x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
- x86_fst_membase (code, X86_ESP, 0, FALSE, TRUE);
- x86_fld_membase (code, X86_ESP, 0, FALSE);
- x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
- }
- break;
- }
default:
break;
}
return code;
}
+gboolean
+mono_x86_have_tls_get (void)
+{
+#ifdef __APPLE__
+ guint32 *ins = (guint32*)pthread_getspecific;
+ /*
+ * We're looking for these two instructions:
+ *
+ * mov 0x4(%esp),%eax
+ * mov %gs:0x48(,%eax,4),%eax
+ */
+ return ins [0] == 0x0424448b && ins [1] == 0x85048b65 && ins [2] == 0x00000048;
+#else
+ return TRUE;
+#endif
+}
+
/*
* mono_x86_emit_tls_get:
* @code: buffer to store code to
guint8*
mono_x86_emit_tls_get (guint8* code, int dreg, int tls_offset)
{
-#ifdef PLATFORM_WIN32
+#if defined(__APPLE__)
+ x86_prefix (code, X86_GS_PREFIX);
+ x86_mov_reg_mem (code, dreg, 0x48 + tls_offset * 4, 4);
+#elif defined(TARGET_WIN32)
/*
* See the Under the Hood article in the May 1996 issue of Microsoft Systems
* Journal and/or a disassembly of the TlsGet () function.
x86_pop_reg (code, X86_EDX); \
x86_pop_reg (code, X86_EAX);
+/* REAL_PRINT_REG does not appear to be used, and was not adapted to work with Native Client. */
+#ifdef __native__client_codegen__
+#define REAL_PRINT_REG(text, reg) g_assert_not_reached()
+#endif
+
/* benchmark and set based on cpu */
#define LOOP_ALIGNMENT 8
#define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
bb->native_offset = cfg->code_len;
}
}
-
+#ifdef __native_client_codegen__
+ {
+ /* For Native Client, all indirect call/jump targets must be */
+ /* 32-byte aligned. Exception handler blocks are jumped to */
+ /* indirectly as well. */
+ gboolean bb_needs_alignment = (bb->flags & BB_INDIRECT_JUMP_TARGET) ||
+ (bb->flags & BB_EXCEPTION_HANDLER);
+
+ /* if ((cfg->code_len & kNaClAlignmentMask) != 0) { */
+ if ( bb_needs_alignment && ((cfg->code_len & kNaClAlignmentMask) != 0)) {
+ int pad = kNaClAlignment - (cfg->code_len & kNaClAlignmentMask);
+ if (pad != kNaClAlignment) code = mono_arch_nacl_pad(code, pad);
+ cfg->code_len += pad;
+ bb->native_offset = cfg->code_len;
+ }
+ }
+#endif /* __native_client_codegen__ */
if (cfg->verbose_level > 2)
g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
mono_debug_open_block (cfg, bb, offset);
+ if (mono_break_at_bb_method && mono_method_desc_full_match (mono_break_at_bb_method, cfg->method) && bb->block_num == mono_break_at_bb_bb_num)
+ x86_breakpoint (code);
+
MONO_BB_FOR_EACH_INS (bb, ins) {
offset = code - cfg->native_code;
max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
- if (G_UNLIKELY (offset > (cfg->code_size - max_len - 16))) {
+#define EXTRA_CODE_SPACE (NACL_SIZE (16, 16 + kNaClAlignment))
+
+ if (G_UNLIKELY (offset > (cfg->code_size - max_len - EXTRA_CODE_SPACE))) {
cfg->code_size *= 2;
- cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
+ cfg->native_code = mono_realloc_native_code(cfg);
code = cfg->native_code + offset;
mono_jit_stats.code_reallocs++;
}
case OP_NOT_REACHED:
case OP_NOT_NULL:
break;
+ case OP_SEQ_POINT: {
+ int i;
+
+ if (cfg->compile_aot)
+ NOT_IMPLEMENTED;
+
+ /*
+ * Read from the single stepping trigger page. This will cause a
+ * SIGSEGV when single stepping is enabled.
+ * We do this _before_ the breakpoint, so single stepping after
+ * a breakpoint is hit will step to the next IL offset.
+ */
+ if (ins->flags & MONO_INST_SINGLE_STEP_LOC)
+ x86_alu_reg_mem (code, X86_CMP, X86_EAX, (guint32)ss_trigger_page);
+
+ mono_add_seq_point (cfg, bb, ins, code - cfg->native_code);
+
+ /*
+ * A placeholder for a possible breakpoint inserted by
+ * mono_arch_set_breakpoint ().
+ */
+ for (i = 0; i < 6; ++i)
+ x86_nop (code);
+ break;
+ }
case OP_ADDCC:
case OP_IADDCC:
case OP_IADD:
*/
x86_alu_reg_reg (code, X86_XOR, X86_EAX, X86_EDX);
x86_alu_reg_reg (code, X86_SUB, X86_EAX, X86_EDX);
+ } else if (power == 0) {
+ x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
} else {
/* Based on gcc code */
x86_mov_reg_imm (code, ins->dreg, 0);
break;
case OP_LOAD_GOTADDR:
- x86_call_imm (code, 0);
- /*
- * The patch needs to point to the pop, since the GOT offset needs
- * to be added to that address.
- */
- mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GOT_OFFSET, NULL);
- x86_pop_reg (code, ins->dreg);
- x86_alu_reg_imm (code, X86_ADD, ins->dreg, 0xf0f0f0f0);
+ g_assert (ins->dreg == MONO_ARCH_GOT_REG);
+ code = mono_arch_emit_load_got_addr (cfg->native_code, code, cfg, NULL);
break;
case OP_GOT_ENTRY:
mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_right->inst_i1, ins->inst_right->inst_p0);
case OP_VOIDCALL_MEMBASE:
case OP_CALL_MEMBASE:
call = (MonoCallInst*)ins;
+
x86_call_membase (code, ins->sreg1, ins->inst_offset);
if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
if (call->stack_usage == 4)
break;
}
case OP_THROW: {
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, MONO_ARCH_FRAME_ALIGNMENT - 4);
x86_push_reg (code, ins->sreg1);
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
(gpointer)"mono_arch_throw_exception");
break;
}
case OP_RETHROW: {
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, MONO_ARCH_FRAME_ALIGNMENT - 4);
x86_push_reg (code, ins->sreg1);
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
(gpointer)"mono_arch_rethrow_exception");
x86_alu_reg_imm (code, X86_SUB, X86_ESP, MONO_ARCH_FRAME_ALIGNMENT - 4);
mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
x86_call_imm (code, 0);
+ mono_cfg_add_try_hole (cfg, ins->inst_eh_block, code, bb);
x86_alu_reg_imm (code, X86_ADD, X86_ESP, MONO_ARCH_FRAME_ALIGNMENT - 4);
break;
case OP_START_HANDLER: {
ins->inst_c0 = code - cfg->native_code;
break;
case OP_BR:
- if (ins->flags & MONO_INST_BRLABEL) {
- if (ins->inst_i0->inst_c0) {
- x86_jump_code (code, cfg->native_code + ins->inst_i0->inst_c0);
- } else {
- mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_LABEL, ins->inst_i0);
- if ((cfg->opt & MONO_OPT_BRANCH) &&
- x86_is_imm8 (ins->inst_i0->inst_c1 - cpos))
- x86_jump8 (code, 0);
- else
- x86_jump32 (code, 0);
- }
+ if (ins->inst_target_bb->native_offset) {
+ x86_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
} else {
- if (ins->inst_target_bb->native_offset) {
- x86_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
- } else {
- mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
- if ((cfg->opt & MONO_OPT_BRANCH) &&
- x86_is_imm8 (ins->inst_target_bb->max_offset - cpos))
- x86_jump8 (code, 0);
- else
- x86_jump32 (code, 0);
- }
+ mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
+ if ((cfg->opt & MONO_OPT_BRANCH) &&
+ x86_is_imm8 (ins->inst_target_bb->max_offset - cpos))
+ x86_jump8 (code, 0);
+ else
+ x86_jump32 (code, 0);
}
break;
case OP_BR_REG:
case OP_STORER8_MEMBASE_REG:
x86_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, TRUE, TRUE);
break;
- case OP_LOADR8_SPILL_MEMBASE:
- x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
- x86_fxch (code, 1);
- break;
case OP_LOADR8_MEMBASE:
x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
break;
case OP_LOADR4_MEMBASE:
x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
break;
- case OP_ICONV_TO_R4: /* FIXME: change precision */
+ case OP_ICONV_TO_R4:
+ x86_push_reg (code, ins->sreg1);
+ x86_fild_membase (code, X86_ESP, 0, FALSE);
+ /* Change precision */
+ x86_fst_membase (code, X86_ESP, 0, FALSE, TRUE);
+ x86_fld_membase (code, X86_ESP, 0, FALSE);
+ x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
+ break;
case OP_ICONV_TO_R8:
x86_push_reg (code, ins->sreg1);
x86_fild_membase (code, X86_ESP, 0, FALSE);
x86_fild_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
break;
case OP_FCONV_TO_R4:
- /* FIXME: nothing to do ?? */
+ /* Change precision */
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
+ x86_fst_membase (code, X86_ESP, 0, FALSE, TRUE);
+ x86_fld_membase (code, X86_ESP, 0, FALSE);
+ x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
break;
case OP_FCONV_TO_I1:
code = emit_float_to_int (cfg, code, ins->dreg, 1, TRUE);
x86_push_reg (code, ins->sreg2);
x86_push_reg (code, ins->sreg1);
x86_fild_membase (code, X86_ESP, 0, TRUE);
+ /* Change precision */
+ x86_fst_membase (code, X86_ESP, 0, TRUE, TRUE);
+ x86_fld_membase (code, X86_ESP, 0, TRUE);
x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
break;
case OP_LCONV_TO_R4_2:
x86_fld_membase (code, X86_ESP, 0, FALSE);
x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
break;
- case OP_LCONV_TO_R_UN:
case OP_LCONV_TO_R_UN_2: {
static guint8 mn[] = { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x40 };
guint8 *br;
/* load 64bit integer to FP stack */
- x86_push_imm (code, 0);
x86_push_reg (code, ins->sreg2);
x86_push_reg (code, ins->sreg1);
x86_fild_membase (code, X86_ESP, 0, TRUE);
- /* store as 80bit FP value */
- x86_fst80_membase (code, X86_ESP, 0);
/* test if lreg is negative */
x86_test_reg_reg (code, ins->sreg2, ins->sreg2);
/* add correction constant mn */
x86_fld80_mem (code, mn);
- x86_fld80_membase (code, X86_ESP, 0);
x86_fp_op_reg (code, X86_FADD, 1, TRUE);
- x86_fst80_membase (code, X86_ESP, 0);
x86_patch (br, code);
- x86_fld80_membase (code, X86_ESP, 0);
- x86_alu_reg_imm (code, X86_ADD, X86_ESP, 12);
+ /* Change precision */
+ x86_fst_membase (code, X86_ESP, 0, TRUE, TRUE);
+ x86_fld_membase (code, X86_ESP, 0, TRUE);
+
+ x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
break;
}
break;
}
- case OP_ATOMIC_EXCHANGE_I4:
- case OP_ATOMIC_CAS_IMM_I4: {
+ case OP_ATOMIC_EXCHANGE_I4: {
guchar *br[2];
int sreg2 = ins->sreg2;
int breg = ins->inst_basereg;
/* We need the EAX reg for the cmpxchg */
if (ins->sreg2 == X86_EAX) {
- x86_push_reg (code, X86_EDX);
- x86_mov_reg_reg (code, X86_EDX, X86_EAX, 4);
- sreg2 = X86_EDX;
+ sreg2 = (breg == X86_EDX) ? X86_EBX : X86_EDX;
+ x86_push_reg (code, sreg2);
+ x86_mov_reg_reg (code, sreg2, X86_EAX, 4);
}
if (breg == X86_EAX) {
- x86_push_reg (code, X86_ESI);
- x86_mov_reg_reg (code, X86_ESI, X86_EAX, 4);
- breg = X86_ESI;
+ breg = (sreg2 == X86_ESI) ? X86_EDI : X86_ESI;
+ x86_push_reg (code, breg);
+ x86_mov_reg_reg (code, breg, X86_EAX, 4);
}
- if (ins->opcode == OP_ATOMIC_CAS_IMM_I4) {
- x86_mov_reg_imm (code, X86_EAX, ins->backend.data);
-
- x86_prefix (code, X86_LOCK_PREFIX);
- x86_cmpxchg_membase_reg (code, breg, ins->inst_offset, sreg2);
- } else {
- x86_mov_reg_membase (code, X86_EAX, breg, ins->inst_offset, 4);
+ x86_mov_reg_membase (code, X86_EAX, breg, ins->inst_offset, 4);
- br [0] = code; x86_prefix (code, X86_LOCK_PREFIX);
- x86_cmpxchg_membase_reg (code, breg, ins->inst_offset, sreg2);
- br [1] = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
- x86_patch (br [1], br [0]);
- }
+ br [0] = code; x86_prefix (code, X86_LOCK_PREFIX);
+ x86_cmpxchg_membase_reg (code, breg, ins->inst_offset, sreg2);
+ br [1] = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
+ x86_patch (br [1], br [0]);
if (breg != ins->inst_basereg)
- x86_pop_reg (code, X86_ESI);
+ x86_pop_reg (code, breg);
if (ins->sreg2 != sreg2)
- x86_pop_reg (code, X86_EDX);
+ x86_pop_reg (code, sreg2);
break;
}
-#ifdef MONO_ARCH_SIMD_INTRINSICS
- case OP_ADDPS:
- x86_sse_alu_ps_reg_reg (code, X86_SSE_ADD, ins->sreg1, ins->sreg2);
+ case OP_ATOMIC_CAS_I4: {
+ g_assert (ins->dreg == X86_EAX);
+ g_assert (ins->sreg3 == X86_EAX);
+ g_assert (ins->sreg1 != X86_EAX);
+ g_assert (ins->sreg1 != ins->sreg2);
+
+ x86_prefix (code, X86_LOCK_PREFIX);
+ x86_cmpxchg_membase_reg (code, ins->sreg1, ins->inst_offset, ins->sreg2);
+ break;
+ }
+#ifdef HAVE_SGEN_GC
+ case OP_CARD_TABLE_WBARRIER: {
+ int ptr = ins->sreg1;
+ int value = ins->sreg2;
+ guchar *br;
+ int nursery_shift, card_table_shift;
+ gpointer card_table_mask;
+ size_t nursery_size;
+ gulong card_table = (gulong)mono_gc_get_card_table (&card_table_shift, &card_table_mask);
+ gulong nursery_start = (gulong)mono_gc_get_nursery (&nursery_shift, &nursery_size);
+
+ /*
+ * We need one register we can clobber, we choose EDX and make sreg1
+ * fixed EAX to work around limitations in the local register allocator.
+ * sreg2 might get allocated to EDX, but that is not a problem since
+ * we use it before clobbering EDX.
+ */
+ g_assert (ins->sreg1 == X86_EAX);
+
+ /*
+ * This is the code we produce:
+ *
+ * edx = value
+ * edx >>= nursery_shift
+ * cmp edx, (nursery_start >> nursery_shift)
+ * jne done
+ * edx = ptr
+ * edx >>= card_table_shift
+ * card_table[edx] = 1
+ * done:
+ */
+
+ if (value != X86_EDX)
+ x86_mov_reg_reg (code, X86_EDX, value, 4);
+ x86_shift_reg_imm (code, X86_SHR, X86_EDX, nursery_shift);
+ x86_alu_reg_imm (code, X86_CMP, X86_EDX, nursery_start >> nursery_shift);
+ br = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
+ x86_mov_reg_reg (code, X86_EDX, ptr, 4);
+ x86_shift_reg_imm (code, X86_SHR, X86_EDX, card_table_shift);
+ x86_mov_membase_imm (code, X86_EDX, card_table, 1, 1);
+ x86_patch (br, code);
+ break;
+ }
+#endif
+#ifdef MONO_ARCH_SIMD_INTRINSICS
+ case OP_ADDPS:
+ x86_sse_alu_ps_reg_reg (code, X86_SSE_ADD, ins->sreg1, ins->sreg2);
break;
case OP_DIVPS:
x86_sse_alu_ps_reg_reg (code, X86_SSE_DIV, ins->sreg1, ins->sreg2);
case OP_XORPD:
x86_sse_alu_pd_reg_reg (code, X86_SSE_XOR, ins->sreg1, ins->sreg2);
break;
+ case OP_SQRTPD:
+ x86_sse_alu_pd_reg_reg (code, X86_SSE_SQRT, ins->dreg, ins->sreg1);
+ break;
case OP_ADDSUBPD:
x86_sse_alu_pd_reg_reg (code, X86_SSE_ADDSUB, ins->sreg1, ins->sreg2);
break;
}
if (G_UNLIKELY ((code - cfg->native_code - offset) > max_len)) {
+#ifndef __native_client_codegen__
g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %d)",
- mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
+ mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
g_assert_not_reached ();
+#endif /* __native_client_codegen__ */
}
cpos += max_len;
int alloc_size, pos, max_offset, i, cfa_offset;
guint8 *code;
gboolean need_stack_frame;
+#ifdef __native_client_codegen__
+ guint alignment_check;
+#endif
- cfg->code_size = MAX (mono_method_get_header (method)->code_size * 4, 10240);
+ cfg->code_size = MAX (cfg->header->code_size * 4, 10240);
if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
cfg->code_size += 512;
+#ifdef __native_client_codegen__
+ /* native_code_alloc is not 32-byte aligned, native_code is. */
+ cfg->native_code_alloc = g_malloc (cfg->code_size + kNaClAlignment);
+
+ /* Align native_code to next nearest kNaclAlignment byte. */
+ cfg->native_code = (guint)cfg->native_code_alloc + kNaClAlignment;
+ cfg->native_code = (guint)cfg->native_code & ~kNaClAlignmentMask;
+
+ code = cfg->native_code;
+
+ alignment_check = (guint)cfg->native_code & kNaClAlignmentMask;
+ g_assert(alignment_check == 0);
+#else
code = cfg->native_code = g_malloc (cfg->code_size);
+#endif
/* Offset between RSP and the CFA */
cfa_offset = 0;
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
x86_patch (buf, code);
-#ifdef PLATFORM_WIN32
+#ifdef TARGET_WIN32
/* The TLS key actually contains a pointer to the MonoJitTlsData structure */
/* FIXME: Add a separate key for LMF to avoid this */
x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
#endif
}
else {
- g_assert (!cfg->compile_aot);
- x86_push_imm (code, cfg->domain);
+ if (cfg->compile_aot) {
+ /*
+ * This goes before the saving of callee saved regs, so save the got reg
+ * ourselves.
+ */
+ x86_push_reg (code, MONO_ARCH_GOT_REG);
+ code = mono_arch_emit_load_got_addr (cfg->native_code, code, cfg, NULL);
+ x86_push_imm (code, 0);
+ } else {
+ x86_push_imm (code, cfg->domain);
+ }
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
+ if (cfg->compile_aot)
+ x86_pop_reg (code, MONO_ARCH_GOT_REG);
}
}
pos += sizeof (MonoLMF);
/* save the current IP */
- mono_add_patch_info (cfg, code + 1 - cfg->native_code, MONO_PATCH_INFO_IP, NULL);
- x86_push_imm_template (code);
+ if (cfg->compile_aot) {
+ /* This pushes the current ip */
+ x86_call_imm (code, 0);
+ } else {
+ mono_add_patch_info (cfg, code + 1 - cfg->native_code, MONO_PATCH_INFO_IP, NULL);
+ x86_push_imm_template (code);
+ }
cfa_offset += sizeof (gpointer);
/* save all caller saved regs */
if (lmf_addr_tls_offset != -1) {
/* Load lmf quicky using the GS register */
code = mono_x86_emit_tls_get (code, X86_EAX, lmf_addr_tls_offset);
-#ifdef PLATFORM_WIN32
+#ifdef TARGET_WIN32
/* The TLS key actually contains a pointer to the MonoJitTlsData structure */
/* FIXME: Add a separate key for LMF to avoid this */
x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
#endif
} else {
+ if (cfg->compile_aot)
+ code = mono_arch_emit_load_got_addr (cfg->native_code, code, cfg, NULL);
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_get_lmf_addr");
}
if (alloc_size) {
/* See mono_emit_stack_alloc */
-#if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
+#if defined(TARGET_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
guint32 remaining_size = alloc_size;
+ /*FIXME handle unbounded code expansion, we should use a loop in case of more than X interactions*/
+ guint32 required_code_size = ((remaining_size / 0x1000) + 1) * 8; /*8 is the max size of x86_alu_reg_imm + x86_test_membase_reg*/
+ guint32 offset = code - cfg->native_code;
+ if (G_UNLIKELY (required_code_size >= (cfg->code_size - offset))) {
+ while (required_code_size >= (cfg->code_size - offset))
+ cfg->code_size *= 2;
+ cfg->native_code = mono_realloc_native_code(cfg);
+ code = cfg->native_code + offset;
+ mono_jit_stats.code_reallocs++;
+ }
while (remaining_size >= 0x1000) {
x86_alu_reg_imm (code, X86_SUB, X86_ESP, 0x1000);
x86_test_membase_reg (code, X86_ESP, 0, X86_ESP);
/* max alignment for loops */
if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
max_offset += LOOP_ALIGNMENT;
-
+#ifdef __native_client_codegen__
+ /* max alignment for native client */
+ max_offset += kNaClAlignment;
+#endif
MONO_BB_FOR_EACH_INS (bb, ins) {
if (ins->opcode == OP_LABEL)
ins->inst_c1 = max_offset;
-
+#ifdef __native_client_codegen__
+ {
+ int space_in_block = kNaClAlignment -
+ ((max_offset + cfg->code_len) & kNaClAlignmentMask);
+ int max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
+ if (space_in_block < max_len && max_len < kNaClAlignment) {
+ max_offset += space_in_block;
+ }
+ }
+#endif /* __native_client_codegen__ */
max_offset += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
}
}
while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
cfg->code_size *= 2;
- cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
+ cfg->native_code = mono_realloc_native_code(cfg);
mono_jit_stats.code_reallocs++;
}
while (cfg->code_len + code_size > (cfg->code_size - 16)) {
cfg->code_size *= 2;
- cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
+ cfg->native_code = mono_realloc_native_code(cfg);
mono_jit_stats.code_reallocs++;
}
guint32 size;
/* Compute size of code following the push <OFFSET> */
+#ifdef __native_client_codegen__
+ code = mono_nacl_align (code);
+ size = kNaClAlignment;
+#else
size = 5 + 5;
+#endif
+ /*This is aligned to 16 bytes by the callee. This way we save a few bytes here.*/
if ((code - cfg->native_code) - throw_ip < 126 - size) {
/* Use the shorter form */
{
if (!tls_offset_inited) {
if (!getenv ("MONO_NO_TLS")) {
-#ifdef PLATFORM_WIN32
+#ifdef TARGET_WIN32
/*
* We need to init this multiple times, since when we are first called, the key might not
* be initialized yet.
*/
appdomain_tls_offset = mono_domain_get_tls_key ();
lmf_tls_offset = mono_get_jit_tls_key ();
- thread_tls_offset = mono_thread_get_tls_key ();
/* Only 64 tls entries can be accessed using inline code */
if (appdomain_tls_offset >= 64)
appdomain_tls_offset = -1;
if (lmf_tls_offset >= 64)
lmf_tls_offset = -1;
- if (thread_tls_offset >= 64)
- thread_tls_offset = -1;
#else
#if MONO_XEN_OPT
optimize_for_xen = access ("/proc/xen", F_OK) == 0;
appdomain_tls_offset = mono_domain_get_tls_offset ();
lmf_tls_offset = mono_get_lmf_tls_offset ();
lmf_addr_tls_offset = mono_get_lmf_addr_tls_offset ();
- thread_tls_offset = mono_thread_get_tls_offset ();
#endif
}
}
//[1 + 5] x86_jump_mem(inst,mem)
#define CMP_SIZE 6
+#ifdef __native_client_codegen__
+/* These constants should be coming from cpu-x86.md */
+/* I suspect the size calculation below is actually incorrect. */
+/* TODO: fix the calculation that uses these sizes. */
+#define BR_SMALL_SIZE 16
+#define BR_LARGE_SIZE 12
+#else
#define BR_SMALL_SIZE 2
#define BR_LARGE_SIZE 5
+#endif /* __native_client_codegen__ */
#define JUMP_IMM_SIZE 6
#define ENABLE_WRONG_METHOD_CHECK 0
+#define DEBUG_IMT 0
static int
imt_branch_distance (MonoIMTCheckItem **imt_entries, int start, int target)
int size = 0;
guint8 *code, *start;
+#ifdef __native_client_codegen__
+ /* g_print("mono_arch_build_imt_thunk needs to be aligned.\n"); */
+#endif
for (i = 0; i < count; ++i) {
MonoIMTCheckItem *item = imt_entries [i];
if (item->is_equals) {
if (fail_tramp)
code = mono_method_alloc_generic_virtual_thunk (domain, size);
else
- code = mono_code_manager_reserve (domain->code_mp, size);
+ code = mono_domain_code_reserve (domain, size);
start = code;
for (i = 0; i < count; ++i) {
MonoIMTCheckItem *item = imt_entries [i];
x86_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)item->key);
item->jmp_code = code;
x86_branch8 (code, X86_CC_NE, 0, FALSE);
- if (fail_tramp)
+ if (item->has_target_code)
x86_jump_code (code, item->value.target_code);
else
x86_jump_mem (code, & (vtable->vtable [item->value.vtable_slot]));
x86_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)item->key);
item->jmp_code = code;
x86_branch8 (code, X86_CC_NE, 0, FALSE);
- x86_jump_code (code, item->value.target_code);
+ if (item->has_target_code)
+ x86_jump_code (code, item->value.target_code);
+ else
+ x86_jump_mem (code, & (vtable->vtable [item->value.vtable_slot]));
x86_patch (item->jmp_code, code);
x86_jump_code (code, fail_tramp);
item->jmp_code = NULL;
item->jmp_code = code;
x86_branch8 (code, X86_CC_NE, 0, FALSE);
#endif
- x86_jump_mem (code, & (vtable->vtable [item->value.vtable_slot]));
+ if (item->has_target_code)
+ x86_jump_code (code, item->value.target_code);
+ else
+ x86_jump_mem (code, & (vtable->vtable [item->value.vtable_slot]));
#if ENABLE_WRONG_METHOD_CHECK
x86_patch (item->jmp_code, code);
x86_breakpoint (code);
if (!fail_tramp)
mono_stats.imt_thunks_size += code - start;
g_assert (code - start <= size);
+
+#if DEBUG_IMT
+ {
+ char *buff = g_strdup_printf ("thunk_for_class_%s_%s_entries_%d", vtable->klass->name_space, vtable->klass->name, count);
+ mono_disassemble_code (NULL, (guint8*)start, code - start, buff);
+ g_free (buff);
+ }
+#endif
+
return start;
}
MonoMethod*
-mono_arch_find_imt_method (gpointer *regs, guint8 *code)
+mono_arch_find_imt_method (mgreg_t *regs, guint8 *code)
{
return (MonoMethod*) regs [MONO_ARCH_IMT_REG];
}
-
-MonoObject*
-mono_arch_find_this_argument (gpointer *regs, MonoMethod *method, MonoGenericSharingContext *gsctx)
-{
- MonoMethodSignature *sig = mono_method_signature (method);
- CallInfo *cinfo = get_call_info (gsctx, NULL, sig, FALSE);
- int this_argument_offset;
- MonoObject *this_argument;
-
- /*
- * this is the offset of the this arg from esp as saved at the start of
- * mono_arch_create_trampoline_code () in tramp-x86.c.
- */
- this_argument_offset = 5;
- if (MONO_TYPE_ISSTRUCT (sig->ret) && (cinfo->ret.storage == ArgOnStack))
- this_argument_offset++;
-
- this_argument = * (MonoObject**) (((guint8*) regs [X86_ESP]) + this_argument_offset * sizeof (gpointer));
-
- g_free (cinfo);
- return this_argument;
-}
#endif
MonoVTable*
-mono_arch_find_static_call_vtable (gpointer *regs, guint8 *code)
+mono_arch_find_static_call_vtable (mgreg_t *regs, guint8 *code)
{
return (MonoVTable*) regs [MONO_ARCH_RGCTX_REG];
}
+GSList*
+mono_arch_get_cie_program (void)
+{
+ GSList *l = NULL;
+
+ mono_add_unwind_op_def_cfa (l, (guint8*)NULL, (guint8*)NULL, X86_ESP, 4);
+ mono_add_unwind_op_offset (l, (guint8*)NULL, (guint8*)NULL, X86_NREG, -4);
+
+ return l;
+}
+
MonoInst*
mono_arch_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
{
return ins;
}
-MonoInst* mono_arch_get_thread_intrinsic (MonoCompile* cfg)
-{
- MonoInst* ins;
-
- if (thread_tls_offset == -1)
- return NULL;
-
- MONO_INST_NEW (cfg, ins, OP_TLS_GET);
- ins->inst_offset = thread_tls_offset;
- return ins;
-}
-
guint32
mono_arch_get_patch_offset (guint8 *code)
{
return can_write;
}
-gpointer
-mono_arch_get_vcall_slot (guint8 *code, gpointer *regs, int *displacement)
-{
- guint8 buf [8];
- guint8 reg = 0;
- gint32 disp = 0;
-
- mono_breakpoint_clean_code (NULL, code, 8, buf, sizeof (buf));
- code = buf + 8;
-
- *displacement = 0;
-
- /* go to the start of the call instruction
- *
- * address_byte = (m << 6) | (o << 3) | reg
- * call opcode: 0xff address_byte displacement
- * 0xff m=1,o=2 imm8
- * 0xff m=2,o=2 imm32
- */
- code -= 6;
-
- /*
- * A given byte sequence can match more than case here, so we have to be
- * really careful about the ordering of the cases. Longer sequences
- * come first.
- * Some of the rules are only needed because the imm in the mov could
- * match the
- * code [2] == 0xe8 case below.
- */
- if ((code [-2] == 0x8b) && (x86_modrm_mod (code [-1]) == 0x2) && (code [4] == 0xff) && (x86_modrm_reg (code [5]) == 0x2) && (x86_modrm_mod (code [5]) == 0x0)) {
- /*
- * This is an interface call
- * 8b 80 0c e8 ff ff mov 0xffffe80c(%eax),%eax
- * ff 10 call *(%eax)
- */
- reg = x86_modrm_rm (code [5]);
- disp = 0;
-#ifdef MONO_ARCH_HAVE_IMT
- } else if ((code [-2] == 0xba) && (code [3] == 0xff) && (x86_modrm_mod (code [4]) == 1) && (x86_modrm_reg (code [4]) == 2) && ((signed char)code [5] < 0)) {
- /* IMT-based interface calls: with MONO_ARCH_IMT_REG == edx
- * ba 14 f8 28 08 mov $0x828f814,%edx
- * ff 50 fc call *0xfffffffc(%eax)
- */
- reg = code [4] & 0x07;
- disp = (signed char)code [5];
-#endif
- } else if ((code [-2] >= 0xb8) && (code [-2] < 0xb8 + 8) && (code [3] == 0xff) && (x86_modrm_reg (code [4]) == 0x2) && (x86_modrm_mod (code [4]) == 0x1)) {
- /*
- * ba e8 e8 e8 e8 mov $0xe8e8e8e8,%edx
- * ff 50 60 callq *0x60(%eax)
- */
- reg = x86_modrm_rm (code [4]);
- disp = *(gint8*)(code + 5);
- } else if ((code [1] != 0xe8) && (code [3] == 0xff) && ((code [4] & 0x18) == 0x10) && ((code [4] >> 6) == 1)) {
- reg = code [4] & 0x07;
- disp = (signed char)code [5];
- } else {
- if ((code [0] == 0xff) && ((code [1] & 0x18) == 0x10) && ((code [1] >> 6) == 2)) {
- reg = code [1] & 0x07;
- disp = *((gint32*)(code + 2));
- } else if ((code [1] == 0xe8)) {
- return NULL;
- } else if ((code [4] == 0xff) && (((code [5] >> 6) & 0x3) == 0) && (((code [5] >> 3) & 0x7) == 2)) {
- /*
- * This is a interface call
- * 8b 40 30 mov 0x30(%eax),%eax
- * ff 10 call *(%eax)
- */
- disp = 0;
- reg = code [5] & 0x07;
- }
- else
- return NULL;
- }
-
- *displacement = disp;
- return regs [reg];
-}
-
-gpointer*
-mono_arch_get_vcall_slot_addr (guint8 *code, gpointer *regs)
+/*
+ * mono_x86_get_this_arg_offset:
+ *
+ * Return the offset of the stack location where this is passed during a virtual
+ * call.
+ */
+guint32
+mono_x86_get_this_arg_offset (MonoGenericSharingContext *gsctx, MonoMethodSignature *sig)
{
- gpointer vt;
- int displacement;
- vt = mono_arch_get_vcall_slot (code, regs, &displacement);
- if (!vt)
- return NULL;
- return (gpointer*)((char*)vt + displacement);
+ return 0;
}
gpointer
mono_arch_get_this_arg_from_call (MonoGenericSharingContext *gsctx, MonoMethodSignature *sig,
- gssize *regs, guint8 *code)
+ mgreg_t *regs, guint8 *code)
{
guint32 esp = regs [X86_ESP];
- CallInfo *cinfo;
+ CallInfo *cinfo = NULL;
gpointer res;
+ int offset;
- if (!gsctx && code)
- gsctx = mono_get_generic_context_from_code (code);
- cinfo = get_call_info (gsctx, NULL, sig, FALSE);
+ offset = 0;
/*
* The stack looks like:
* <other args>
* <this=delegate>
- * <possible vtype return address>
* <return addr>
* <4 pointers pushed by mono_arch_create_trampoline_code ()>
*/
- res = (((MonoObject**)esp) [5 + (cinfo->args [0].offset / 4)]);
- g_free (cinfo);
+ res = (((MonoObject**)esp) [5 + (offset / 4)]);
+ if (cinfo)
+ g_free (cinfo);
return res;
}
#define MAX_ARCH_DELEGATE_PARAMS 10
-gpointer
-mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
+static gpointer
+get_delegate_invoke_impl (gboolean has_target, guint32 param_count, guint32 *code_len)
{
guint8 *code, *start;
- if (sig->param_count > MAX_ARCH_DELEGATE_PARAMS)
- return NULL;
-
- /* FIXME: Support more cases */
- if (MONO_TYPE_ISSTRUCT (sig->ret))
- return NULL;
-
/*
* The stack contains:
* <delegate>
*/
if (has_target) {
- static guint8* cached = NULL;
- if (cached)
- return cached;
-
start = code = mono_global_codeman_reserve (64);
/* Replace the this argument with the target */
x86_jump_membase (code, X86_EAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
g_assert ((code - start) < 64);
-
- mono_debug_add_delegate_trampoline (start, code - start);
-
- mono_memory_barrier ();
-
- cached = start;
} else {
- static guint8* cache [MAX_ARCH_DELEGATE_PARAMS + 1] = {NULL};
int i = 0;
/* 8 for mov_reg and jump, plus 8 for each parameter */
- int code_reserve = 8 + (sig->param_count * 8);
-
- for (i = 0; i < sig->param_count; ++i)
- if (!mono_is_regsize_var (sig->params [i]))
- return NULL;
-
- code = cache [sig->param_count];
- if (code)
- return code;
-
+#ifdef __native_client_codegen__
+ /* TODO: calculate this size correctly */
+ int code_reserve = 13 + (param_count * 8) + 2 * kNaClAlignment;
+#else
+ int code_reserve = 8 + (param_count * 8);
+#endif /* __native_client_codegen__ */
/*
* The stack contains:
* <args in reverse order>
x86_mov_reg_membase (code, X86_ECX, X86_ESP, 4, 4);
/* move args up */
- for (i = 0; i < sig->param_count; ++i) {
+ for (i = 0; i < param_count; ++i) {
x86_mov_reg_membase (code, X86_EAX, X86_ESP, (i+2)*4, 4);
x86_mov_membase_reg (code, X86_ESP, (i+1)*4, X86_EAX, 4);
}
x86_jump_membase (code, X86_ECX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
g_assert ((code - start) < code_reserve);
+ }
+
+ mono_debug_add_delegate_trampoline (start, code - start);
+
+ if (code_len)
+ *code_len = code - start;
+
+ return start;
+}
+
+GSList*
+mono_arch_get_delegate_invoke_impls (void)
+{
+ GSList *res = NULL;
+ guint8 *code;
+ guint32 code_len;
+ int i;
+
+ code = get_delegate_invoke_impl (TRUE, 0, &code_len);
+ res = g_slist_prepend (res, mono_tramp_info_create (g_strdup ("delegate_invoke_impl_has_target"), code, code_len, NULL, NULL));
+
+ for (i = 0; i < MAX_ARCH_DELEGATE_PARAMS; ++i) {
+ code = get_delegate_invoke_impl (FALSE, i, &code_len);
+ res = g_slist_prepend (res, mono_tramp_info_create (g_strdup_printf ("delegate_invoke_impl_target_%d", i), code, code_len, NULL, NULL));
+ }
+
+ return res;
+}
+
+gpointer
+mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
+{
+ guint8 *code, *start;
+
+ if (sig->param_count > MAX_ARCH_DELEGATE_PARAMS)
+ return NULL;
+
+ /* FIXME: Support more cases */
+ if (MONO_TYPE_ISSTRUCT (sig->ret))
+ return NULL;
+
+ /*
+ * The stack contains:
+ * <delegate>
+ * <return addr>
+ */
+
+ if (has_target) {
+ static guint8* cached = NULL;
+ if (cached)
+ return cached;
+
+ if (mono_aot_only)
+ start = mono_aot_get_trampoline ("delegate_invoke_impl_has_target");
+ else
+ start = get_delegate_invoke_impl (TRUE, 0, NULL);
+
+ mono_memory_barrier ();
+
+ cached = start;
+ } else {
+ static guint8* cache [MAX_ARCH_DELEGATE_PARAMS + 1] = {NULL};
+ int i = 0;
+
+ for (i = 0; i < sig->param_count; ++i)
+ if (!mono_is_regsize_var (sig->params [i]))
+ return NULL;
+
+ code = cache [sig->param_count];
+ if (code)
+ return code;
- mono_debug_add_delegate_trampoline (start, code - start);
+ if (mono_aot_only) {
+ char *name = g_strdup_printf ("delegate_invoke_impl_target_%d", sig->param_count);
+ start = mono_aot_get_trampoline (name);
+ g_free (name);
+ } else {
+ start = get_delegate_invoke_impl (FALSE, sig->param_count, NULL);
+ }
mono_memory_barrier ();
MonoInst *fconv;
int dreg, src_opcode;
- if (!(cfg->opt & MONO_OPT_SSE2) || !(cfg->opt & MONO_OPT_SIMD))
+ if (!(cfg->opt & MONO_OPT_SSE2) || !(cfg->opt & MONO_OPT_SIMD) || COMPILE_LLVM (cfg))
return;
switch (src_opcode = ins->opcode) {
ins->backend.source_opcode = src_opcode;
}
+#endif /* #ifdef MONO_ARCH_SIMD_INTRINSICS */
+
void
mono_arch_decompose_long_opts (MonoCompile *cfg, MonoInst *long_ins)
{
MonoInst *ins;
int vreg;
+
+ if (long_ins->opcode == OP_LNEG) {
+ ins = long_ins;
+ MONO_EMIT_NEW_UNALU (cfg, OP_INEG, ins->dreg + 1, ins->sreg1 + 1);
+ MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ADC_IMM, ins->dreg + 2, ins->sreg1 + 2, 0);
+ MONO_EMIT_NEW_UNALU (cfg, OP_INEG, ins->dreg + 2, ins->dreg + 2);
+ NULLIFY_INS (ins);
+ return;
+ }
+
+#ifdef MONO_ARCH_SIMD_INTRINSICS
+
if (!(cfg->opt & MONO_OPT_SIMD))
return;
long_ins->opcode = OP_NOP;
break;
}
+#endif /* MONO_ARCH_SIMD_INTRINSICS */
}
+
+/*MONO_ARCH_HAVE_HANDLER_BLOCK_GUARD*/
+gpointer
+mono_arch_install_handler_block_guard (MonoJitInfo *ji, MonoJitExceptionInfo *clause, MonoContext *ctx, gpointer new_value)
+{
+ int offset;
+ gpointer *sp, old_value;
+ char *bp;
+ const unsigned char *handler;
+
+ /*Decode the first instruction to figure out where did we store the spvar*/
+ /*Our jit MUST generate the following:
+ mov %esp, -?(%ebp)
+ Which is encoded as: 0x89 mod_rm.
+ mod_rm (esp, ebp, imm) which can be: (imm will never be zero)
+ mod (reg + imm8): 01 reg(esp): 100 rm(ebp): 101 -> 01100101 (0x65)
+ mod (reg + imm32): 10 reg(esp): 100 rm(ebp): 101 -> 10100101 (0xA5)
+ */
+ handler = clause->handler_start;
+
+ if (*handler != 0x89)
+ return NULL;
+
+ ++handler;
+
+ if (*handler == 0x65)
+ offset = *(signed char*)(handler + 1);
+ else if (*handler == 0xA5)
+ offset = *(int*)(handler + 1);
+ else
+ return NULL;
+
+ /*Load the spvar*/
+ bp = MONO_CONTEXT_GET_BP (ctx);
+ sp = *(gpointer*)(bp + offset);
+
+ old_value = *sp;
+ if (old_value < ji->code_start || (char*)old_value > ((char*)ji->code_start + ji->code_size))
+ return old_value;
+
+ *sp = new_value;
+
+ return old_value;
+}
+
+/*
+ * mono_aot_emit_load_got_addr:
+ *
+ * Emit code to load the got address.
+ * On x86, the result is placed into EBX.
+ */
+guint8*
+mono_arch_emit_load_got_addr (guint8 *start, guint8 *code, MonoCompile *cfg, MonoJumpInfo **ji)
+{
+ x86_call_imm (code, 0);
+ /*
+ * The patch needs to point to the pop, since the GOT offset needs
+ * to be added to that address.
+ */
+ if (cfg)
+ mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GOT_OFFSET, NULL);
+ else
+ *ji = mono_patch_info_list_prepend (*ji, code - start, MONO_PATCH_INFO_GOT_OFFSET, NULL);
+ x86_pop_reg (code, MONO_ARCH_GOT_REG);
+ x86_alu_reg_imm (code, X86_ADD, MONO_ARCH_GOT_REG, 0xf0f0f0f0);
+
+ return code;
+}
+
+/*
+ * mono_ppc_emit_load_aotconst:
+ *
+ * Emit code to load the contents of the GOT slot identified by TRAMP_TYPE and
+ * TARGET from the mscorlib GOT in full-aot code.
+ * On x86, the GOT address is assumed to be in EBX, and the result is placed into
+ * EAX.
+ */
+guint8*
+mono_arch_emit_load_aotconst (guint8 *start, guint8 *code, MonoJumpInfo **ji, int tramp_type, gconstpointer target)
+{
+ /* Load the mscorlib got address */
+ x86_mov_reg_membase (code, X86_EAX, MONO_ARCH_GOT_REG, sizeof (gpointer), 4);
+ *ji = mono_patch_info_list_prepend (*ji, code - start, tramp_type, target);
+ /* arch_emit_got_access () patches this */
+ x86_mov_reg_membase (code, X86_EAX, X86_EAX, 0xf0f0f0f0, 4);
+
+ return code;
+}
+
+/* Can't put this into mini-x86.h */
+gpointer
+mono_x86_get_signal_exception_trampoline (MonoTrampInfo **info, gboolean aot);
+
+GSList *
+mono_arch_get_trampolines (gboolean aot)
+{
+ MonoTrampInfo *info;
+ GSList *tramps = NULL;
+
+ mono_x86_get_signal_exception_trampoline (&info, aot);
+
+ tramps = g_slist_append (tramps, info);
+
+ return tramps;
+}
+
+
+#if __APPLE__
+#define DBG_SIGNAL SIGBUS
+#else
+#define DBG_SIGNAL SIGSEGV
+#endif
+
+/* Soft Debug support */
+#ifdef MONO_ARCH_SOFT_DEBUG_SUPPORTED
+
+/*
+ * mono_arch_set_breakpoint:
+ *
+ * Set a breakpoint at the native code corresponding to JI at NATIVE_OFFSET.
+ * The location should contain code emitted by OP_SEQ_POINT.
+ */
+void
+mono_arch_set_breakpoint (MonoJitInfo *ji, guint8 *ip)
+{
+ guint8 *code = ip;
+
+ /*
+ * In production, we will use int3 (has to fix the size in the md
+ * file). But that could confuse gdb, so during development, we emit a SIGSEGV
+ * instead.
+ */
+ g_assert (code [0] == 0x90);
+ x86_alu_reg_mem (code, X86_CMP, X86_EAX, (guint32)bp_trigger_page);
+}
+
+/*
+ * mono_arch_clear_breakpoint:
+ *
+ * Clear the breakpoint at IP.
+ */
+void
+mono_arch_clear_breakpoint (MonoJitInfo *ji, guint8 *ip)
+{
+ guint8 *code = ip;
+ int i;
+
+ for (i = 0; i < 6; ++i)
+ x86_nop (code);
+}
+
+/*
+ * mono_arch_start_single_stepping:
+ *
+ * Start single stepping.
+ */
+void
+mono_arch_start_single_stepping (void)
+{
+ mono_mprotect (ss_trigger_page, mono_pagesize (), 0);
+}
+
+/*
+ * mono_arch_stop_single_stepping:
+ *
+ * Stop single stepping.
+ */
+void
+mono_arch_stop_single_stepping (void)
+{
+ mono_mprotect (ss_trigger_page, mono_pagesize (), MONO_MMAP_READ);
+}
+
+/*
+ * mono_arch_is_single_step_event:
+ *
+ * Return whenever the machine state in SIGCTX corresponds to a single
+ * step event.
+ */
+gboolean
+mono_arch_is_single_step_event (void *info, void *sigctx)
+{
+#ifdef TARGET_WIN32
+ EXCEPTION_RECORD* einfo = (EXCEPTION_RECORD*)info; /* Sometimes the address is off by 4 */
+ if ((einfo->ExceptionInformation[1] >= ss_trigger_page && (guint8*)einfo->ExceptionInformation[1] <= (guint8*)ss_trigger_page + 128))
+ return TRUE;
+ else
+ return FALSE;
+#else
+ siginfo_t* sinfo = (siginfo_t*) info;
+ /* Sometimes the address is off by 4 */
+ if (sinfo->si_signo == DBG_SIGNAL && (sinfo->si_addr >= ss_trigger_page && (guint8*)sinfo->si_addr <= (guint8*)ss_trigger_page + 128))
+ return TRUE;
+ else
+ return FALSE;
+#endif
+}
+
+gboolean
+mono_arch_is_breakpoint_event (void *info, void *sigctx)
+{
+#ifdef TARGET_WIN32
+ EXCEPTION_RECORD* einfo = (EXCEPTION_RECORD*)info; /* Sometimes the address is off by 4 */
+ if ((einfo->ExceptionInformation[1] >= bp_trigger_page && (guint8*)einfo->ExceptionInformation[1] <= (guint8*)bp_trigger_page + 128))
+ return TRUE;
+ else
+ return FALSE;
+#else
+ siginfo_t* sinfo = (siginfo_t*)info;
+ /* Sometimes the address is off by 4 */
+ if (sinfo->si_signo == DBG_SIGNAL && (sinfo->si_addr >= bp_trigger_page && (guint8*)sinfo->si_addr <= (guint8*)bp_trigger_page + 128))
+ return TRUE;
+ else
+ return FALSE;
+#endif
+}
+
+/*
+ * mono_arch_get_ip_for_breakpoint:
+ *
+ * See mini-amd64.c for docs.
+ */
+guint8*
+mono_arch_get_ip_for_breakpoint (MonoJitInfo *ji, MonoContext *ctx)
+{
+ guint8 *ip = MONO_CONTEXT_GET_IP (ctx);
+
+ return ip;
+}
+
+#define BREAKPOINT_SIZE 6
+
+/*
+ * mono_arch_get_ip_for_single_step:
+ *
+ * See mini-amd64.c for docs.
+ */
+guint8*
+mono_arch_get_ip_for_single_step (MonoJitInfo *ji, MonoContext *ctx)
+{
+ guint8 *ip = MONO_CONTEXT_GET_IP (ctx);
+
+ /* Size of x86_alu_reg_imm */
+ ip += 6;
+
+ return ip;
+}
+
+/*
+ * mono_arch_skip_breakpoint:
+ *
+ * See mini-amd64.c for docs.
+ */
+void
+mono_arch_skip_breakpoint (MonoContext *ctx)
+{
+ MONO_CONTEXT_SET_IP (ctx, (guint8*)MONO_CONTEXT_GET_IP (ctx) + BREAKPOINT_SIZE);
+}
+
+/*
+ * mono_arch_skip_single_step:
+ *
+ * See mini-amd64.c for docs.
+ */
+void
+mono_arch_skip_single_step (MonoContext *ctx)
+{
+ MONO_CONTEXT_SET_IP (ctx, (guint8*)MONO_CONTEXT_GET_IP (ctx) + 6);
+}
+
+/*
+ * mono_arch_get_seq_point_info:
+ *
+ * See mini-amd64.c for docs.
+ */
+gpointer
+mono_arch_get_seq_point_info (MonoDomain *domain, guint8 *code)
+{
+ NOT_IMPLEMENTED;
+ return NULL;
+}
+
#endif