asmjit::x86::Compiler Class Reference [¶]
Inheritance diagram for asmjit::x86::Compiler:
X86/X64 compiler implementation.
Compiler Basics
The first x86::Compiler example shows how to generate a function that simply returns an integer value. It's an analogy to the first Assembler example:
#include <asmjit/x86.h>
#include <stdio.h>
using namespace asmjit;
int main() {
using Func = int (*)(void); // Signature of the generated function.
JitRuntime rt; // Runtime specialized for JIT code execution.
rt.cpu_features());
cc.add_func(FuncSignature::build<int>());// Begin a function of `int fn(void)` signature.
x86::Gp virt_reg = cc.new_gp32(); // Create a 32-bit general purpose register.
cc.mov(virt_reg, 1); // Move one to our virtual register `virt_reg`.
cc.ret(virt_reg); // Return `virt_reg` from the function.
cc.end_func(); // End of the function body.
cc.finalize(); // Translate and assemble the whole 'cc' content.
// ----> x86::Compiler is no longer needed from here and can be destroyed <----
Func fn;
if (err != Error::kOk) {
return 1; // Handle a possible error returned by AsmJit.
}
// ----> CodeHolder is no longer needed from here and can be destroyed <----
int result = fn(); // Execute the generated code.
printf("%d\n", result); // Print the resulting "1".
rt.release(fn); // Explicitly remove the function from the runtime.
return 0;
}
The BaseCompiler::add_func() and BaseCompiler::end_func() functions are used to define the function and its end. Both must be called per function, but the body doesn't have to be generated in sequence. An example of generating two functions will be shown later. The next example shows more complicated code that contain a loop and generates a simple memory copy function that uses uint32_t items:
#include <asmjit/x86.h>
#include <stdio.h>
using namespace asmjit;
int main() {
// Signature of the generated function.
using MemCpy32 = void (*)(uint32_t* dst, const uint32_t* src, size_t count);
JitRuntime rt; // Runtime specialized for JIT code execution.
rt.cpu_features());
FuncNode* func_node = cc.add_func( // Begin the function of the following signature:
uint32_t*, // 1st argument - uint32_t* (machine reg-size).
const uint32_t*, // 2nd argument - uint32_t* (machine reg-size).
size_t>()); // 3rd argument - size_t (machine reg-size).
Label L_Loop = cc.new_label(); // Start of the loop.
Label L_Exit = cc.new_label(); // Used to exit early.
func_node->set_arg(0, dst); // Assign `dst` argument.
func_node->set_arg(1, src); // Assign `src` argument.
func_node->set_arg(2, i); // Assign `i` argument.
cc.test(i, i); // Early exit if length is zero.
cc.jz(L_Exit);
cc.bind(L_Loop); // Bind the beginning of the loop here.
cc.mov(tmp, x86::dword_ptr(src)); // Load DWORD from [src] address.
cc.mov(x86::dword_ptr(dst), tmp); // Store DWORD to [dst] address.
cc.add(src, 4); // Increment `src`.
cc.add(dst, 4); // Increment `dst`.
cc.dec(i); // Loop until `i` is non-zero.
cc.jnz(L_Loop);
cc.bind(L_Exit); // Label used by early exit.
cc.end_func(); // End of the function body.
cc.finalize(); // Translate and assemble the whole 'cc' content.
// ----> x86::Compiler is no longer needed from here and can be destroyed <----
// Add the generated code to the runtime.
MemCpy32 memcpy32;
if (err != Error::kOk) {
return 1; // Handle a possible error returned by AsmJit.
}
// ----> CodeHolder is no longer needed from here and can be destroyed <----
// Test the generated code.
uint32_t input[6] = { 1, 2, 3, 5, 8, 13 };
uint32_t output[6];
memcpy32(output, input, 6);
for (uint32_t i = 0; i < 6; i++) {
printf("%d\n", output[i]);
}
rt.release(memcpy32);
return 0;
}
AVX and AVX-512
AVX and AVX-512 code generation must be explicitly enabled via FuncFrame to work properly. If it's not setup correctly then Prolog & Epilog would use SSE instead of AVX instructions to work with SIMD registers. In addition, Compiler requires explicitly enable AVX-512 via FuncFrame in order to use all 32 SIMD registers.
#include <asmjit/x86.h>
#include <stdio.h>
using namespace asmjit;
int main() {
using Func = void (*)(void*); // Signature of the generated function.
JitRuntime rt; // Runtime specialized for JIT code execution.
rt.cpu_features());
FuncNode* func_node = cc.add_func(FuncSignature::build<void, void*>());
// Use the following to enable AVX and/or AVX-512.
func_node->frame().set_avx_enabled();
func_node->frame().set_avx512_enabled();
// Do something with the input pointer.
x86::Gp addr = cc.new_gp_ptr("addr");
x86::Vec vreg = cc.new_zmm("vreg");
func_node->set_arg(0, addr);
cc.vmovdqu32(vreg, x86::ptr(addr));
cc.vpaddq(vreg, vreg, vreg);
cc.vmovdqu32(x86::ptr(addr), vreg);
cc.end_func(); // End of the function body.
cc.finalize(); // Translate and assemble the whole 'cc' content.
// ----> x86::Compiler is no longer needed from here and can be destroyed <----
Func fn;
if (err != Error::kOk) {
return 1; // Handle a possible error returned by AsmJit.
}
// ----> CodeHolder is no longer needed from here and can be destroyed <----
// Execute the generated code and print some output.
uint64_t data[] = { 1, 2, 3, 4, 5, 6, 7, 8 };
fn(data);
printf("%llu\n", (unsigned long long)data[0]);
rt.release(fn); // Explicitly remove the function from the runtime.
return 0;
}
Recursive Functions
It's possible to create more functions by using the same x86::Compiler instance and make links between them. In such case it's important to keep the pointer to FuncNode.
The example below creates a simple Fibonacci function that calls itself recursively:
#include <asmjit/x86.h>
#include <stdio.h>
using namespace asmjit;
int main() {
using FibFn = uint32_t (*)(uint32_t x); // Signature of the generated function.
JitRuntime rt; // Runtime specialized for JIT code execution.
rt.cpu_features());
FuncNode* func_node = cc.add_func( // Begin of the Fibonacci function, add_func()
FuncSignature::build<int, int>()); // Returns a pointer to the FuncNode node.
Label L_Exit = cc.new_label(); // Exit label.
x86::Gp x = cc.new_gp32(); // Function x argument.
x86::Gp y = cc.new_gp32(); // Temporary.
func_node->set_arg(0, x);
cc.cmp(x, 3); // Return x if less than 3.
cc.jb(L_Exit);
cc.mov(y, x); // Make copy of the original x.
cc.dec(x); // Decrease x.
InvokeNode* invoke_node; // Function invocation:
cc.invoke(Out(invoke_node), // - InvokeNode (output).
func_node->label(), // - Function address or Label.
FuncSignature::build<int, int>()); // - Function signature.
invoke_node->set_arg(0, x); // Assign x as the first argument.
invoke_node->set_ret(0, x); // Assign x as a return value as well.
cc.add(x, y); // Combine the return value with y.
cc.bind(L_Exit);
cc.ret(x); // Return x.
cc.end_func(); // End of the function body.
cc.finalize(); // Translate and assemble the whole 'cc' content.
// ----> x86::Compiler is no longer needed from here and can be destroyed <----
FibFn fib;
if (err != Error::kOk) {
return 1; // Handle a possible error returned by AsmJit.
}
// ----> CodeHolder is no longer needed from here and can be destroyed <----
// Test the generated code.
printf("Fib(%u) -> %u\n", 8, fib(8));
rt.release(fib);
return 0;
}
Stack Management
Function's stack-frame is managed automatically, which is used by the register allocator to spill virtual registers. It also provides an interface to allocate user-defined block of the stack, which can be used as a temporary storage by the generated function. In the following example a stack of 256 bytes size is allocated, filled by bytes starting from 0 to 255 and then iterated again to sum all the values.
#include <asmjit/x86.h>
#include <stdio.h>
using namespace asmjit;
int main() {
using Func = int (*)(void); // Signature of the generated function.
JitRuntime rt; // Runtime specialized for JIT code execution.
rt.cpu_features());
cc.add_func(FuncSignature::build<int>()); // Create a function that returns int.
x86::Gp p = cc.new_gp_ptr("p");
x86::Gp i = cc.new_gp_ptr("i");
// Allocate 256 bytes on the stack aligned to 4 bytes.
x86::Mem stack = cc.new_stack(256, 4);
x86::Mem stack_idx(stack); // Copy of stack with i added.
stack_idx.set_index(i); // stack_idx <- stack[i].
stack_idx.set_size(1); // stack_idx <- byte ptr stack[i].
// Load a stack address to `p`. This step is purely optional and shows
// that `lea` is useful to load a memory operands address (even absolute)
// to a general purpose register.
cc.lea(p, stack);
// Clear i (xor is a C++ keyword, hence 'xor_' is used instead).
cc.xor_(i, i);
Label L1 = cc.new_label();
Label L2 = cc.new_label();
cc.bind(L1); // First loop, fill the stack.
cc.mov(stack_idx, i.r8()); // stack[i] = uint8_t(i).
cc.inc(i); // i++;
cc.cmp(i, 256); // if (i < 256)
cc.jb(L1); // goto L1;
// Second loop, sum all bytes stored in `stack`.
x86::Gp sum = cc.new_gp32("sum");
x86::Gp val = cc.new_gp32("val");
cc.xor_(i, i);
cc.xor_(sum, sum);
cc.bind(L2);
cc.movzx(val, stack_idx); // val = uint32_t(stack[i]);
cc.add(sum, val); // sum += val;
cc.inc(i); // i++;
cc.cmp(i, 256); // if (i < 256)
cc.jb(L2); // goto L2;
cc.ret(sum); // Return the `sum` of all values.
cc.end_func(); // End of the function body.
cc.finalize(); // Translate and assemble the whole 'cc' content.
// ----> x86::Compiler is no longer needed from here and can be destroyed <----
Func func;
if (err != Error::kOk) {
return 1; // Handle a possible error returned by AsmJit.
}
// ----> CodeHolder is no longer needed from here and can be destroyed <----
return 0;
}
Constant Pool
Compiler provides two constant pools for a general purpose code generation:
- Local constant pool - Part of FuncNode, can be only used by a single function and added after the function epilog sequence (after
retinstruction). - Global constant pool - Part of BaseCompiler, flushed at the end of the generated code by BaseEmitter::finalize().
The example below illustrates how a built-in constant pool can be used:
#include <asmjit/x86.h>
using namespace asmjit;
x86::Gp v0 = cc.new_gp32("v0");
x86::Gp v1 = cc.new_gp32("v1");
cc.mov(v0, c0);
cc.mov(v1, c1);
cc.add(v0, v1);
cc.ret(v0);
cc.end_func();
}
Jump Tables
x86::Compiler supports jmp instruction with reg/mem operand, which is a commonly used pattern to implement indirect jumps within a function, for example to implement switch() statement in a programming languages. By default AsmJit assumes that every basic block can be a possible jump target as it's unable to deduce targets from instruction's operands. This is a very pessimistic default that should be avoided if possible as it's costly and very unfriendly to liveness analysis and register allocation.
Instead of relying on such pessimistic default behavior, let's use JumpAnnotation to annotate a jump where all targets are known:
#include <asmjit/x86.h>
using namespace asmjit;
// Function arguments
x86::Vec a = cc.new_xmm_ss("a");
x86::Vec b = cc.new_xmm_ss("b");
x86::Gp op = cc.new_gp32("op");
x86::Gp target = cc.new_gp_ptr("target");
x86::Gp offset = cc.new_gp_ptr("offset");
func_node->set_arg(0, a);
func_node->set_arg(1, b);
func_node->set_arg(2, op);
// Jump annotation is a building block that allows to annotate all possible targets where `jmp()` can
// jump. It then drives the CFG construction and liveness analysis, which impacts register allocation.
JumpAnnotation* annotation = cc.new_jump_annotation();
annotation->add_label(L_Add);
annotation->add_label(L_Sub);
annotation->add_label(L_Mul);
annotation->add_label(L_Div);
// Most likely not the common indirect jump approach, but it
// doesn't really matter how final address is calculated. The
// most important path using JumpAnnotation with `jmp()`.
cc.lea(offset, x86::ptr(L_Table));
if (cc.is_64bit())
cc.movsxd(target, x86::dword_ptr(offset, op.clone_as(offset), 2));
else
cc.mov(target, x86::dword_ptr(offset, op.clone_as(offset), 2));
cc.add(target, offset);
cc.jmp(target, annotation);
// Acts like a switch() statement in C.
cc.bind(L_Add);
cc.addss(a, b);
cc.jmp(L_End);
cc.bind(L_Sub);
cc.subss(a, b);
cc.jmp(L_End);
cc.bind(L_Mul);
cc.mulss(a, b);
cc.jmp(L_End);
cc.bind(L_Div);
cc.divss(a, b);
cc.bind(L_End);
cc.ret(a);
cc.end_func();
// Relative int32_t offsets of `L_XXX - L_Table`.
cc.bind(L_Table);
cc.embed_label_delta(L_Add, L_Table, 4);
cc.embed_label_delta(L_Sub, L_Table, 4);
cc.embed_label_delta(L_Mul, L_Table, 4);
cc.embed_label_delta(L_Div, L_Table, 4);
}