X-Git-Url: http://wien.tomnetworks.com/gitweb/?a=blobdiff_plain;f=mono%2Fmini%2Ftramp-amd64.c;h=634203bc2e0c92404a2551b3241ccf6c47afae64;hb=2015a4ab214a8ec73e6a132da76d9681aaa99e06;hp=8e054b99dadf150db4a8e4631565b94862e145f9;hpb=2dc628d36f9ebd44d25126bfba00cffc6c2df8e4;p=mono.git diff --git a/mono/mini/tramp-amd64.c b/mono/mini/tramp-amd64.c index 8e054b99dad..634203bc2e0 100644 --- a/mono/mini/tramp-amd64.c +++ b/mono/mini/tramp-amd64.c @@ -1,13 +1,16 @@ -/* - * tramp-amd64.c: JIT trampoline code for amd64 +/** + * \file + * JIT trampoline code for amd64 * * Authors: * Dietmar Maurer (dietmar@ximian.com) * Zoltan Varga (vargaz@gmail.com) + * Johan Lorensson (lateralusx.github@gmail.com) * * (C) 2001 Ximian, Inc. * Copyright 2003-2011 Novell, Inc (http://www.novell.com) * Copyright 2011 Xamarin, Inc (http://www.xamarin.com) + * Licensed under the MIT license. See LICENSE file in the project root for full license information. */ #include @@ -17,7 +20,6 @@ #include #include #include -#include #include #include #include @@ -28,15 +30,15 @@ #include "mini-amd64.h" #include "debugger-agent.h" -#if defined(__native_client_codegen__) && defined(__native_client__) -#include -#include +#ifdef ENABLE_INTERPRETER +#include "interp/interp.h" #endif #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1)) #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f)) +#ifndef DISABLE_JIT /* * mono_arch_get_unbox_trampoline: * @m: method pointer @@ -51,13 +53,13 @@ mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr) { guint8 *code, *start; GSList *unwind_ops; - int this_reg, size = NACL_SIZE (20, 32); + int this_reg, size = 20; MonoDomain *domain = mono_domain_get (); this_reg = mono_arch_get_this_arg_reg (NULL); - start = code = (guint8 *)mono_domain_code_reserve (domain, size); + start = code = (guint8 *)mono_domain_code_reserve (domain, size + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)); unwind_ops = mono_arch_get_cie_program (); @@ -66,8 +68,7 @@ mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr) amd64_mov_reg_imm (code, AMD64_RAX, addr); amd64_jump_reg (code, AMD64_RAX); g_assert ((code - start) < size); - - nacl_domain_code_validate (domain, &start, size, &code); + g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0))); mono_arch_flush_icache (start, code - start); mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_UNBOX_TRAMPOLINE, m); @@ -80,10 +81,10 @@ mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr) /* * mono_arch_get_static_rgctx_trampoline: * - * Create a trampoline which sets RGCTX_REG to MRGCTX, then jumps to ADDR. + * Create a trampoline which sets RGCTX_REG to ARG, then jumps to ADDR. */ gpointer -mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericContext *mrgctx, gpointer addr) +mono_arch_get_static_rgctx_trampoline (gpointer arg, gpointer addr) { guint8 *code, *start; GSList *unwind_ops; @@ -96,20 +97,20 @@ mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericCo #else /* AOTed code could still have a non-32 bit address */ if ((((guint64)addr) >> 32) == 0) - buf_len = NACL_SIZE (16, 32); + buf_len = 16; else - buf_len = NACL_SIZE (30, 32); + buf_len = 30; #endif - start = code = (guint8 *)mono_domain_code_reserve (domain, buf_len); + start = code = (guint8 *)mono_domain_code_reserve (domain, buf_len + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)); unwind_ops = mono_arch_get_cie_program (); - amd64_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, mrgctx); + amd64_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, arg); amd64_jump_code (code, addr); g_assert ((code - start) < buf_len); + g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0))); - nacl_domain_code_validate (domain, &start, buf_len, &code); mono_arch_flush_icache (start, code - start); mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL); @@ -117,6 +118,7 @@ mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericCo return start; } +#endif /* !DISABLE_JIT */ #ifdef _WIN64 // Workaround lack of Valgrind support for 64-bit Windows @@ -132,7 +134,6 @@ mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericCo void mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr) { -#if defined(__default_codegen__) guint8 *code; guint8 buf [16]; gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf)); @@ -177,40 +178,9 @@ mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr) VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer)); } } -#elif defined(__native_client__) - /* These are essentially the same 2 cases as above, modified for NaCl*/ - - /* Target must be bundle-aligned */ - g_assert (((guint32)addr & kNaClAlignmentMask) == 0); - /* Return target must be bundle-aligned */ - g_assert (((guint32)orig_code & kNaClAlignmentMask) == 0); - - if (orig_code[-5] == 0xe8) { - /* Direct call */ - int ret; - gint32 offset = (gint32)addr - (gint32)orig_code; - guint8 buf[sizeof(gint32)]; - *((gint32*)(buf)) = offset; - ret = nacl_dyncode_modify (orig_code - sizeof(gint32), buf, sizeof(gint32)); - g_assert (ret == 0); - } - - else if (is_nacl_call_reg_sequence (orig_code - 10) && orig_code[-16] == 0x41 && orig_code[-15] == 0xbb) { - int ret; - guint8 buf[sizeof(gint32)]; - *((gint32 *)(buf)) = addr; - /* orig_code[-14] is the start of the immediate. */ - ret = nacl_dyncode_modify (orig_code - 14, buf, sizeof(gint32)); - g_assert (ret == 0); - } - else { - g_assert_not_reached (); - } - - return; -#endif } +#ifndef DISABLE_JIT guint8* mono_arch_create_llvm_native_thunk (MonoDomain *domain, guint8 *addr) { @@ -230,6 +200,7 @@ mono_arch_create_llvm_native_thunk (MonoDomain *domain, guint8 *addr) mono_profiler_code_buffer_new (thunk_start, thunk_code - thunk_start, MONO_PROFILER_CODE_BUFFER_HELPER, NULL); return addr; } +#endif /* !DISABLE_JIT */ void mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr) @@ -237,7 +208,6 @@ mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *a gint32 disp; gpointer *plt_jump_table_entry; -#if defined(__default_codegen__) /* A PLT entry: jmp *(%rip) */ g_assert (code [0] == 0xff); g_assert (code [1] == 0x25); @@ -245,27 +215,11 @@ mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *a disp = *(gint32*)(code + 2); plt_jump_table_entry = (gpointer*)(code + 6 + disp); -#elif defined(__native_client_codegen__) - /* A PLT entry: */ - /* mov (%rip), %r11d */ - /* nacljmp *%r11 */ - - /* Verify the 'mov' */ - g_assert (code [0] == 0x45); - g_assert (code [1] == 0x8b); - g_assert (code [2] == 0x1d); - - disp = *(gint32*)(code + 3); - - /* 7 = 3 (mov opcode) + 4 (disp) */ - /* This needs to resolve to the target of the RIP-relative offset */ - plt_jump_table_entry = (gpointer*)(code + 7 + disp); - -#endif /* __native_client_codegen__ */ InterlockedExchangePointer (plt_jump_table_entry, addr); } +#ifndef DISABLE_JIT static void stack_unaligned (MonoTrampolineType tramp_type) { @@ -277,24 +231,20 @@ guchar* mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot) { char *tramp_name; - guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code; - int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, tramp_offset, ctx_offset, saved_regs_offset; + guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code, *br_ex_check; + int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, ex_offset, tramp_offset, ctx_offset, saved_regs_offset; int r11_save_offset, saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset; gboolean has_caller; GSList *unwind_ops = NULL; MonoJumpInfo *ji = NULL; - const guint kMaxCodeSize = NACL_SIZE (600, 600*2); - -#if defined(__native_client_codegen__) - const guint kNaClTrampOffset = 17; -#endif + const guint kMaxCodeSize = 630; if (tramp_type == MONO_TRAMPOLINE_JUMP || tramp_type == MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD) has_caller = FALSE; else has_caller = TRUE; - code = buf = (guint8 *)mono_global_codeman_reserve (kMaxCodeSize); + code = buf = (guint8 *)mono_global_codeman_reserve (kMaxCodeSize + MONO_MAX_TRAMPOLINE_UNWINDINFO_SIZE); /* Compute stack frame size and offsets */ offset = 0; @@ -303,6 +253,9 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf offset += sizeof(mgreg_t); rax_offset = -offset; + offset += sizeof(mgreg_t); + ex_offset = -offset; + offset += sizeof(mgreg_t); r11_save_offset = -offset; @@ -360,16 +313,13 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf orig_rsp_to_rbp_offset -= sizeof(mgreg_t); amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t)); mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP); + mono_add_unwind_op_fp_alloc (unwind_ops, code, buf, AMD64_RBP, 0); amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize); /* Compute the trampoline address from the return address */ if (aot) { -#if defined(__default_codegen__) /* 7 = length of call *(rip) */ amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7); -#elif defined(__native_client_codegen__) - amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, kNaClTrampOffset); -#endif } else { /* 5 = length of amd64_call_membase () */ amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5); @@ -406,12 +356,13 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf amd64_mov_reg_membase (code, i, AMD64_RSP, r11_save_offset + orig_rsp_to_rbp_offset, sizeof(mgreg_t)); amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t)); } + /* cfa = rbp + cfa_offset */ + mono_add_unwind_op_offset (unwind_ops, code, buf, i, - cfa_offset + saved_regs_offset + (i * sizeof (mgreg_t))); } for (i = 0; i < 8; ++i) amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)), i); /* Check that the stack is aligned */ -#if defined(__default_codegen__) amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof (mgreg_t)); amd64_alu_reg_imm (code, X86_AND, AMD64_R11, 15); amd64_alu_reg_imm (code, X86_CMP, AMD64_R11, 0); @@ -427,14 +378,12 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf } mono_amd64_patch (br [0], code); //amd64_breakpoint (code); -#endif if (tramp_type != MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD) { /* Obtain the trampoline argument which is encoded in the instruction stream */ if (aot) { /* Load the GOT offset */ amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer)); -#if defined(__default_codegen__) /* * r11 points to a call *(%rip) instruction, load the * pc-relative offset from the instruction itself. @@ -442,18 +391,12 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 3, 4); /* 7 is the length of the call, 8 is the offset to the next got slot */ amd64_alu_reg_imm_size (code, X86_ADD, AMD64_RAX, 7 + sizeof (gpointer), sizeof(gpointer)); -#elif defined(__native_client_codegen__) - /* The arg is hidden in a "push imm32" instruction, */ - /* add one to skip the opcode. */ - amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, kNaClTrampOffset+1, 4); -#endif /* Compute the address of the GOT slot */ amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, sizeof(gpointer)); /* Load the value */ amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, sizeof(gpointer)); } else { amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer)); -#if defined(__default_codegen__) amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1); amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE); amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1); @@ -467,10 +410,6 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf mono_amd64_patch (br [0], code); amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8); mono_amd64_patch (br [1], code); -#elif defined(__native_client_codegen__) - /* All args are 32-bit pointers in NaCl */ - amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4); -#endif } amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer)); } else { @@ -538,21 +477,7 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf amd64_mov_reg_imm (code, AMD64_R11, tramp); } amd64_call_reg (code, AMD64_R11); - - /* Check for thread interruption */ - /* This is not perf critical code so no need to check the interrupt flag */ - /* - * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack. - */ amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, sizeof(mgreg_t)); - if (aot) { - code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint"); - } else { - amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint); - } - amd64_call_reg (code, AMD64_R11); - - amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, sizeof(mgreg_t)); /* Restore LMF */ amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer)); @@ -564,8 +489,59 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf * Save rax to the stack, after the leave instruction, this will become part of * the red zone. */ + amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, sizeof(mgreg_t)); amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, sizeof(mgreg_t)); + /* Check for thread interruption */ + /* This is not perf critical code so no need to check the interrupt flag */ + /* + * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack. + */ + if (aot) { + code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint_noraise"); + } else { + amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint_noraise); + } + amd64_call_reg (code, AMD64_R11); + + amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX); + br_ex_check = code; + amd64_branch8 (code, X86_CC_Z, -1, 1); + + /* + * Exception case: + * We have an exception we want to throw in the caller's frame, so pop + * the trampoline frame and throw from the caller. + */ +#if TARGET_WIN32 + amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, 0); + amd64_pop_reg (code, AMD64_RBP); + mono_add_unwind_op_same_value (unwind_ops, code, buf, AMD64_RBP); +#else + amd64_leave (code); +#endif + /* We are in the parent frame, the exception is in rax */ + /* + * EH is initialized after trampolines, so get the address of the variable + * which contains throw_exception, and load it from there. + */ + if (aot) { + /* Not really a jit icall */ + code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "throw_exception_addr"); + } else { + amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_get_throw_exception_addr ()); + } + amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, sizeof(gpointer)); + amd64_mov_reg_reg (code, AMD64_ARG_REG1, AMD64_RAX, sizeof(mgreg_t)); + /* + * We still have the original return value on the top of the stack, so the + * throw trampoline will use that as the throw site. + */ + amd64_jump_reg (code, AMD64_R11); + + /* Normal case */ + mono_amd64_patch (br_ex_check, code); + /* Restore argument registers, r10 (imt method/rgxtx) and rax (needed for direct calls to C vararg functions). */ for (i = 0; i < AMD64_NREG; ++i) @@ -575,11 +551,16 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t))); /* Restore stack */ +#if TARGET_WIN32 + amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, 0); + amd64_pop_reg (code, AMD64_RBP); + mono_add_unwind_op_same_value (unwind_ops, code, buf, AMD64_RBP); +#else amd64_leave (code); +#endif cfa_offset -= sizeof (mgreg_t); mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, cfa_offset); - if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) { /* Load result */ amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - sizeof(mgreg_t), sizeof(mgreg_t)); @@ -590,8 +571,7 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf } g_assert ((code - buf) <= kMaxCodeSize); - - nacl_global_codeman_validate (&buf, kMaxCodeSize, &code); + g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_MAX_TRAMPOLINE_UNWINDINFO_SIZE)); mono_arch_flush_icache (buf, code - buf); mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL); @@ -612,7 +592,6 @@ mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_ty tramp = mono_get_trampoline_code (tramp_type); -#if defined(__default_codegen__) if ((((guint64)arg1) >> 32) == 0) size = 5 + 1 + 4; else @@ -628,15 +607,6 @@ mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_ty size += 16; code = buf = (guint8 *)mono_domain_code_reserve_align (domain, size, 1); } -#elif defined(__native_client_codegen__) - size = 5 + 1 + 4; - /* Aligning the call site below could */ - /* add up to kNaClAlignment-1 bytes */ - size += (kNaClAlignment-1); - size = NACL_BUNDLE_ALIGN_UP (size); - buf = mono_domain_code_reserve_align (domain, size, kNaClAlignment); - code = buf; -#endif if (far_addr) { amd64_mov_reg_imm (code, AMD64_R11, tramp); @@ -645,7 +615,6 @@ mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_ty amd64_call_code (code, tramp); } /* The trampoline code will obtain the argument from the instruction stream */ -#if defined(__default_codegen__) if ((((guint64)arg1) >> 32) == 0) { *code = 0x4; *(guint32*)(code + 1) = (gint64)arg1; @@ -655,20 +624,12 @@ mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_ty *(guint64*)(code + 1) = (gint64)arg1; code += 9; } -#elif defined(__native_client_codegen__) - /* For NaCl, all tramp args are 32-bit because they're pointers */ - *code = 0x68; /* push imm32 */ - *(guint32*)(code + 1) = (gint32)arg1; - code += 5; -#endif g_assert ((code - buf) <= size); if (code_len) *code_len = size; - nacl_domain_code_validate(domain, &buf, size, &code); - mono_arch_flush_icache (buf, size); mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_SPECIFIC_TRAMPOLINE, mono_get_generic_trampoline_simple_name (tramp_type)); @@ -700,9 +661,9 @@ mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info index -= size - 1; } - tramp_size = NACL_SIZE (64 + 8 * depth, 128 + 8 * depth); + tramp_size = 64 + 8 * depth; - code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size); + code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)); unwind_ops = mono_arch_get_cie_program (); @@ -762,11 +723,11 @@ mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info amd64_jump_code (code, tramp); } - nacl_global_codeman_validate (&buf, tramp_size, &code); mono_arch_flush_icache (buf, code - buf); mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL); g_assert (code - buf <= tramp_size); + g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0))); char *name = mono_get_rgctx_fetch_trampoline_name (slot); *info = mono_tramp_info_create (name, buf, code - buf, ji, unwind_ops); @@ -775,6 +736,42 @@ mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info return buf; } +gpointer +mono_arch_create_general_rgctx_lazy_fetch_trampoline (MonoTrampInfo **info, gboolean aot) +{ + guint8 *code, *buf; + int tramp_size; + MonoJumpInfo *ji = NULL; + GSList *unwind_ops; + + g_assert (aot); + tramp_size = 64; + + code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)); + + unwind_ops = mono_arch_get_cie_program (); + + // FIXME: Currently, we always go to the slow path. + /* This receives a in the rgctx arg reg. */ + /* Load trampoline addr */ + amd64_mov_reg_membase (code, AMD64_R11, MONO_ARCH_RGCTX_REG, 8, 8); + /* move the rgctx pointer to the VTABLE register */ + amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer)); + /* Jump to the trampoline */ + amd64_jump_reg (code, AMD64_R11); + + mono_arch_flush_icache (buf, code - buf); + mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL); + + g_assert (code - buf <= tramp_size); + g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0))); + + if (info) + *info = mono_tramp_info_create ("rgctx_fetch_trampoline_general", buf, code - buf, ji, unwind_ops); + + return buf; +} + void mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg) { @@ -787,59 +784,86 @@ mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg) x86_push_imm (code, (guint64)func_arg); amd64_call_reg (code, AMD64_R11); } +#endif /* !DISABLE_JIT */ - -static void -handler_block_trampoline_helper (gpointer *ptr) +gpointer +mono_amd64_handler_block_trampoline_helper (void) { - MonoJitTlsData *jit_tls = (MonoJitTlsData *)mono_native_tls_get_value (mono_jit_tls_id); - *ptr = jit_tls->handler_block_return_address; + MonoJitTlsData *jit_tls = (MonoJitTlsData *)mono_tls_get_jit_tls (); + return jit_tls->handler_block_return_address; } +#ifndef DISABLE_JIT gpointer mono_arch_create_handler_block_trampoline (MonoTrampInfo **info, gboolean aot) { - guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD); guint8 *code, *buf; int tramp_size = 64; MonoJumpInfo *ji = NULL; GSList *unwind_ops; - g_assert (!aot); - - code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size); + code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)); unwind_ops = mono_arch_get_cie_program (); /* - This trampoline restore the call chain of the handler block then jumps into the code that deals with it. - */ - if (mono_get_jit_tls_offset () != -1) { - code = mono_amd64_emit_tls_get (code, MONO_AMD64_ARG_REG1, mono_get_jit_tls_offset ()); - amd64_mov_reg_membase (code, MONO_AMD64_ARG_REG1, MONO_AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoJitTlsData, handler_block_return_address), 8); - /* Simulate a call */ - amd64_push_reg (code, AMD64_RAX); - mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, 16); - amd64_jump_code (code, tramp); + * This trampoline restore the call chain of the handler block then jumps into the code that deals with it. + * We get here from the ret emitted by CEE_ENDFINALLY. + * The stack is misaligned. + */ + /* Align the stack before the call to mono_amd64_handler_block_trampoline_helper() */ +#ifdef TARGET_WIN32 + /* Also make room for the "register parameter stack area" as specified by the Windows x64 ABI (4 64-bit registers) */ + amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8 + 4 * 8); +#else + amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8); +#endif + if (aot) { + code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_amd64_handler_block_trampoline_helper"); + amd64_call_reg (code, AMD64_R11); + } else { + amd64_mov_reg_imm (code, AMD64_RAX, mono_amd64_handler_block_trampoline_helper); + amd64_call_reg (code, AMD64_RAX); + } + /* Undo stack alignment */ +#ifdef TARGET_WIN32 + amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8 + 4 * 8); +#else + amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8); +#endif + /* Save the result to the stack */ + amd64_push_reg (code, AMD64_RAX); +#ifdef TARGET_WIN32 + /* Make room for the "register parameter stack area" as specified by the Windows x64 ABI (4 64-bit registers) */ + amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 4 * 8); +#endif + if (aot) { + char *name = g_strdup_printf ("trampoline_func_%d", MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD); + code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, name); + amd64_mov_reg_reg (code, AMD64_RAX, AMD64_R11, 8); } else { - /*Slow path uses a c helper*/ - amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RSP, 8); - amd64_mov_reg_imm (code, AMD64_RAX, tramp); - amd64_push_reg (code, AMD64_RAX); - mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, 16); - amd64_push_reg (code, AMD64_RAX); - mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, 24); - amd64_jump_code (code, handler_block_trampoline_helper); + amd64_mov_reg_imm (code, AMD64_RAX, mono_get_trampoline_func (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD)); } + /* The stack is aligned */ + amd64_call_reg (code, AMD64_RAX); +#ifdef TARGET_WIN32 + amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 4 * 8); +#endif + /* Load return address */ + amd64_pop_reg (code, AMD64_RAX); + /* The stack is misaligned, thats what the code we branch to expects */ + amd64_jump_reg (code, AMD64_RAX); mono_arch_flush_icache (buf, code - buf); mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL); g_assert (code - buf <= tramp_size); + g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0))); *info = mono_tramp_info_create ("handler_block_trampoline", buf, code - buf, ji, unwind_ops); return buf; } +#endif /* !DISABLE_JIT */ /* * mono_arch_get_call_target: @@ -867,15 +891,10 @@ mono_arch_get_call_target (guint8 *code) guint32 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code) { -#if defined(__native_client__) || defined(__native_client_codegen__) - /* 18 = 3 (mov opcode) + 4 (disp) + 10 (nacljmp) + 1 (push opcode) */ - /* See aot-compiler.c arch_emit_plt_entry for details. */ - return *(guint32*)(plt_entry + 18); -#else return *(guint32*)(plt_entry + 6); -#endif } +#ifndef DISABLE_JIT /* * mono_arch_create_sdb_trampoline: * @@ -886,13 +905,13 @@ mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code) guint8* mono_arch_create_sdb_trampoline (gboolean single_step, MonoTrampInfo **info, gboolean aot) { - int tramp_size = 256; + int tramp_size = 512; int i, framesize, ctx_offset, cfa_offset, gregs_offset; guint8 *code, *buf; GSList *unwind_ops = NULL; MonoJumpInfo *ji = NULL; - code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size); + code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size + MONO_MAX_TRAMPOLINE_UNWINDINFO_SIZE); framesize = 0; #ifdef TARGET_WIN32 @@ -918,6 +937,7 @@ mono_arch_create_sdb_trampoline (gboolean single_step, MonoTrampInfo **info, gbo amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t)); mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP); + mono_add_unwind_op_fp_alloc (unwind_ops, code, buf, AMD64_RBP, 0); amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize); gregs_offset = ctx_offset + MONO_STRUCT_OFFSET (MonoContext, gregs); @@ -960,13 +980,21 @@ mono_arch_create_sdb_trampoline (gboolean single_step, MonoTrampInfo **info, gbo amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, gregs_offset + (AMD64_RIP * sizeof (mgreg_t)), sizeof (mgreg_t)); amd64_mov_membase_reg (code, AMD64_RBP, sizeof (mgreg_t), AMD64_R11, sizeof (mgreg_t)); +#if TARGET_WIN32 + amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, 0); + amd64_pop_reg (code, AMD64_RBP); + mono_add_unwind_op_same_value (unwind_ops, code, buf, AMD64_RBP); +#else amd64_leave (code); +#endif cfa_offset -= sizeof (mgreg_t); mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, cfa_offset); amd64_ret (code); mono_arch_flush_icache (code, code - buf); + mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL); g_assert (code - buf <= tramp_size); + g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_MAX_TRAMPOLINE_UNWINDINFO_SIZE)); const char *tramp_name = single_step ? "sdb_single_step_trampoline" : "sdb_breakpoint_trampoline"; *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops); @@ -974,8 +1002,236 @@ mono_arch_create_sdb_trampoline (gboolean single_step, MonoTrampInfo **info, gbo return buf; } -#if defined(ENABLE_GSHAREDVT) && defined(MONO_ARCH_GSHAREDVT_SUPPORTED) +/* + * mono_arch_get_enter_icall_trampoline: + * + * A trampoline that handles the transition from interpreter into native + * world. It requiers to set up a descriptor (InterpMethodArguments), so the + * trampoline can translate the arguments into the native calling convention. + * + * See also `build_args_from_sig ()` in interp.c. + */ +gpointer +mono_arch_get_enter_icall_trampoline (MonoTrampInfo **info) +{ +#ifdef ENABLE_INTERPRETER + const int gregs_num = INTERP_ICALL_TRAMP_IARGS; + const int fregs_num = INTERP_ICALL_TRAMP_FARGS; + guint8 *start = NULL, *code, *label_gexits [gregs_num], *label_fexits [fregs_num], *label_leave_tramp [3], *label_is_float_ret; + MonoJumpInfo *ji = NULL; + GSList *unwind_ops = NULL; + static int farg_regs[] = {AMD64_XMM0, AMD64_XMM1, AMD64_XMM2}; + int buf_len, i, framesize = 0, off_rbp, off_methodargs, off_targetaddr; + + buf_len = 512 + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0); + start = code = (guint8 *) mono_global_codeman_reserve (buf_len); + + off_rbp = -framesize; + + framesize += sizeof (mgreg_t); + off_methodargs = -framesize; + + framesize += sizeof (mgreg_t); + off_targetaddr = -framesize; + + framesize += (gregs_num - PARAM_REGS) * sizeof (mgreg_t); + + amd64_push_reg (code, AMD64_RBP); + amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof (mgreg_t)); + amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, ALIGN_TO (framesize, MONO_ARCH_FRAME_ALIGNMENT)); + + /* save InterpMethodArguments* onto stack */ + amd64_mov_membase_reg (code, AMD64_RBP, off_methodargs, AMD64_ARG_REG2, sizeof (mgreg_t)); + + /* save target address on stack */ + amd64_mov_membase_reg (code, AMD64_RBP, off_targetaddr, AMD64_ARG_REG1, sizeof (mgreg_t)); + + /* load pointer to InterpMethodArguments* into R11 */ + amd64_mov_reg_reg (code, AMD64_R11, AMD64_ARG_REG2, 8); + + /* move flen into RAX */ + amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, flen), sizeof (mgreg_t)); + /* load pointer to fargs into R11 */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, fargs), sizeof (mgreg_t)); + + for (i = 0; i < fregs_num; ++i) { + amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX); + label_fexits [i] = code; + x86_branch8 (code, X86_CC_Z, 0, FALSE); + + amd64_sse_movsd_reg_membase (code, farg_regs [i], AMD64_R11, i * sizeof (double)); + amd64_dec_reg_size (code, AMD64_RAX, 1); + } + + for (i = 0; i < fregs_num; i++) + x86_patch (label_fexits [i], code); + + /* load pointer to InterpMethodArguments* into R11 */ + amd64_mov_reg_reg (code, AMD64_R11, AMD64_ARG_REG2, sizeof (mgreg_t)); + /* move ilen into RAX */ + amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, ilen), sizeof (mgreg_t)); + + int stack_offset = 0; + for (i = 0; i < gregs_num; i++) { + amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX); + label_gexits [i] = code; + x86_branch32 (code, X86_CC_Z, 0, FALSE); + + /* load pointer to InterpMethodArguments* into R11 */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_methodargs, sizeof (mgreg_t)); + /* load pointer to iargs into R11 */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, iargs), sizeof (mgreg_t)); + + if (i < PARAM_REGS) { + amd64_mov_reg_membase (code, param_regs [i], AMD64_R11, i * sizeof (mgreg_t), sizeof (mgreg_t)); + } else { + amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, i * sizeof (mgreg_t), sizeof (mgreg_t)); + amd64_mov_membase_reg (code, AMD64_RSP, stack_offset, AMD64_R11, sizeof (mgreg_t)); + stack_offset += sizeof (mgreg_t); + } + amd64_dec_reg_size (code, AMD64_RAX, 1); + } + + for (i = 0; i < gregs_num; i++) + x86_patch (label_gexits [i], code); + + /* load target addr */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_targetaddr, sizeof (mgreg_t)); + + /* call into native function */ + amd64_call_reg (code, AMD64_R11); + + /* load InterpMethodArguments */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_methodargs, sizeof (mgreg_t)); + + /* load is_float_ret */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, is_float_ret), sizeof (mgreg_t)); + + /* check if a float return value is expected */ + amd64_test_reg_reg (code, AMD64_R11, AMD64_R11); + + label_is_float_ret = code; + x86_branch8 (code, X86_CC_NZ, 0, FALSE); + + /* greg return */ + /* load InterpMethodArguments */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_methodargs, sizeof (mgreg_t)); + /* load retval */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, retval), sizeof (mgreg_t)); + + amd64_test_reg_reg (code, AMD64_R11, AMD64_R11); + label_leave_tramp [0] = code; + x86_branch8 (code, X86_CC_Z, 0, FALSE); + + amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RAX, sizeof (mgreg_t)); + + label_leave_tramp [1] = code; + x86_jump8 (code, 0); + + /* freg return */ + x86_patch (label_is_float_ret, code); + /* load InterpMethodArguments */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_methodargs, sizeof (mgreg_t)); + /* load retval */ + amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, retval), sizeof (mgreg_t)); + + amd64_test_reg_reg (code, AMD64_R11, AMD64_R11); + label_leave_tramp [2] = code; + x86_branch8 (code, X86_CC_Z, 0, FALSE); + + amd64_sse_movsd_membase_reg (code, AMD64_R11, 0, AMD64_XMM0); + + for (i = 0; i < 3; i++) + x86_patch (label_leave_tramp [i], code); + + amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, ALIGN_TO (framesize, MONO_ARCH_FRAME_ALIGNMENT)); + amd64_pop_reg (code, AMD64_RBP); + amd64_ret (code); + + g_assert (code - start < buf_len); + + mono_arch_flush_icache (start, code - start); + mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_EXCEPTION_HANDLING, NULL); + + if (info) + *info = mono_tramp_info_create ("enter_icall_trampoline", start, code - start, ji, unwind_ops); + + return start; +#else + g_assert_not_reached (); + return NULL; +#endif /* ENABLE_INTERPRETER */ +} +#endif /* !DISABLE_JIT */ + +#ifdef DISABLE_JIT +gpointer +mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr) +{ + g_assert_not_reached (); + return NULL; +} + +gpointer +mono_arch_get_static_rgctx_trampoline (gpointer arg, gpointer addr) +{ + g_assert_not_reached (); + return NULL; +} + +gpointer +mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot) +{ + g_assert_not_reached (); + return NULL; +} + +guchar* +mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot) +{ + g_assert_not_reached (); + return NULL; +} + +gpointer +mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len) +{ + g_assert_not_reached (); + return NULL; +} + +gpointer +mono_arch_create_general_rgctx_lazy_fetch_trampoline (MonoTrampInfo **info, gboolean aot) +{ + g_assert_not_reached (); + return NULL; +} -#include "../../../mono-extensions/mono/mini/tramp-amd64-gsharedvt.c" +gpointer +mono_arch_create_handler_block_trampoline (MonoTrampInfo **info, gboolean aot) +{ + g_assert_not_reached (); + return NULL; +} -#endif /* !ENABLE_GSHAREDVT */ +void +mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg) +{ + g_assert_not_reached (); + return; +} + +guint8* +mono_arch_create_sdb_trampoline (gboolean single_step, MonoTrampInfo **info, gboolean aot) +{ + g_assert_not_reached (); + return NULL; +} + +gpointer +mono_arch_get_enter_icall_trampoline (MonoTrampInfo **info) +{ + g_assert_not_reached (); + return NULL; +} +#endif /* DISABLE_JIT */