Open64 Compiler: WHIRL 中间表示形式与 SYMTAB

需积分: 9 124 浏览量更新于2024-07-21 收藏 584KB PDF 举报

"Open64 Compiler的WHIRL（WHirl Intermediate Representation）文档，主要涉及WHIRL抽象语法树、编译目标、WHIRL的层次结构以及各种语言元素的详细描述，包括操作符、数据类型、控制流语句、调用、内存访问等。" Open64是一个开源的优化编译器，它使用WHIRL作为其中间表示（Intermediate Representation, IR）。WHIRL是一种静态单赋值形式（Static Single Assignment, SSA）的中间语言，用于编译过程中的分析和优化。以下是WHIRL的一些核心概念和特性： 1. **Whirl抽象语法树**：WHIRL以树形结构表示源代码，每个节点代表一个编程语言的语义单元，如操作、变量、类型等。 2. **编译目标**：Open64编译器设计的目标是支持多种平台和架构，WHIRL作为通用的IR，使得编译器能够生成针对不同目标机器的高效代码。 3. **WHIRL的层次结构**： - **Very High (VH) WHIRL**：最接近源代码的级别，保留了源代码的大部分结构。 - **High (H) WHIRL**：进一步抽象，但仍包含高级语言特性。 - **Mid (M) WHIRL**：更接近底层，但仍然保留一些高级结构。 - **Low (L) WHIRL**：更接近机器码，包含更多与特定架构相关的细节。 - **Very Low (VL) WHIRL**：最底层，几乎完全与机器指令对应。 4. **WHIRL的组件**： - **Operators**：表示各种操作，如算术、逻辑、比较等。 - **Result and Descriptor Types**：定义结果类型和描述符类型。 - **Supported Data Types**：包括基本类型和复杂类型。 - **Kid Pointers**：指向子节点，构建树结构。 - **Next and Previous Pointers**：用于链表结构，方便遍历。 - **Offset**：节点在内存中的偏移量。 - **Mapping Mechanism**：用于映射源代码位置到WHIRL节点。 - **Source Position Information**：记录源代码的位置信息。 - **Additional Fields**：其他自定义或特定用途的字段。 - **WHIRL Node Layout**：节点布局，描述节点如何在内存中存储。 5. **控制流语句**：包括结构化和非结构化的控制流，如循环、条件分支、异常处理等。 6. **调用**：表示函数调用，包括参数传递、返回值处理等。 7. **其他语句**：如声明、初始化、赋值等。 8. **内存访问**：表示对内存的读写操作，包括数组访问和指针操作。 9. **位域表示**：用于处理位级别的操作。 10. **伪寄存器**：在优化阶段使用的临时存储区域，不直接对应硬件寄存器。 11. **其他叶节点操作**：如常量、空值等简单操作。 12. **类型转换**：在不同数据类型间进行转换。 13. **高阶类型规格**：用于描述复杂类型，如结构体、联合体等。 14. **表达式操作**：包括一元、二元、三元及N元操作，覆盖了各种运算。这些内容构成了WHIRL的骨架，为Open64编译器提供了强大的中间表示能力，使得编译器可以进行复杂的优化，如死代码消除、循环展开、并行性挖掘等，最终生成高效的机器代码。

1.6. OTHER CONTROL FLOW STATEMENTS 13

black boxes while working on the current compilation unit. REGION nodes cannot contain references

and deﬁnitions of pseudo-registers that are live-in or live-out with respect to the node. Values can be

passed in and out of the black boxes via dedicated registers at the region boundaries. The WHIRL level

of a region must be lower than or equal to the level of its enclosing region. A compilation component

may choose to ignore a region boundary at which the WHIRL level does not change, in which case

it will optimize the code of the region together with the enclosing region. In general, nested regions

should be compiled inside-out. An additional use of this node is to specify a region to be parallelized. In

the course of compilation, a segment of code to be parallelized is ﬁrst marked as a parallel region. The

lowering process will spawn oﬀ the region as a nested procedure that will be called via synchronization

routines during parallel execution.

• DO LOOP [VH–H]

This operator has the semantics of a Fortran Do loop. Kid 0 is an IDNAME represe nting the index

variable, which must be of type integer. Kid 1 must be an STID statement initializing the index variable,

which must not be null. Kid 2 is a comparison e xpression for the end condition. The com parison must

use GE, GT, LE or LT, and any content other than the induction variable in this expression must

be loop invariant. Kid 3 must be an STID statement that increments the index variable via an ADD

by a step amount. The step must be an expression that is loop invariant. Kid 4 is a BLOCK node

representing the body of the do loop. If Kid 5 is present, it must be a LOOP INFO that gives additional

information about the loop.

• DO WHILE [VH–H]

A do-while loop. Kid 0 is a boolean expression. Kid 1 is a BLOCK node representing the block of

statements that is executed while kid 0 returns nonzero. The condition is tested at the end of the loop,

so the block is executed at least once.

• WHILE DO [VH–H]

A while loop. Kid 0 is a boolean expression. Kid 1 is a BLOCK representing the block of statements

that is executed while Kid 0 returns non-zero. The condition is tested at the start of the loop.

• IF [VH–H]

This operator represents a structured logical if statement. Kid 0 is an expression, and both kids 1 and

2 must b e BLOCKs. Kid 1 gives a block of statements that is executed if Kid 0 evaluates to some

non-zero value. Kid 2 gives another block of s tatements that is executed if Kid 0 evaluates to zero. If

this s tateme nt has no else part, the block for Kid 2 has an empty statement list. The ﬂags ﬁeld is used

to provide c ompilation-related information for this node.

DO LOOP, DO WHILE, WHILE DO and IF represent only well-formed high-level control constructs.

The blocks associated with them cannot be the target of jumps from outside. To make it easier for the

front-ends, we do tolerate illegal high-level control constructs in the front-ends’ output. Such illegal

high-level control constructs will be screened out and converted to use ordinary control ﬂow constructs

by the ﬁrst optimization phase.

1.6 Other Control Flow Statements

This section describes the remaining control ﬂow statements in WHIRL, which are not hierarchical. They

are allowed at all levels of WHIRL. All of these operators use the prev and next ﬁelds.

• GOTO [VH–VL]

An unconditional branch to the label in the current procedure as given by label number.

• GOTO OUTER BLOCK [VH–VL]

An unconditional branch from a nested procedure to the label in a parent proc edure as given by

label number. It involves unwinding of the procedure call stack.

14 CHAPTER 1. WHIRL ABSTRACT SYNTAX TREE

• SWITCH [VH]

A switch statement in a form close to the source code. An internal ﬁeld, number entries, gives the

number of cases in the jump table. Another ﬁeld, last label, gives the label that marks the end of the

code compiled from the switch statement in the source program. Kid 0 is the switch expression, which

must be of type integer. Kid 1 is a BLOCK that deﬁnes the jump table by a list of CASEGOTOs, the

number of which equals number entries. Kid 2 is a GOTO giving the default jump target. If there is

no default target (i.e. the front-end guarantees that a match case can be found), then Kid 2 does not

exist. This statement will be lowered to the control ﬂow constructs that most eﬃciently implement

the switch.

• CASEGOTO [VH]

This is used only within a SWITCH to specify jump targets for individual case values. The const val

ﬁeld gives the integer case value. The label number ﬁeld gives the target of the jump if the switch

expression evaluates to the given case value.

• COMPGOTO [VH–M]

A non-structured computed goto statement. An internal ﬁeld, number entries, gives the number of

entries in the jump table. Another ﬁeld, last label, gives the label that marks the end of the code

compiled from the switch statement in the source program; a value of 0 means no information, and is

used in the case of a FORTRAN computed/assignedgoto, in which the jump targets are not contiguous.

Kid 0 is the switch value, and must evaluate to a 0-based integer index. Kid 1 is a BLOCK that deﬁnes

the jump table by its list of GOTO’s. The number of GOTO nodes must equal number entries. For

index value 0, the ﬁrst GOTO is executed; for the next index value, the next GOTO is executed, etc.

Kid 2 is a GOTO giving the default jump target. If there is no default target (i.e. the front-end

guarantees that the switch value is in range), then Kid 2 does not exist.

• XGOTO [L–VL]

This is formed out of lowering a COMPGOTO. st idx gives the symbol table entry of the allocated

jump table. Kid 0 is an expression that evaluates to the address to be jumped to, starting with the

base address of the allocated jump table. Kid 1 is the same as Kid 1 in COMPGOTO. Numb e r-entries

gives the number of entries in the jump table as in COMPGOTO.

There is no default jump target. The default jump target in the original COMPGOTO must be handled

by additional code ge nerated during lowering.

• AGOTO [VH–VL]

An assigned or indirect unconditional branch. The ﬂow of control is transferred to the address evaluated

by Kid 0.

• REGION EXIT [VH–VL]

This must exist within a REGION block, and speciﬁes an exit out of the region. The label number

speciﬁes the label outside the region that the ﬂow of control will transfer to. Exit out of a region can

only be eﬀected by executing this statement, and fall-through out of a region is not allowed. Other

jump s tateme nts in the region must have their targets located inside the region.

• ALTENTRY [VH–VL]

An alternate entry for the function, as translated from multiple entry subroutines in Fortran. st idx

names the entry point. Kid 0..n-1 are IDNAME leaves as in FUNC ENTRY. However, there is no

BLOCK, and control ﬂows to the next statement. The code that appears before this operator must

always ends with a GOTO to jump around the alternate entry, because the prolog code generated from

lowering ALTENTRY is not to be executed unless control is entered via the alternate entry point.

• TRUEBR [VH–VL]

A non-structured conditional branch. This node contains a label number. Kid 0 is an expression that

must evaluate to an integral value. If it evaluates to non-zero, control is transferred to the previously

mentioned label. Otherwise, control ﬂows to the next statement.

1.7. CALLS 15

• FALSEBR [VH–VL]

A non-structured conditional branch. This node contains a label number. Kid 0 is an expression that

must evaluate to an integral value. If it evaluates to zero, control is transferred to the previously

mentioned label. Otherwise, control ﬂows to the next statement.

• RETURN [VH–VL]

Return from this procedure. There can be any number of return statements in a program unit. If a

value is being returned, RETURN VAL must be used instead. All return points must be explicitly

speciﬁed via RETURN or RETURN VAL even if it is the end of the function body.

• RETURN VAL res=any [VH–H]

Return from this function with the return value speciﬁed by Kid 0. This is lowered to RETURN with

associated store statements in M WHIRL.

• LABEL [VH–VL]

Deﬁne a label. This node contains a label number. Any branch to the label will transfer control to

the statement following this one. A ﬂags ﬁeld gives attributes about the label. In particular, one

bit speciﬁes that the label marks the start of an exception handler, in which case the label has to be

treated as an entry point to the program unit. A LOOP INFO may be attached to this node as Kid

0. Otherwise, Kid 0 must be NULL.

• LOOP INFO [H–VL]

Not a statement node, but e xists as a kid of DO LOOP in H WHIRL and LABEL otherwise. It

provides information about a loop and does not translate into any executable code. In the case of

being attached to a LABEL, it sp e ciﬁes the labe l as marking the start of the loop body, and the actual

extent of the loop can be determined by ﬁnding all the basic blocks dominated by the label up to a

branch back to that same label. The trip est ﬁeld is a 16-bit ﬁeld that gives the estimated trip count

of the loop; if it is larger than 16-bits, it should be represented as a large 16-bit number instead of

being truncated; if 0, the information is not provided. The depth ﬁeld gives the loop nesting depth of

the content of the loop. The ﬂags ﬁeld provides various information about the loop, like innermost,

loop wind-down, etc. Kid 0 must be an LDID that gives the induction variable of the loop. If Kid 0

is NULL, the loop has no induction variable. Kid 1 is an expression that evaluates to the exact trip

count of the loop. If Kid 1 is NULL, the exact trip count cannot be speciﬁed or is unknown, as in the

case of a WHILE DO. If Kid 0 is NULL, Kid 1 must also be NULL. The trip count expression is for

information only, and does not need to be optimized, since it replicates the executable code elsewhere

that computes the trip count.

1.7 Calls

Because function calls can incur side eﬀects, they are classiﬁed as statements rather than expression tree s.

Programming languages allow arbitrary nesting of function calls inside expressions. In VH WHIRL, those

nestings are preserved by allowing call statements as nodes in an expression. The lowerer to H WHIRL has

to unnest calls from expression trees in order to obey H WHIRL semantics. This also includes ﬂattening out

nested calls. Calls unnested from an expression need to be generated sequentially, and their return values

need to be stored in pseudo-registers(pregs). In VH and H WHIRL, return values from calls reside in the

special pseudo-register speciﬁed by preg -1. This conforms to C language convention, in which only a single

item can be returned, though it may be a composite item.

In lowering to M WHIRL, the actual return mechanism conforming to the target ISA and ABI is mani-

fested. The actual return mechanism may involve multiple registers speciﬁed by dedicated pregs. The actual

return mechanism may also create an implicit parameter that points to the memory block designated by the

caller for returning a struct. res in the callnode indicates the return type. Type V must be used for res if

there is no subsequent read of the return pregs. The code to read the return values in the pregs must be in

the statements immediately after the call. If there is one return value, it must be in the ﬁrst statement after

剩余74页未读，继续阅读

qq_21255579

粉丝: 0
资源: 2

Open64 Compiler: WHIRL 中间表示形式与 SYMTAB

3Planesoft Screensaver Manager注册码

WHIRL

WHIRL SSA doc

ember-cli-whirl:Ember CLI 插件包含 whirl.css

旋转：Whirl，Rust中的开源世界服务器实现

pymarl:PyMARL是WhiRL的深度多智能体强化学习框架

whirl：用于实现LL（1）解析器的仅标头的库

whirl:柯里化的，接受对象的，数据最后版本的地图，每个，过滤和减少

whirl:用于循环作业的淡化宝石。 预装了 Redis 和 HTTParty。 考虑到 Heroku 的使用情况而构建

最新资源

whirl:用于循环作业的淡化宝石。预装了 Redis 和 HTTParty。考虑到 Heroku 的使用情况而构建