当前位置:网站首页>Compiler optimization (4): inductive variables
Compiler optimization (4): inductive variables
2022-07-07 19:52:00 【openEuler】
0. Basic knowledge inventory
0.1 loop (loop)
Definition
loop(llvm It is understood as natural loop) Is defined in CFG A node set in L, And has the following properties [1][2]:
There is a single entry node ( be called header), This node governs loop All the nodes in ; There is a back edge that enters the loop head ;
Related terms
entering block: A non loop The inner node has an edge connected to loop. When there is only one entering block And only one side of it is connected to header, be called preheader; Act as not loop Nodal peheader Dominate the whole loop; latch: There is an edge connected to header; backedge: It's called back side , One from latch To header The edge of ; exiting edge: One side from loop Inward to loop Outside , The starting point of the edge is called exiting block, The target node is called exit block;
In the right picture above , The yellow area is a loop, The red area is not , Why? ?
Because the red area a and c Are all entry nodes , Does not satisfy the nature of a single entry node .
0.2 Scalar Evolution(SCEV)
Definition
SCEV It is the optimization of the compiler to analyze variables ( Often only for integer types ), It is mainly used to analyze how variables are updated in the loop , Then optimize according to this information .
Loop chain
As shown in the figure , Inductive variables in the loop var
Starting at start
, The way of iteration is ϕ
, In steps of step
;
Its circular chain (chrec,Chains of Recurrences) as follows :
var = {start, ϕ , step}
// ϕ∈{+,∗}
// start: starting value
// step: step in each iteration
for instance :
int m = 0;
for (int i = 0; i < n; i++) {
m = m + n;
*res = m;
}
that m The cycle chain of is :m = {0,+,n}
.
1. Induction Variable( Inductive variables )
1.1 Definition
Each iteration of the loop increases or decreases a fixed amount of variables , Or another linear function of inductive variables .
for instance [3], In the following cycle i and j Are inductive variables :
for (i = 0; i < 10; ++i) {
j = 17 * i;
}
1.2 benefit
Summarize the benefits of variable optimization , There are but not limited to the following points :
Replace the original calculation method with simpler instructions .
such as , Inductive variables are identified in the above example , Replace the corresponding multiplication with a less expensive addition .j = -17;
for (i = 0; i < 10; ++i) {
j = j + 17;
}Reduce the number of inductive variables , Reduce register pressure .
extern int sum;
int foo(int n) {
int i, j;
j = 5;
for (i = 0; i < n; ++i) {
j += 2;
sum += j;
}
return sum;
}Current loop There are two inductive variables :i、j, Use one of the variables to express the other post , as follows :
extern int sum;
int foo(int n) {
int i;
for (i = 0; i < n; ++i) {
sum += 5 + 2 * (i + 1);
}
return sum;
}Inductive variable substitution , Make the relationship between variables and circular indexes clear , It is convenient for other optimization analysis ( Such as dependency analysis ). Examples are as follows , take c Expressed as a function related to circular index :
int c, i;
c = 10;
for (i = 0; i < 10; i++) {
c = c + 5; // c is incremented by 5 for each loop iteration
}Convert to :
int c, i;
c = 10;
for (i = 0; i < 10; i++) {
c = 10 + 5 * (i + 1); // c is explicitly expressed as a function of loop index
}
2. practice
2.1 Related compilation options
compiler | option |
---|---|
gcc | -fivopt |
Bi Sheng | -indvars |
2.2 Optimize use cases
Optimization of inductive variables (ivs) stay llvm The position in is :llvm\lib\Transforms\Scalar\IndVarSimplify.cpp
Let's pass a use case , Take a look at the optimization process of Bisheng compiler .
Here's the picture , Suppose that func
The inner part is the code to be optimized , below func Inside is the expected result :
its IR Use cases test.ll
yes :
The compile command is :
opt test.ll -indvars -S
In the current example ,header
、latch
and exiting block
It's all the same BB, namely bb5.
Step one : basis def-use
Relationship , Traverse loop Of ExitBlock
in phi The source of the operand of the node , Calculate the final value and replace it , Then replace the phi Use of nodes .
In the example , Calculation %tmp2.lcssa
, Its only operand is %tmp2 = add nuw nsw i32 %i.01.0, 3
, Where the expression is located loop yes bb5, here %tmp2
The cycle chain of is
%tmp2 = {3,+,3}<nuw><nsw><%bb5>
Get current loop The maximum value of not exiting the loop is 199999
, Now %tmp2=add(3, mul(3,199999))=600000
; Next, we will see that the current replacement is not expensive ( The calculation of cost will vary according to different architectures ), At the same time phi Nodal user
Replace the value in . The optimization results are as follows :
Step two : Traverse ExitingBlock
, Calculate the jump condition , basis def-use
The relationship between , Delete the corresponding instruction .
In the example , To calculate the br i1 %0, label %bb5, label %bb7
Of %0
yes false
, After the jump instruction is replaced ,%0 = icmp ult i32 %tmp4,200000
non-existent user
, Add it to “ Dead order ” in . The optimization results are as follows :
Step three : Delete all “ Dead order ”, And see if his operands should be deleted .
In the example , As %0
Of operands %tmp4
And others user %x.03.0
, So it can't be regarded as “ Dead order ” Be deleted . The optimization results are as follows :
Step four : Delete HeaderBlock
Medium “ die ”phi node .
In the example , %tmp4
and phi node %x.03.0
It forms a cycle without results , Will delete them , Delete... In the same way %tmp2
and %i.01.0
. The optimization results are as follows :
Reference resources
[1] https://llvm.org/docs/LoopTerminology.html
[2] 《 Compiler principle 》 [ beautiful ]Alfred V.Aho,[ beautiful ]Monica S.Lam,[ beautiful ]Ravi Sethi Waiting , Zhao Jianhua , Translated by Zheng Tao, et al
[3] https://en.wikipedia.org/wiki/Induction_variable
Click on Read the original Start using Bisheng compiler
This article is from WeChat official account. - openEuler(openEulercommunity).
If there is any infringement , Please contact the [email protected] Delete .
Participation of this paper “OSC Source creation plan ”, You are welcome to join us , share .
边栏推荐
- Solve the problem of remote rviz error reporting
- Interpretation of transpose convolution theory (input-output size analysis)
- R语言ggplot2可视化:使用ggpubr包的ggstripchart函数可视化分组点状条带图(dot strip plot)、设置position参数配置不同分组数据点的分离程度
- [confluence] JVM memory adjustment
- What does "true" mean
- Kirin Xin'an won the bid for the new generation dispatching project of State Grid!
- State mode - Unity (finite state machine)
- PMP对工作有益吗?怎么选择靠谱平台让备考更省心省力!!!
- 杰理之快速配对,不支持取消配对【篇】
- R语言ggplot2可视化:使用ggpubr包的ggdensity函数可视化分组密度图、使用stat_overlay_normal_density函数为每个分组的密度图叠加正太分布曲线
猜你喜欢
[Verilog advanced challenge of Niuke network question brushing series] ~ multi bit MUX synchronizer
LeetCode_7_5
Netease Yunxin participated in the preparation of the standard "real time audio and video service (RTC) basic capability requirements and evaluation methods" issued by the Chinese Academy of Communica
Jerry's headphones with the same channel are not allowed to pair [article]
谷歌seo外链Backlinks研究工具推荐
【STL】vector
杰理之手动配对方式【篇】
关于ssh登录时卡顿30s左右的问题调试处理
Kunpeng developer summit 2022 | Kirin Xin'an and Kunpeng jointly build a new ecosystem of computing industry
爬虫实战(七):爬王者英雄图片
随机推荐
“本真”是什么意思
Training IX basic configuration of network services
How to buy stocks on your mobile phone and open an account? Is it safe to open an account
LC:字符串转换整数 (atoi) + 外观数列 + 最长公共前缀
吞吐量Throughout
PMP每日一练 | 考试不迷路-7.7
LeetCode 515(C#)
LeetCode 535(C#)
ASP.NET幼儿园连锁管理系统源码
一锅乱炖,npm、yarn cnpm常用命令合集
杰理之关于 TWS 配对方式配置【篇】
2022.07.04
[confluence] JVM memory adjustment
Make insurance more "safe"! Kirin Xin'an one cloud multi-core cloud desktop won the bid of China Life Insurance, helping the innovation and development of financial and insurance information technolog
华南X99平台打鸡血教程
UCloud是基础云计算服务提供商
Netease Yunxin participated in the preparation of the standard "real time audio and video service (RTC) basic capability requirements and evaluation methods" issued by the Chinese Academy of Communica
RESTAPI 版本控制策略【eolink 翻译】
what‘s the meaning of inference
位运算介绍