The lcc 4.1 Code-Generation Interface

Christopher W. Fraser and David R. Hanson, Microsoft Research

Contents

Introduction

Version 4.1 is the latest release of lcc, the ANSI C compiler described in our book A Retargetable C Compiler: Design and Implementation (Addison-Wesley, 1995, ISBN 0-8053-1670-1). This document summarizes the differences between the 4.1 code-generation interface and the 3.x interface described in Chap. 5 of A Retargetable C Compiler.

Previous versions of lcc supported only three sizes of integers, two sizes of floats, and insisted that pointers fit in unsigned integers (see Sec. 5.1 of A Retargetable C Compiler). These assumptions simplified the compiler, and were suitable for 32-bit architectures. But on 64-bit architectures, such as the DEC ALPHA, it's natural to have four sizes of integers and perhaps three sizes of floats, and on 16-bit architectures, 32-bit pointers don't fit in unsigned integers. Also, the 3.x constaints limited the use of lcc's back ends for other languages, such as Java.

Version 4.x removes all of these restrictions: It supports any number of sizes for integers and floats, and the size of pointers need not be related to the size of any of the integer types. The major changes in the code-generation interface are:

• The number of type suffixes has been reduced to 6.
• Dag operators are composed of a generic operator, a type suffix, and a size.
• Unsigned variants of several operators have been added.
• Several interface functions have new signatures.

In addition, version 4.x is written in ANSI C and uses the standard I/O library and other standard C functions.

The sections below parallel the subsections of Chap. 5 of A Retargetable C Compiler and summarize the differences between the 3.x and 4.x code-generation interface. Unaffected subsections are omitted. Page citations refer to pages in A Retargetable C Compiler.

5.1 Type Metrics

There are now 10 metrics in an interface record:

Metrics charmetric;
Metrics shortmetric;
Metrics intmetric;
Metrics longmetric;
Metrics longlongmetric;
Metrics floatmetric;
Metrics doublemetric;
Metrics longdoublemetric;
Metrics ptrmetric;
Metrics structmetric;

Each of these specifies the size and alignment of the corresponding type. ptrmetric describes all pointers.

5.3 Symbols

The actual value of a constant is stored in the u.c.v field of a symbol, which holds a Value:

typedef union value {
	long i;
	unsigned long u;
	long double d;
	void *p;
	void (*g)(void);
} Value;

The value is stored in the appropriate field according to its type, which is given by the symbol's type field.

5.5 Dag Operators

The op field a of node structure holds a dag operator, which consists of a generic operator, a type suffix, and a size indicator. The type suffixes are:

enum {
	F=FLOAT,
	I=INT,
	U=UNSIGNED,
	P=POINTER,
	V=VOID,
	B=STRUCT
};

#define sizeop(n) ((n)<<10)

Given a generic operator o, a type suffix t, and a size s, a type- and size-specific operator is formed by o+t+sizeop(s). For example, ADD+F+sizeop(4) forms the operator ADDF4, which denotes the sum of two 4-byte floats. Similarly, ADD+F+sizeop(8) forms ADDF8, which denotes 8-byte floating addition. In the 3.x code-generation interface, ADDF and ADDD denoted these operations. There was no size indicator in the 3.x operators because the type suffix supplied both a type and a size.

Table 5.1 lists each generic operator, its valid type suffixes, and the number of kids and syms that it uses; multiple values for kids indicate type-specific variants. The notations in the syms column give the number of syms values and a one-letter code that suggests their uses: 1V indicates that syms[0] points to a symbol for a variable, 1C indicates that syms[0] is a constant, and 1L indicates that syms[0] is a label. For 1S, syms[0] is a constant whose value is a size in bytes; 2S adds syms[1], which is a constant whose value is an alignment. For most operators, the type suffix and size indicator denote the type and size of operation to perform and the type and size of the result.

Table 5.1Node Operators.
syms	kids	Operator	Type Suffixes	Sizes	Operation
1V	0	ADDRF	...P..	p	address of a parameter
1V	0	ADDRG	...P..	p	address of a global
1V	0	ADDRL	...P..	p	address of a local
1C	0	CNST	FIUP..	fdx csilh p	constant
	1	BCOM	.IU...	ilh	bitwise complement
1S	1	CVF	FI....	fdx ilh	convert from float
1S	1	CVI	FIU...	fdx csilh csilhp	convert from signed integer
1S	1	CVP	..U..	p	convert from pointer
1S	1	CVU	.IUP..	csilh p	convert from unsigned integer
	1	INDIR	FIUP.B	fdx csilh p	fetch
	1	NEG	FI....	fdx ilh	negation
	2	ADD	FIUP..	fdx ilh ilhp p	addition
	2	BAND	.IU...	ilh	bitwise AND
	2	BOR	.IU...	ilh	bitwise inclusive OR
	2	BXOR	.IU...	ilh	bitwise exclusive OR
	2	DIV	FIU...	fdx ilh	division
	2	LSH	.IU...	ilh	left shift
	2	MOD	.IU...	ilh	modulus
	2	MUL	FIU...	fdx ilh	multiplication
	2	RSH	.IU...	ilh	right shift
	2	SUB	FIUP..	fdx ilh ilhp p	subtraction
2S	2	ASGN	FIUP.B	fdx csilh p	assignment
1L	2	EQ	FIU...	fdx ilh ilhp	jump if equal
1L	2	GE	FIU...	fdx ilh ilhp	jump if greater than or equal
1L	2	GT	FIU...	fdx ilh ilhp	jump if greater than
1L	2	LE	FIU...	fdx ilh ilhp	jump if less than or equal
1L	2	LT	FIU...	fdx ilh ilhp	jump if less than
1L	2	NE	FIU...	fdx ilh ilhp	jump if not equal
2S	1	ARG	FIUP.B	fdx ilh p	argument
1	1 or 2	CALL	FIUPVB	fdx ilh p	function call
	1	RET	FIUPV.	fdx ilh p	return from function
	1	JUMP	....V.		unconditional jump
1L	0	LABEL	....V.		label definition

The entries in the Sizes column indicate sizes of the operators that back ends must implement. Letters denote the size of float (f), double (d), long double (x), character (c), short integer (s), integer (i), long integer (l), "long long" integer (h) , and pointer (p). These sizes are separated into sets for each type suffix, except that a single set is used for both I and U when the set for I is identical to the set for U.

The actual values for the size indicators, fdxcsilhp, depend on the target. A specification like ADDFf denotes the operator ADD+F+sizeop(f), where "f" is replaced by a target-dependent value, e.g., ADDF4 and ADDF8. For example, back ends must implement the following CVI and MUL operators.

CVIFf CVIFd CVIFx
CVIIc CVIIs CVIIi CVIIl CVIIh
CVIUc CVIUs CVIUi CVIUl CVIUh CVIUp

MULFf MULFd MULFx
MULIi MULIl MULIh
MULUi MULUl MULUh

On most platforms, there are fewer than three sizes of floats and six sizes of integers, and pointers are usually the same size as one of the integers. And lcc doesn't support the "long long" type, so h is not currently used. So the set of platform-specific operators is usually smaller than the list above suggests. For example, the X86, SPARC, and MIPS back ends implement the following CVI and MUL operators.

CVIF4 CVIF8
CVII1 CVII2 CVII4
CVIU1 CVIU2 CVIU4

MULF4 MULF8
MULI4
MULU4

The set of operators is thus target-dependent; for example, ADDI8 appears only if the target supports an 8-byte integer type. ops.c is a program that, given a set of sizes, prints the required operators and their values, e.g.,

% ops c=1 s=2 i=4 l=4 h=4 f=4 d=8 x=8 p=4
...
 CVIF4=4225 CVIF8=8321
 CVII1=1157 CVII2=2181 CVII4=4229
 CVIU1=1158 CVIU2=2182 CVIU4=4230
...
 MULF4=4561 MULF8=8657
 MULI4=4565
 MULU4=4566
...
131 operators

The type suffix for a conversion operator denotes the type of the result and the size indicator gives the size of the result. For example, CVUI4 converts an unsigned (U) to a 4-byte signed integer (I4). The syms[0] field points to a symbol-table entry for a integer constant that gives the size of the source operand. For example, if syms[0] in a CVUI4 points to a symbol-table entry for 2, the conversion widens a 2-byte unsigned integer to a 4-byte signed integer. Conversions that widen unsigned integers zero-extend; those that widen signed integers sign-extend.

The front end composes conversions between types T₁ and T₂ by widening T₁ to its "supertype", if necessary, converting that result to T₂'s supertype, then narrowing the result to T₂, if necessary. The following table lists the supertypes; omitted entries are their own supertypes.

Type		Supertype
signed char		int
signed short		int
unsigned char		int, if sizeof (char) < sizeof (int) unsigned, otherwise
unsigned short		int, if sizeof (short) < sizeof (int) unsigned, otherwise
void *		an unsigned type as large as a pointer

Pointers are converted to an unsigned type of the same size, even when that type is not one of the integer types.

For example, the front end converts a signed short to a float by first converting it to an int and then to a float. It converts an unsigned short to an int with a single CVUIi conversion, when shorts are smaller than ints.

There are now signed and unsigned variants of ASGN, INDIR, BCOM, BOR, BXOR, BAND, ARG, CALL, and RET to simplify code generation on platforms that use different instructions or register set for signed and unsigned operations. Likewise there are now pointer variants of ASGN, INDIR, ARG, CALL, and RET.

5.6 Interface Flags

unsigned unsigned_char:1;

tells the front end whether plain characters are signed or unsigned. If it's zero, char is a signed type; otherwise, char is an unsigned type.

All the interface flags can be set by command-line options, e.g., -Wf-unsigned_char=1 causes plain characters to be unsigned.

5.8 Definitions

The front end announces local variables by calling

void (*local)(Symbol);

It announces temporaries likewise; these have the symbol's temporary flag set, which indicates that the symbol will be used only in the next call to gen. If a temporary's u.t.cse field is nonnull, it points to the node that computes the value assigned to the temporary; see page 346.

The front end calls

void (*address)(Symbol p, Symbol q, long n);

to initialize q to a symbol that represents an address of the form x+n, where x is the address represented by p and the long integer n is positive or negative.

5.9 Constants

The interface function

void (*defconst)(int suffix, int size, Value v);

initializes constants. defconst emits directives to define a cell and initialize it to a constant value. v is the constant value, suffix identifies the type of the value, and size is the size of the value in bytes. The value of suffix indicates which field of v holds the value, as shown in the following table.

suffix		v Field		size
F		v.d		float, double, long double
I		v.i		signed char, signed short, signed int, signed long
U		v.u		unsigned char, unsigned short, unsigned int, unsigned long
P		v.p		void *

defconst must narrow v.x when size is less than sizeof v.x; e.g., to emit an unsigned char, defconst should emit (unsigned char)v.i.

5.12 Upcalls

lcc 4.x uses standard I/O and its I/O functions have been changed accordingly. lcc reads input from the standard input, emits code to the standard output, and writes diagnostics to the standard error output. It uses freopen to redirect these streams to explicit files, when necessary.

bp, outflush, and outs have been eliminated.

extern void fprint(FILE *f, const char *fmt, ...);
extern void  print(const char *fmt, ...);

print formatted data to file f (fprint) or the standard output (print). These functions are like standard C's printf and fprintf, but support only some of the standard conversion specifiers and do not support flags, precision, and field-width specifications. They support the following new conversion specifiers in addition to those described on page 99.

Specifiers		Corresponding printf Specifiers
%c		%c
%d %D		%d %ld
%u %U		%u %lu
%x %X		%x %lx
%f %e %g		%e %f %g
%p		Converts the corresponding void * argument to unsigned long and prints it with the printf %#x specifier or just %x when the argument is null.
%I		Prints the number of spaces given by the corresponding argument.

#define generic(op)  ((op)&0x3F0)
#define specific(op) ((op)&0x3FF)

generic(op) returns the generic variant of op; that is, without its type suffix and size indicator. specific(op) returns the type-specific variant of op; that is, without its size indicator.

newconst has been replaced by

extern Symbol intconst(int n);

which installs the integer constant n in the symbol table, if necessary, and returns a pointer to the symbol-table entry.