ch11-llvm-ir-中间代码生成
LLVM介绍
LLVM官网：The LLVM Compiler Infrastructure Project
与gcc相比，LLVM的模块化更好
Low Level Virtual Machine：x no
获得了ACM Software System Award 2012
LLVM IR (Intermediate Representation)
经过前端之后形成了中间表示：不一定是代码，也可能是有向图之类的
生成中间表示就可以做优化了
然后针对不同平台生成不同的代码
“IR 设计的优秀与否决定着整个编译器的好坏” ——《华为方舟编译器之美》
8 章技术内容, 其中 4 章介绍 Maple IR, 另外 4 章基于 Maple IR
IR: Intermediate Representation定义
LLVM IR: 带类型的、介于高级程序设计语言与汇编语言之间
结构
- .ll文件的外层和module对应
- .ll文件本身就是一个module
- 一个module有多个函数
- 一个函数由多个基本块构成
- 一个基本块由指令构成
不开优化
clang -S -emit-llvm factorial0.c -o f0-opt0.ll
-
-S -emit-llvm：只生成ll中间代码，不生成目标文件
-
-o f0：不开优化
-
target的意思是:虽然现在只是ll中间文件，但是后面需要生成目标文件。
和目标平台相关，支持的数据类型不同x86还是ARM
-
%x是虚拟的寄存器：x是变量名：所以抄袭的时候只改变量名是没用的
-
@是全局变量， %是局部变量
-
i32：类型为int，32位
-
alloc：分配空间
-
align 4：要求按照4byte对齐
-
store：往内存存信息，
0存入%3，%3的类型是i32的指针，即i32*
%0（传入的第一个参数）的信息放入%4
%1的类型是i8**，所以%5的类型是i8***
-
调用函数call：全局函数@factorial，参数是i32，值是2
调用结果放到%6中
-
mul：相乘
-
nsw：（no signed wrap）整型数相乘，有溢出风险，nsw表示不会出现有符号的溢出的行为
-
icmp：专门做整型数的compare
比较的结果放入%8，类型是i1
-
zext：zero-extension，零拓展（高位补0），return要求i32，类型不符
Static Single Assignment (SSA)静态单赋值
SSA静态单赋值:任何变量,最多只能赋一次值，(存入%7) :
- 好处：当我使用变量的时候，我明确知道在哪定义，可以建立直接连接
- 坏处：需要很多寄存器；本例子中寄存器都是软件的虚拟的
一级优化
clang -S -emit-llvm factorial0.c -o f0-opt1.ll -O1 -g0
- 不先分配空间，而是直接用寄存器计算
分支语句
函数调用图
控制流图 Control Flow Graph (CFG)
https://en.wikipedia.org/wiki/Frances_Allen
一个函数内部的指令的控制流关系
Definition (CFG）
例子1
%2: store the return value (in different branches)
- 基本块的最后一条语句是一个branch：根据比较的结果决定是跳转到%5还是%6
- 只有最后一条可以是跳转语句：可以有或无条件跳转
- %2：把返回值在两个分支中放到同一个地方%2，最后一个基本块从%2中load出来返回值返回
如果用clang生成中间代码的时候打开优化：
-
开优化，store和load被优化，只使用寄存器
-
产生了新的问题：LLVM的难点：
- 如果是true跳转来的，应该return 1;
- 如果是false跳转来的，应该return %6。
-
LLVM中间代码静态单赋值，不能在两个分支中都重新赋值给%3，然后在label%7中返回%3。
-
于是引入phi φ/Φ指令：做选择，
- 如果是从%3跳转来的就选%6寄存器
- 如果是从%1跳转来的，就选择1，装入%8，返回%8。
- 问题没有得到解决：phi指令的实现是没有底层实现的，是需要再做一步转换的，即回退到上一个方案，这是一个虚拟的实现
例子2
- 第一段代码不符合SSA
- 第二段满足SSA
基本思想: 将 ϕ 指令转换成若干赋值指令, 上推至前驱基本块中
SSA 形式的构建与消去
《毕昇编译器原理与实践》SECTION 4.3
《编译器设计》SECTION 9.3
[compilers-papers-we-love/TOPLAS1991 Efficiently Computing Static Single Assignment Form and the Control Dependence Graph.pdf at master · courses-at-nju-by-hfwei/compilers-papers-we-love · GitHub](https://github.com/courses-at-nju-by-hfwei/compilers-papers-we-love/blob/master/ir/TOPLAS1991 Efficiently Computing Static Single Assignment Form and the Control Dependence Graph.pdf)
factorial1 (opt3)三级优化
调用图
允许分配内存，内存不是寄存器，可以用store
例子3
fctorial2.c
在实验实现中
对于if语句，生成相应的branch语句
对于for循环，生成以下的中间代码
未开优化
- %5中sle：less or equal 小于等于
- %3寄存器存着ret值
开一级优化
- slt：大于等于，sle的相反方向
开三级优化
▶ 为什么基本块的中间某条指令可以是 call 指令?
▶ 一个函数里可以包含多个含有 ret 指令的基本块吗?
▶ 如果可以, 那不就可以解决 ϕ 指令相关的问题了吗?
▶ 课程实验中具体如何生成 ϕ 指令? “课上不讲, 课后自学” 吗?
▶ 为什么是 “%0, %1” 之后是 “%3”?
如何用编程的方式生成 LLVM IR?
LLVM JAVA API使用手册：Maven Repository: org.bytedeco » llvm-platform » 15.0.3-1.5.8 (mvnrepository.com)
- 构造一个语法分析树，当我碰到一个if，用编程的方式生成LLVM IR的字典，LLVM不能帮你做这个过程
- 原生API是c++，这个是用java包装的
package llvm.factorial;
import org.bytedeco.javacpp.Pointer;
import org.bytedeco.javacpp.PointerPointer;
import org.bytedeco.llvm.LLVM.*;
import static org.bytedeco.llvm.global.LLVM.*;
public class Factorial1 {
static void factorial1opt0() {
LLVMInitializeCore(LLVMGetGlobalPassRegistry());
LLVMLinkInMCJIT();
LLVMInitializeNativeAsmPrinter();
LLVMInitializeNativeAsmParser();
LLVMInitializeNativeTarget();
LLVMContextRef context = LLVMContextCreate();
LLVMModuleRef module = LLVMModuleCreateWithNameInContext("factorial1", context);
//builder指令生成器
LLVMBuilderRef builder = LLVMCreateBuilderInContext(context);
//类型和常量
LLVMTypeRef i32Type = LLVMInt32TypeInContext(context);
LLVMTypeRef i8Type = LLVMInt8TypeInContext(context);
LLVMTypeRef i8ArrayType = LLVMPointerType(i8Type, 0);
LLVMTypeRef i8ArrayArrayType = LLVMPointerType(i8ArrayType, 0);
LLVMValueRef zero = LLVMConstInt(i32Type, 0, 0);
LLVMValueRef one = LLVMConstInt(i32Type, 1, 0);
//准备参数
PointerPointer<Pointer> mainParamTypes = new PointerPointer<>(2)
.put(0, i32Type)
.put(1, i8ArrayArrayType);
PointerPointer<Pointer> factorialParamTypes = new PointerPointer<>(1)
.put(0, i32Type);
// i32 main(i32, i8**)
LLVMTypeRef mainRetType = LLVMFunctionType(i32Type, mainParamTypes, 2, 0);
// i32 factorial(i32)
LLVMTypeRef factorialRetType = LLVMFunctionType(i32Type, factorialParamTypes, 1, 0);
// define function main and factorial
LLVMValueRef mainFunction = LLVMAddFunction(module, "main", mainRetType);
LLVMValueRef factorialFunction = LLVMAddFunction(module, "factorial", factorialRetType);
// append code at main function
//基本块
LLVMBasicBlockRef mainEntry = LLVMAppendBasicBlock(mainFunction, "");
LLVMPositionBuilderAtEnd(builder, mainEntry);
// get function param
LLVMValueRef arg1 = LLVMGetParam(mainFunction, 0);
LLVMValueRef arg2 = LLVMGetParam(mainFunction, 1);
// alloca and store param
LLVMValueRef alloca1 = LLVMBuildAlloca(builder, i32Type, "");
LLVMValueRef alloca2 = LLVMBuildAlloca(builder, i32Type, "");
LLVMValueRef alloca3 = LLVMBuildAlloca(builder, i8ArrayArrayType, "");
LLVMBuildStore(builder, zero, alloca1);
LLVMBuildStore(builder, arg1, alloca2);
LLVMBuildStore(builder, arg2, alloca3);
// call factorial(2)
PointerPointer<Pointer> argsv1 = new PointerPointer<>(1)
.put(0, LLVMConstInt(i32Type, 2, 0));
LLVMValueRef call1 = LLVMBuildCall(builder, factorialFunction, argsv1, 1, "");
// factorial(2) * 7
LLVMValueRef mul1 = LLVMBuildMul(builder, call1, LLVMConstInt(i32Type, 7, 0), "");
// factorial(2) * 7 == 42
LLVMValueRef icmp1 = LLVMBuildICmp(builder, LLVMIntEQ, mul1, LLVMConstInt(i32Type, 42, 0), "");
// i1 -> i32
LLVMValueRef zext1 = LLVMBuildZExt(builder, icmp1, i32Type, "");
// return
LLVMBuildRet(builder, zext1);
// append code at factorial function
LLVMBasicBlockRef factorialEntry = LLVMAppendBasicBlock(factorialFunction, "");
LLVMPositionBuilderAtEnd(builder, factorialEntry);
// alloca i32 and store arg
LLVMValueRef arg3 = LLVMGetParam(factorialFunction, 0);
LLVMValueRef alloca4 = LLVMBuildAlloca(builder, i32Type, "");
LLVMValueRef alloca5 = LLVMBuildAlloca(builder, i32Type, "");
LLVMBuildStore(builder, arg3, alloca5);
LLVMValueRef load1 = LLVMBuildLoad(builder, alloca5, "");
// val == 0
LLVMValueRef icmp2 = LLVMBuildICmp(builder, LLVMIntEQ, load1, zero, "");
// append 3 basic block
LLVMBasicBlockRef entry1 = LLVMAppendBasicBlock(factorialFunction, "");
LLVMBasicBlockRef entry2 = LLVMAppendBasicBlock(factorialFunction, "");
LLVMBasicBlockRef entry3 = LLVMAppendBasicBlock(factorialFunction, "");
// if val == 0 goto entry1 else goto entry2
LLVMBuildCondBr(builder, icmp2, entry1, entry2);
// append code at entry1
LLVMPositionBuilderAtEnd(builder, entry1);
// store 1 to alloca4
LLVMBuildStore(builder, one, alloca4);
// goto entry3
LLVMBuildBr(builder, entry3);
// append code at entry2
LLVMPositionBuilderAtEnd(builder, entry2);
// load2, load3 = val
LLVMValueRef load2 = LLVMBuildLoad(builder, alloca5, "");
LLVMValueRef load3 = LLVMBuildLoad(builder, alloca5, "");
// val - 1
LLVMValueRef sub1 = LLVMBuildSub(builder, load3, one, "");
// call factorial(val - 1)
PointerPointer<Pointer> argsv2 = new PointerPointer<>(1)
.put(0, sub1);
LLVMValueRef call2 = LLVMBuildCall(builder, factorialFunction, argsv2, 1, "");
// val * factorial(val - 1)
LLVMValueRef mul2 = LLVMBuildMul(builder, load2, call2, "");
// store result to alloca4
LLVMBuildStore(builder, mul2, alloca4);
// goto entry3
LLVMBuildBr(builder, entry3);
// append code at entry3
LLVMPositionBuilderAtEnd(builder, entry3);
// return value at alloca4
LLVMValueRef load4 = LLVMBuildLoad(builder, alloca4, "");
LLVMBuildRet(builder, load4);
LLVMDumpModule(module);
}
static void factorial1opt1() {
LLVMInitializeCore(LLVMGetGlobalPassRegistry());
LLVMLinkInMCJIT();
LLVMInitializeNativeAsmPrinter();
LLVMInitializeNativeAsmParser();
LLVMInitializeNativeTarget();
LLVMContextRef context = LLVMContextCreate();
LLVMModuleRef module = LLVMModuleCreateWithNameInContext("factorial1", context);
LLVMBuilderRef builder = LLVMCreateBuilderInContext(context);
LLVMTypeRef i32Type = LLVMInt32TypeInContext(context);
LLVMTypeRef i8Type = LLVMInt8TypeInContext(context);
LLVMTypeRef i8ArrayType = LLVMPointerType(i8Type, 0);
LLVMTypeRef i8ArrayArrayType = LLVMPointerType(i8ArrayType, 0);
LLVMValueRef zero = LLVMConstInt(i32Type, 0, 0);
LLVMValueRef one = LLVMConstInt(i32Type, 1, 0);
PointerPointer<Pointer> mainParamTypes = new PointerPointer<>(2)
.put(0, i32Type)
.put(1, i8ArrayArrayType);
PointerPointer<Pointer> factorialParamTypes = new PointerPointer<>(1)
.put(0, i32Type);
LLVMTypeRef mainRetType = LLVMFunctionType(i32Type, mainParamTypes, 2, 0);
LLVMTypeRef factorialRetType = LLVMFunctionType(i32Type, factorialParamTypes, 1, 0);
LLVMValueRef mainFunction = LLVMAddFunction(module, "main", mainRetType);
LLVMValueRef factorialFunction = LLVMAddFunction(module, "factorial", factorialRetType);
LLVMBasicBlockRef mainEntry = LLVMAppendBasicBlock(mainFunction, "");
LLVMPositionBuilderAtEnd(builder, mainEntry);
// 相比于opt0, opt1并没有store main函数参数的中间代码，因为main函数的参数并没有用到
// call factorial(2)
PointerPointer<Pointer> argsv1 = new PointerPointer<>(1)
.put(0, LLVMConstInt(i32Type, 2, 0));
LLVMValueRef call1 = LLVMBuildCall(builder, factorialFunction, argsv1, 1, "");
// factorial(2) * 7
LLVMValueRef mul1 = LLVMBuildMul(builder, call1, LLVMConstInt(i32Type, 7, 0), "");
// factorial(2) * 7 == 42
LLVMValueRef icmp1 = LLVMBuildICmp(builder, LLVMIntEQ, mul1, LLVMConstInt(i32Type, 42, 0), "");
// i1 -> i32
LLVMValueRef zext1 = LLVMBuildZExt(builder, icmp1, i32Type, "");
// return
LLVMBuildRet(builder, zext1);
// append code at factorial
LLVMBasicBlockRef factorialEntry = LLVMAppendBasicBlock(factorialFunction, "");
LLVMPositionBuilderAtEnd(builder, factorialEntry);
// get arg(val)
LLVMValueRef arg3 = LLVMGetParam(factorialFunction, 0);
// val == 0
LLVMValueRef icmp2 = LLVMBuildICmp(builder, LLVMIntEQ, arg3, zero, "");
LLVMBasicBlockRef entry1 = LLVMAppendBasicBlock(factorialFunction, "");
LLVMBasicBlockRef entry2 = LLVMAppendBasicBlock(factorialFunction, "");
// if val == 0 goto entry2 else goto entry1
LLVMBuildCondBr(builder, icmp2, entry2, entry1);
// append code at entry1
LLVMPositionBuilderAtEnd(builder, entry1);
// val - 1
LLVMValueRef add1 = LLVMBuildAdd(builder, arg3, LLVMConstInt(i32Type, -1, 0), "");
// call factorial(val - 1)
PointerPointer<Pointer> argsv2 = new PointerPointer<>(1)
.put(0, add1);
LLVMValueRef call2 = LLVMBuildCall(builder, factorialFunction, argsv2, 1, "");
// val * factorial(val - 1)
LLVMValueRef mul2 = LLVMBuildMul(builder, call2, arg3, "");
// goto entry2
LLVMBuildBr(builder, entry2);
// append code at entry2
LLVMPositionBuilderAtEnd(builder, entry2);
// build phi: if prev block is entry1, phi1 = mul2; if prev block is factorialEntry, phi1 = 1
// more about phi: https://llvm.org/docs/LangRef.html#phi-instruction
LLVMValueRef phi1 = LLVMBuildPhi(builder, i32Type, "");
PointerPointer<Pointer> phiValues = new PointerPointer<>(2)
.put(0, mul2)
.put(1, one);
PointerPointer<Pointer> phiBlocks = new PointerPointer<>(2)
.put(0, entry1)
.put(1, factorialEntry);
LLVMAddIncoming(phi1, phiValues, phiBlocks, 2);
// return result
LLVMBuildRet(builder, phi1);
LLVMDumpModule(module);
}
public static void main(String[] args) {
factorial1opt0();
factorial1opt1();
}
}
- 如果不会写LLVM，就先写一段测试代码，然后用clang生成IR
package llvm.factorial;
import org.bytedeco.javacpp.Pointer;
import org.bytedeco.javacpp.PointerPointer;
import org.bytedeco.llvm.LLVM.*;
import static org.bytedeco.llvm.global.LLVM.*;
public class Factorial2 {
public static void main(String[] args) {
LLVMInitializeCore(LLVMGetGlobalPassRegistry());
LLVMLinkInMCJIT();
LLVMInitializeNativeAsmPrinter();
LLVMInitializeNativeAsmParser();
LLVMInitializeNativeTarget();
LLVMModuleRef module = LLVMModuleCreateWithName("factorial2");
LLVMBuilderRef builder = LLVMCreateBuilder();
LLVMTypeRef i32Type = LLVMInt32Type();
LLVMTypeRef returnType = i32Type;
LLVMValueRef one = LLVMConstInt(i32Type, 1, 0);
LLVMValueRef two = LLVMConstInt(i32Type, 2, 0);
LLVMValueRef seven = LLVMConstInt(i32Type, 7, 0);
LLVMValueRef fortyTwo = LLVMConstInt(i32Type, 42, 0);
// 创建factorial函数
LLVMTypeRef factorial2Type = LLVMFunctionType(returnType, i32Type, 1, 0);
LLVMValueRef factorial2 = LLVMAddFunction(module, "factorial2", factorial2Type);
// 构建函数体
LLVMBasicBlockRef entry = LLVMAppendBasicBlock(factorial2, "entry");
LLVMBasicBlockRef forBranch = LLVMAppendBasicBlock(factorial2, "forBranch");
LLVMBasicBlockRef trueBranch = LLVMAppendBasicBlock(factorial2, "true");
LLVMBasicBlockRef exit = LLVMAppendBasicBlock(factorial2, "exit");
// 获取参数 val
LLVMValueRef val = LLVMGetParam(factorial2, /* parameterIndex */0);
// 设置在entry块后插入代码
LLVMPositionBuilderAtEnd(builder, entry);
// int temp = 1;
LLVMValueRef temp = LLVMBuildAlloca(builder, i32Type, "temp");
LLVMBuildStore(builder, one, temp);
// int i = 2;
LLVMValueRef i = LLVMBuildAlloca(builder, i32Type, "i");
LLVMBuildStore(builder, two, temp);
// 无条件跳转到 forBranch
LLVMBuildBr(builder, forBranch);
LLVMPositionBuilderAtEnd(builder, forBranch);
// i <= val
LLVMValueRef iLoad = LLVMBuildLoad(builder, i, "iLoad");
LLVMValueRef condition = LLVMBuildICmp(builder, LLVMIntSLE, iLoad, val, "condition=i<=val");
// 条件为真，到trueBranch，否则到exit
LLVMBuildCondBr(builder, condition, trueBranch, exit);
// 设置在trueBranch块后插入代码
LLVMPositionBuilderAtEnd(builder, trueBranch);
// temp *= i;
iLoad = LLVMBuildLoad(builder, i, "iLoad");
LLVMValueRef tempLoad = LLVMBuildLoad(builder, temp, "tempLoad");
LLVMValueRef mulRes = LLVMBuildMul(builder, tempLoad, iLoad, "temp=temp*i");
LLVMBuildStore(builder, mulRes, temp);
// i++;
iLoad = LLVMBuildLoad(builder, i, "iLoad");
LLVMValueRef addRes = LLVMBuildAdd(builder, iLoad, one, "i=i+1");
LLVMBuildStore(builder, addRes, i);
// 无条件跳转到forBranch块，继续迭代
LLVMBuildBr(builder, forBranch);
// 设置在exit块后插入代码
LLVMPositionBuilderAtEnd(builder, exit);
tempLoad = LLVMBuildLoad(builder, temp, "tempLoad");
LLVMBuildRet(builder, tempLoad);
// 创建main函数
// char -> int8; 指向char的指针
LLVMTypeRef charPtr = LLVMPointerType(LLVMInt8Type(), 0);
// 指向char指针的指针
LLVMTypeRef argumentType = LLVMPointerType(charPtr, 0);
// 创建参数类型
PointerPointer<Pointer> argumentTypes = new PointerPointer<>(2)
.put(0, i32Type)
.put(1, argumentType);
// 创建函数类型
LLVMTypeRef mainType = LLVMFunctionType(returnType, argumentTypes, 2, 0);
LLVMValueRef mainFunc = LLVMAddFunction(module, "main", mainType);
// 构建函数体
LLVMBasicBlockRef mainEntry = LLVMAppendBasicBlock(mainFunc, "mainEntry");
// 设置在mainEntry块后插入代码
LLVMPositionBuilderAtEnd(builder, mainEntry);
// 构建实参
PointerPointer<Pointer> arguments = new PointerPointer<>(1)
.put(0, two);
// factorial(2)
LLVMValueRef callFactorial2 = LLVMBuildCall(builder, factorial2, arguments, 1, "factorial(2)");
// factorial(2) * 7
LLVMValueRef factorial2Mul7 = LLVMBuildMul(builder, callFactorial2, seven, "factorial(2)*7");
// factorial(2) * 7 == 42
LLVMValueRef ifEqual = LLVMBuildICmp(builder, LLVMIntEQ, factorial2Mul7, fortyTwo, "factorial(2)*7==42");
// 尤其要注意的：ifEqual是i1类型，而main函数要求返回i32类型，所以这里要做类型转换
LLVMValueRef result = LLVMBuildZExt(builder, ifEqual, i32Type, "result");
LLVMBuildRet(builder, result);
// 导出生成的LLVM IR
LLVMDumpModule(module);
//释放资源
LLVMDisposeBuilder(builder);
}
}
// factorial(2) * 7
LLVMValueRef factorial2Mul7 = LLVMBuildMul(builder, callFactorial2, seven, "factorial(2)*7");
// factorial(2) * 7 == 42
LLVMValueRef ifEqual = LLVMBuildICmp(builder, LLVMIntEQ, factorial2Mul7, fortyTwo, "factorial(2)*7==42");
// 尤其要注意的：ifEqual是i1类型，而main函数要求返回i32类型，所以这里要做类型转换
LLVMValueRef result = LLVMBuildZExt(builder, ifEqual, i32Type, "result");
LLVMBuildRet(builder, result);
// 导出生成的LLVM IR
LLVMDumpModule(module);
//释放资源
LLVMDisposeBuilder(builder);
}
}
1083
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
