/* * Stack-less Just-In-Time compiler * * Copyright 2009-2012 Zoltan Herczeg (hzmester@freemail.hu). All rights reserved. * * Redistribution and use in source and binary forms, with or without modification, are * permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this list of * conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, this list * of conditions and the following disclaimer in the documentation and/or other materials * provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) AND CONTRIBUTORS ``AS IS'' AND ANY * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT * SHALL THE COPYRIGHT HOLDER(S) OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #ifndef _SLJIT_LIR_H_ #define _SLJIT_LIR_H_ /* ------------------------------------------------------------------------ Stack-Less JIT compiler for multiple architectures (x86, ARM, PowerPC) ------------------------------------------------------------------------ Short description Advantages: - The execution can be continued from any LIR instruction. In other words, it is possible to jump to any label from anywhere, even from a code fragment, which is compiled later, if both compiled code shares the same context. See sljit_emit_enter for more details - Supports self modifying code: target of (conditional) jump and call instructions and some constant values can be dynamically modified during runtime - although it is not suggested to do it frequently - can be used for inline caching: save an important value once in the instruction stream - since this feature limits the optimization possibilities, a special flag must be passed at compile time when these instructions are emitted - A fixed stack space can be allocated for local variables - The compiler is thread-safe - The compiler is highly configurable through preprocessor macros. You can disable unneeded features (multithreading in single threaded applications), and you can use your own system functions (including memory allocators). See sljitConfig.h Disadvantages: - No automatic register allocation, and temporary results are not stored on the stack. (hence the name comes) - Limited number of registers (only 6+4 integer registers, max 3+2 scratch, max 3+2 saved and 6 floating point registers) In practice: - This approach is very effective for interpreters - One of the saved registers typically points to a stack interface - It can jump to any exception handler anytime (even if it belongs to another function) - Hot paths can be modified during runtime reflecting the changes of the fastest execution path of the dynamic language - SLJIT supports complex memory addressing modes - mainly position and context independent code (except some cases) For valgrind users: - pass --smc-check=all argument to valgrind, since JIT is a "self-modifying code" */ #if !(defined SLJIT_NO_DEFAULT_CONFIG && SLJIT_NO_DEFAULT_CONFIG) #include "sljitConfig.h" #endif /* The following header file defines useful macros for fine tuning sljit based code generators. They are listed in the beginning of sljitConfigInternal.h */ #include "sljitConfigInternal.h" /* --------------------------------------------------------------------- */ /* Error codes */ /* --------------------------------------------------------------------- */ /* Indicates no error. */ #define SLJIT_SUCCESS 0 /* After the call of sljit_generate_code(), the error code of the compiler is set to this value to avoid future sljit calls (in debug mode at least). The complier should be freed after sljit_generate_code(). */ #define SLJIT_ERR_COMPILED 1 /* Cannot allocate non executable memory. */ #define SLJIT_ERR_ALLOC_FAILED 2 /* Cannot allocate executable memory. Only for sljit_generate_code() */ #define SLJIT_ERR_EX_ALLOC_FAILED 3 /* return value for SLJIT_CONFIG_UNSUPPORTED empty architecture. */ #define SLJIT_ERR_UNSUPPORTED 4 /* --------------------------------------------------------------------- */ /* Registers */ /* --------------------------------------------------------------------- */ #define SLJIT_UNUSED 0 /* Scratch (temporary) registers whose may not preserve their values across function calls. */ #define SLJIT_SCRATCH_REG1 1 #define SLJIT_SCRATCH_REG2 2 #define SLJIT_SCRATCH_REG3 3 /* Note: extra registers cannot be used for memory addressing. */ /* Note: on x86-32, these registers are emulated (using stack loads & stores). */ #define SLJIT_TEMPORARY_EREG1 4 #define SLJIT_TEMPORARY_EREG2 5 /* Saved registers whose preserve their values across function calls. */ #define SLJIT_SAVED_REG1 6 #define SLJIT_SAVED_REG2 7 #define SLJIT_SAVED_REG3 8 /* Note: extra registers cannot be used for memory addressing. */ /* Note: on x86-32, these registers are emulated (using stack loads & stores). */ #define SLJIT_SAVED_EREG1 9 #define SLJIT_SAVED_EREG2 10 /* Read-only register (cannot be the destination of an operation). Only SLJIT_MEM1(SLJIT_LOCALS_REG) addressing mode is allowed since several ABIs has certain limitations about the stack layout. However sljit_get_local_base() can be used to obtain the offset of a value on the stack. */ #define SLJIT_LOCALS_REG 11 /* Number of registers. */ #define SLJIT_NO_TMP_REGISTERS 5 #define SLJIT_NO_GEN_REGISTERS 5 #define SLJIT_NO_REGISTERS 11 /* Return with machine word. */ #define SLJIT_RETURN_REG SLJIT_SCRATCH_REG1 /* x86 prefers specific registers for special purposes. In case of shift by register it supports only SLJIT_SCRATCH_REG3 for shift argument (which is the src2 argument of sljit_emit_op2). If another register is used, sljit must exchange data between registers which cause a minor slowdown. Other architectures has no such limitation. */ #define SLJIT_PREF_SHIFT_REG SLJIT_SCRATCH_REG3 /* --------------------------------------------------------------------- */ /* Floating point registers */ /* --------------------------------------------------------------------- */ /* Note: SLJIT_UNUSED as destination is not valid for floating point operations, since they cannot be used for setting flags. */ /* Floating point operations are performed on double or single precision values. */ #define SLJIT_FLOAT_REG1 1 #define SLJIT_FLOAT_REG2 2 #define SLJIT_FLOAT_REG3 3 #define SLJIT_FLOAT_REG4 4 #define SLJIT_FLOAT_REG5 5 #define SLJIT_FLOAT_REG6 6 #define SLJIT_NO_FLOAT_REGISTERS 6 /* --------------------------------------------------------------------- */ /* Main structures and functions */ /* --------------------------------------------------------------------- */ struct sljit_memory_fragment { struct sljit_memory_fragment *next; sljit_uw used_size; /* Must be aligned to sljit_sw. */ sljit_ub memory[1]; }; struct sljit_label { struct sljit_label *next; sljit_uw addr; /* The maximum size difference. */ sljit_uw size; }; struct sljit_jump { struct sljit_jump *next; sljit_uw addr; sljit_sw flags; union { sljit_uw target; struct sljit_label* label; } u; }; struct sljit_const { struct sljit_const *next; sljit_uw addr; }; struct sljit_compiler { sljit_si error; struct sljit_label *labels; struct sljit_jump *jumps; struct sljit_const *consts; struct sljit_label *last_label; struct sljit_jump *last_jump; struct sljit_const *last_const; struct sljit_memory_fragment *buf; struct sljit_memory_fragment *abuf; /* Used local registers. */ sljit_si scratches; /* Used saved registers. */ sljit_si saveds; /* Local stack size. */ sljit_si local_size; /* Code size. */ sljit_uw size; /* For statistical purposes. */ sljit_uw executable_size; #if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32) sljit_si args; sljit_si locals_offset; sljit_si scratches_start; sljit_si saveds_start; #endif #if (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64) sljit_si mode32; #endif #if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32) || (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64) sljit_si flags_saved; #endif #if (defined SLJIT_CONFIG_ARM_V5 && SLJIT_CONFIG_ARM_V5) /* Constant pool handling. */ sljit_uw *cpool; sljit_ub *cpool_unique; sljit_uw cpool_diff; sljit_uw cpool_fill; /* Other members. */ /* Contains pointer, "ldr pc, [...]" pairs. */ sljit_uw patches; #endif #if (defined SLJIT_CONFIG_ARM_V5 && SLJIT_CONFIG_ARM_V5) || (defined SLJIT_CONFIG_ARM_V7 && SLJIT_CONFIG_ARM_V7) /* Temporary fields. */ sljit_uw shift_imm; sljit_si cache_arg; sljit_sw cache_argw; #endif #if (defined SLJIT_CONFIG_ARM_THUMB2 && SLJIT_CONFIG_ARM_THUMB2) sljit_si cache_arg; sljit_sw cache_argw; #endif #if (defined SLJIT_CONFIG_ARM_64 && SLJIT_CONFIG_ARM_64) sljit_si locals_offset; sljit_si cache_arg; sljit_sw cache_argw; #endif #if (defined SLJIT_CONFIG_PPC_32 && SLJIT_CONFIG_PPC_32) || (defined SLJIT_CONFIG_PPC_64 && SLJIT_CONFIG_PPC_64) sljit_sw imm; sljit_si cache_arg; sljit_sw cache_argw; #endif #if (defined SLJIT_CONFIG_MIPS_32 && SLJIT_CONFIG_MIPS_32) || (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64) sljit_si delay_slot; sljit_si cache_arg; sljit_sw cache_argw; #endif #if (defined SLJIT_CONFIG_SPARC_32 && SLJIT_CONFIG_SPARC_32) sljit_si delay_slot; sljit_si cache_arg; sljit_sw cache_argw; #endif #if (defined SLJIT_CONFIG_TILEGX && SLJIT_CONFIG_TILEGX) sljit_si cache_arg; sljit_sw cache_argw; #endif #if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) FILE* verbose; #endif #if (defined SLJIT_DEBUG && SLJIT_DEBUG) /* Local size passed to the functions. */ sljit_si logical_local_size; #endif #if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) || (defined SLJIT_DEBUG && SLJIT_DEBUG) sljit_si skip_checks; #endif }; /* --------------------------------------------------------------------- */ /* Main functions */ /* --------------------------------------------------------------------- */ /* Creates an sljit compiler. Returns NULL if failed. */ SLJIT_API_FUNC_ATTRIBUTE struct sljit_compiler* sljit_create_compiler(void); /* Free everything except the compiled machine code. */ SLJIT_API_FUNC_ATTRIBUTE void sljit_free_compiler(struct sljit_compiler *compiler); /* Returns the current error code. If an error is occurred, future sljit calls which uses the same compiler argument returns early with the same error code. Thus there is no need for checking the error after every call, it is enough to do it before the code is compiled. Removing these checks increases the performance of the compiling process. */ static SLJIT_INLINE sljit_si sljit_get_compiler_error(struct sljit_compiler *compiler) { return compiler->error; } /* Allocate a small amount of memory. The size must be <= 64 bytes on 32 bit, and <= 128 bytes on 64 bit architectures. The memory area is owned by the compiler, and freed by sljit_free_compiler. The returned pointer is sizeof(sljit_sw) aligned. Excellent for allocating small blocks during the compiling, and no need to worry about freeing them. The size is enough to contain at most 16 pointers. If the size is outside of the range, the function will return with NULL. However, this return value does not indicate that there is no more memory (does not set the current error code of the compiler to out-of-memory status). */ SLJIT_API_FUNC_ATTRIBUTE void* sljit_alloc_memory(struct sljit_compiler *compiler, sljit_si size); #if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) /* Passing NULL disables verbose. */ SLJIT_API_FUNC_ATTRIBUTE void sljit_compiler_verbose(struct sljit_compiler *compiler, FILE* verbose); #endif SLJIT_API_FUNC_ATTRIBUTE void* sljit_generate_code(struct sljit_compiler *compiler); SLJIT_API_FUNC_ATTRIBUTE void sljit_free_code(void* code); /* After the machine code generation is finished we can retrieve the allocated executable memory size, although this area may not be fully filled with instructions depending on some optimizations. This function is useful only for statistical purposes. Before a successful code generation, this function returns with 0. */ static SLJIT_INLINE sljit_uw sljit_get_generated_code_size(struct sljit_compiler *compiler) { return compiler->executable_size; } /* Instruction generation. Returns with any error code. If there is no error, they return with SLJIT_SUCCESS. */ /* The executable code is basically a function call from the viewpoint of the C language. The function calls must obey to the ABI (Application Binary Interface) of the platform, which specify the purpose of machine registers and stack handling among other things. The sljit_emit_enter function emits the necessary instructions for setting up a new context for the executable code and moves function arguments to the saved registers. The number of arguments are specified in the "args" parameter and the first argument goes to SLJIT_SAVED_REG1, the second goes to SLJIT_SAVED_REG2 and so on. The number of scratch and saved registers are passed in "scratches" and "saveds" arguments respectively. Since the saved registers contains the arguments, "args" must be less or equal than "saveds". The sljit_emit_enter is also capable of allocating a stack space for local variables. The "local_size" argument contains the size in bytes of this local area and its staring address is stored in SLJIT_LOCALS_REG. However the SLJIT_LOCALS_REG is not necessary the machine stack pointer. The memory bytes between SLJIT_LOCALS_REG (inclusive) and SLJIT_LOCALS_REG + local_size (exclusive) can be modified freely until the function returns. The stack space is uninitialized. Note: every call of sljit_emit_enter and sljit_set_context overwrites the previous context. */ #define SLJIT_MAX_LOCAL_SIZE 65536 SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_enter(struct sljit_compiler *compiler, sljit_si args, sljit_si scratches, sljit_si saveds, sljit_si local_size); /* The machine code has a context (which contains the local stack space size, number of used registers, etc.) which initialized by sljit_emit_enter. Several functions (like sljit_emit_return) requres this context to be able to generate the appropriate code. However, some code fragments (like inline cache) may have no normal entry point so their context is unknown for the compiler. Using the function below we can specify their context. Note: every call of sljit_emit_enter and sljit_set_context overwrites the previous context. */ SLJIT_API_FUNC_ATTRIBUTE void sljit_set_context(struct sljit_compiler *compiler, sljit_si args, sljit_si scratches, sljit_si saveds, sljit_si local_size); /* Return from machine code. The op argument can be SLJIT_UNUSED which means the function does not return with anything or any opcode between SLJIT_MOV and SLJIT_MOV_P (see sljit_emit_op1). As for src and srcw they must be 0 if op is SLJIT_UNUSED, otherwise see below the description about source and destination arguments. */ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_return(struct sljit_compiler *compiler, sljit_si op, sljit_si src, sljit_sw srcw); /* Fast calling mechanism for utility functions (see SLJIT_FAST_CALL). All registers and even the stack frame is passed to the callee. The return address is preserved in dst/dstw by sljit_emit_fast_enter (the type of the value stored by this function is sljit_p), and sljit_emit_fast_return can use this as a return value later. */ /* Note: only for sljit specific, non ABI compilant calls. Fast, since only a few machine instructions are needed. Excellent for small uility functions, where saving registers and setting up a new stack frame would cost too much performance. However, it is still possible to return to the address of the caller (or anywhere else). */ /* Note: flags are not changed (unlike sljit_emit_enter / sljit_emit_return). */ /* Note: although sljit_emit_fast_return could be replaced by an ijump, it is not suggested, since many architectures do clever branch prediction on call / return instruction pairs. */ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_enter(struct sljit_compiler *compiler, sljit_si dst, sljit_sw dstw); SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_si src, sljit_sw srcw); /* Source and destination values for arithmetical instructions imm - a simple immediate value (cannot be used as a destination) reg - any of the registers (immediate argument must be 0) [imm] - absolute immediate memory address [reg+imm] - indirect memory address [reg+(reg<addr; } static SLJIT_INLINE sljit_uw sljit_get_jump_addr(struct sljit_jump *jump) { return jump->addr; } static SLJIT_INLINE sljit_uw sljit_get_const_addr(struct sljit_const *const_) { return const_->addr; } /* Only the address is required to rewrite the code. */ SLJIT_API_FUNC_ATTRIBUTE void sljit_set_jump_addr(sljit_uw addr, sljit_uw new_addr); SLJIT_API_FUNC_ATTRIBUTE void sljit_set_const(sljit_uw addr, sljit_sw new_constant); /* --------------------------------------------------------------------- */ /* Miscellaneous utility functions */ /* --------------------------------------------------------------------- */ #define SLJIT_MAJOR_VERSION 0 #define SLJIT_MINOR_VERSION 91 /* Get the human readable name of the platform. Can be useful on platforms like ARM, where ARM and Thumb2 functions can be mixed, and it is useful to know the type of the code generator. */ SLJIT_API_FUNC_ATTRIBUTE SLJIT_CONST char* sljit_get_platform_name(void); /* Portable helper function to get an offset of a member. */ #define SLJIT_OFFSETOF(base, member) ((sljit_sw)(&((base*)0x10)->member) - 0x10) #if (defined SLJIT_UTIL_GLOBAL_LOCK && SLJIT_UTIL_GLOBAL_LOCK) /* This global lock is useful to compile common functions. */ SLJIT_API_FUNC_ATTRIBUTE void SLJIT_CALL sljit_grab_lock(void); SLJIT_API_FUNC_ATTRIBUTE void SLJIT_CALL sljit_release_lock(void); #endif #if (defined SLJIT_UTIL_STACK && SLJIT_UTIL_STACK) /* The sljit_stack is a utiliy feature of sljit, which allocates a writable memory region between base (inclusive) and limit (exclusive). Both base and limit is a pointer, and base is always <= than limit. This feature uses the "address space reserve" feature of modern operating systems. Basically we don't need to allocate a huge memory block in one step for the worst case, we can start with a smaller chunk and extend it later. Since the address space is reserved, the data never copied to other regions, thus it is safe to store pointers here. */ /* Note: The base field is aligned to PAGE_SIZE bytes (usually 4k or more). Note: stack growing should not happen in small steps: 4k, 16k or even bigger growth is better. Note: this structure may not be supported by all operating systems. Some kind of fallback mechanism is suggested when SLJIT_UTIL_STACK is not defined. */ struct sljit_stack { /* User data, anything can be stored here. Starting with the same value as base. */ sljit_uw top; /* These members are read only. */ sljit_uw base; sljit_uw limit; sljit_uw max_limit; }; /* Returns NULL if unsuccessful. Note: limit and max_limit contains the size for stack allocation Note: the top field is initialized to base. */ SLJIT_API_FUNC_ATTRIBUTE struct sljit_stack* SLJIT_CALL sljit_allocate_stack(sljit_uw limit, sljit_uw max_limit); SLJIT_API_FUNC_ATTRIBUTE void SLJIT_CALL sljit_free_stack(struct sljit_stack* stack); /* Can be used to increase (allocate) or decrease (free) the memory area. Returns with a non-zero value if unsuccessful. If new_limit is greater than max_limit, it will fail. It is very easy to implement a stack data structure, since the growth ratio can be added to the current limit, and sljit_stack_resize will do all the necessary checks. The fields of the stack are not changed if sljit_stack_resize fails. */ SLJIT_API_FUNC_ATTRIBUTE sljit_sw SLJIT_CALL sljit_stack_resize(struct sljit_stack* stack, sljit_uw new_limit); #endif /* (defined SLJIT_UTIL_STACK && SLJIT_UTIL_STACK) */ #if !(defined SLJIT_INDIRECT_CALL && SLJIT_INDIRECT_CALL) /* Get the entry address of a given function. */ #define SLJIT_FUNC_OFFSET(func_name) ((sljit_sw)func_name) #else /* !(defined SLJIT_INDIRECT_CALL && SLJIT_INDIRECT_CALL) */ /* All JIT related code should be placed in the same context (library, binary, etc.). */ #define SLJIT_FUNC_OFFSET(func_name) (*(sljit_sw*)(void*)func_name) /* For powerpc64, the function pointers point to a context descriptor. */ struct sljit_function_context { sljit_sw addr; sljit_sw r2; sljit_sw r11; }; /* Fill the context arguments using the addr and the function. If func_ptr is NULL, it will not be set to the address of context If addr is NULL, the function address also comes from the func pointer. */ SLJIT_API_FUNC_ATTRIBUTE void sljit_set_function_context(void** func_ptr, struct sljit_function_context* context, sljit_sw addr, void* func); #endif /* !(defined SLJIT_INDIRECT_CALL && SLJIT_INDIRECT_CALL) */ #endif /* _SLJIT_LIR_H_ */