Incomputer programming, aninline assembler is a feature of somecompilers that allows low-level code written inassembly language to be embedded within a program, among code that otherwise has been compiled from ahigher-level language such asC orAda.
The embedding of assembly language code is usually done for one of these reasons:[1]
On the other hand, inline assembler poses a direct problem for the compiler itself as it complicates the analysis of what is done to each variable, a key part of register allocation.[2] This means the performance might actually decrease. Inline assembler also complicates future porting and maintenance of a program.[1]
Alternative facilities are often provided as a way to simplify the work for both the compiler and the programmer.Intrinsic functions for special instructions are provided by most compilers and C-function wrappers for arbitrary system calls are available on everyUnix platform.
The ISO C++ standard and ISO C standards (annex J) specify a conditionally supported syntax for inline assembler:
An asm declaration has the form
asm-declaration:
asm (string-literal ) ;
The asm declaration is conditionally-supported; its meaning is implementation-defined.[3]
This definition, however, is rarely used in actual C, as it is simultaneously too liberal (in the interpretation) and too restricted (in the use of one string literal only).
In practical use, inline assembly operating on values is rarely standalone as free-floating code. Since the programmer cannot predict what register a variable is assigned to, compilers typically provide a way to substitute them in as an extension.
There are, in general, two types of inline assembly supported by C/C++ compilers:
The two families of extensions represent different understandings of division of labor in processing inline assembly. The GCC form preserves the overall syntax of the language and compartmentizes what the compiler needs to know: what is needed and what is changed. It does not explicitly require the compiler to understand instruction names, as the compiler is only needed to substitute its register assignments, plus a fewmov operations, to handle the input requirements. However, the user is prone to specifying clobbered registers incorrectly. The MSVC form of an embeddeddomain-specific language provides ease of writing, but it requires the compiler itself to know about opcode names and their clobbering properties, demanding extra attention in maintenance and porting.[7] It is still possible to check GCC-style assembly for clobber mistakes with knowledge of the instruction set.[8]
GNAT (Ada language frontend of the GCC suite), andLLVM uses the GCC syntax.[9][10] TheD programming language uses a DSL similar to the MSVC extension officially for x86_64,[11] but the LLVM-based LDC also provides the GCC-style syntax on every architecture.[12] MSVC only supports inline assembler on 32-bit x86.[5]
The Rust language has since migrated to a syntax abstracting away inline assembly options further than the LLVM (GCC-style) version. It provides enough information to allow transforming the block into an externally-assembled function if the backend could not handle embedded assembly.[7]
Calling an operating system directly is generally not possible under a system using protected memory. The OS runs at a more privileged level (kernel mode) than the user (user mode); a (software)interrupt is used to make requests to the operating system. This is rarely a feature in a higher-level language, and sowrapper functions for system calls are written using inline assembler.
The following C code example shows an x86 system call wrapper inAT&T assembler syntax, using theGNU Assembler. Such calls are normally written with the aid of macros; the full code is included for clarity. In this particular case, the wrapper performs a system call of a number given by the caller with three operands, returning the result.[13]
To recap, GCC supports bothbasic andextended assembly. The former simply passes text verbatim to the assembler, while the latter performs some substitutions for register locations.[4]
externinterrno;intsyscall3(intnum,intarg1,intarg2,intarg3){intres;__asm__("int $0x80"/* make the request to the OS */:"=a"(res),/* return result in eax ("a") */"+b"(arg1),/* pass arg1 in ebx ("b") [as a "+" output because the syscall may change it] */"+c"(arg2),/* pass arg2 in ecx ("c") [ditto] */"+d"(arg3)/* pass arg3 in edx ("d") [ditto] */:"a"(num)/* pass system call number in eax ("a") */:"memory","cc",/* announce to the compiler that the memory and condition codes have been modified */"esi","edi","ebp");/* these registers are clobbered [changed by the syscall] too *//* The operating system will return a negative value on error; * wrappers return -1 on error and set the errno global variable */if(-125<=res&&res<0){errno=-res;res=-1;}returnres;}
This example of inline assembly from theD programming language shows code that computes the tangent of x using thex86'sFPU (x87) instructions.
// Compute the tangent of xrealtan(realx){asm{fldx[EBP];// load xfxam;// test for oddball valuesfstswAX;sahf;jctrigerr;// C0 = 1: x is NAN, infinity, or empty// 387's can handle denormalsSC18:fptan;fstpST(0);// dump X, which is always 1fstswAX;sahf;// if (!(fp_status & 0x20)) goto LretjnpLret;// C2 = 1: x is out of range, do argument reductionfldpi;// load pifxch;SC17:fprem1;// reminder (partial)fstswAX;sahf;jpSC17;// C2 = 1: partial reminder, need to loopfstpST(1);// remove pi from stackjmpSC18;}trigerr:returnreal.nan;Lret:// No need to manually return anything as the value is already on FP stack;}
Thefstsw-sahf followed by conditional jump idiom is used to access the x87 FPU status word bits C0 and C2.fstsw stores the status in a general-purpose register; sahf sets theFLAGS register to the higher 8 bits of the register; and the jump is used to judge on whatever flag bit that happens to correspond to the FPU status bit.[14]
However it is possible to implement support for inline assembly without support from the compiler backend by using an external assembler instead.Pull Request for status tracking
syscall(2) – Linux Programmer'sManual – System CallsThe FNSTSW AX form of the instruction is used primarily in conditional branching...