| GNU Compiler Collection (GCC) Internals: Register Classes |
|---|
Next: Stack and Calling, Previous: Registers, Up: Target Macros [Contents][Index]
On many machines, the numbered registers are not all equivalent. For example, certain registers may not be allowed for indexed addressing; certain registers may not be allowed in some instructions. These machine restrictions are described to the compiler using register classes.
You define a number of register classes, giving each one a name and saying which of the registers belong to it. Then you can specify register classes that are allowed as operands to particular instruction patterns.
In general, each register will belong to several classes. In fact, one
class must be named ALL_REGS and contain all the registers. Another
class must be named NO_REGS and contain no registers. Often the
union of two classes will be another class; however, this is not required.
One of the classes must be named GENERAL_REGS. There is nothing
terribly special about the name, but the operand constraint letters
‘r’ and ‘g’ specify this class. If GENERAL_REGS is
the same as ALL_REGS, just define it as a macro which expands
to ALL_REGS.
Order the classes so that if class x is contained in class y then x has a lower class number than y.
The way classes other than GENERAL_REGS are specified in operand
constraints is through machine-dependent operand constraint letters.
You can define such letters to correspond to various classes, then use
them in operand constraints.
You must define the narrowest register classes for allocatable registers, so that each class either has no subclasses, or that for some mode, the move cost between registers within the class is cheaper than moving a register in the class to or from memory (see Costs).
You should define a class for the union of two classes whenever some
instruction allows both classes. For example, if an instruction allows
either a floating point (coprocessor) register or a general register for a
certain operand, you should define a class FLOAT_OR_GENERAL_REGS
which includes both of them. Otherwise you will get suboptimal code,
or even internal compiler errors when reload cannot find a register in the
class computed via reg_class_subunion.
You must also specify certain redundant information about the register classes: for each class, which classes contain it and which ones are contained in it; for each pair of classes, the largest class contained in their union.
When a value occupying several consecutive registers is expected in a
certain class, all the registers used must belong to that class.
Therefore, register classes cannot be used to enforce a requirement for
a register pair to start with an even-numbered register. The way to
specify this requirement is with HARD_REGNO_MODE_OK.
Register classes used for input-operands of bitwise-and or shift
instructions have a special requirement: each such class must have, for
each fixed-point machine mode, a subclass whose registers can transfer that
mode to or from memory. For example, on some machines, the operations for
single-byte values (QImode) are limited to certain registers. When
this is so, each register class that is used in a bitwise-and or shift
instruction must have a subclass consisting of registers from which
single-byte values can be loaded or stored. This is so that
PREFERRED_RELOAD_CLASS can always have a possible value to return.
An enumerated type that must be defined with all the register class names
as enumerated values. NO_REGS must be first. ALL_REGS
must be the last register class, followed by one more enumerated value,
LIM_REG_CLASSES, which is not a register class but rather
tells how many classes there are.
Each register class has a number, which is the value of casting
the class name to type int. The number serves as an index
in many of the tables described below.
The number of distinct register classes, defined as follows:
#define N_REG_CLASSES (int) LIM_REG_CLASSES
An initializer containing the names of the register classes as C string constants. These names are used in writing some of the debugging dumps.
An initializer containing the contents of the register classes, as integers
which are bit masks. The nth integer specifies the contents of class
n. The way the integer mask is interpreted is that
register r is in the class if
mask & (1 << r) is 1.
When the machine has more than 32 registers, an integer does not suffice.
Then the integers are replaced by sub-initializers, braced groupings containing
several integers. Each sub-initializer must be suitable as an initializer
for the type HARD_REG_SET which is defined in hard-reg-set.h.
In this situation, the first integer in each sub-initializer corresponds to
registers 0 through 31, the second integer to registers 32 through 63, and
so on.
A C expression whose value is a register class containing hard register regno. In general there is more than one such class; choose a class which is minimal, meaning that no smaller class also contains the register.
A macro whose definition is the name of the class to which a valid base register must belong. A base register is one used in an address which is the register value plus a displacement.
This is a variation of the BASE_REG_CLASS macro which allows
the selection of a base register in a mode dependent manner. If
mode is VOIDmode then it should return the same value as
BASE_REG_CLASS.
A C expression whose value is the register class to which a valid base register must belong in order to be used in a base plus index register address. You should define this macro if base plus index addresses have different requirements than other base register uses.
A C expression whose value is the register class to which a valid
base register for a memory reference in mode mode to address
space address_space must belong. outer_code and index_code
define the context in which the base register occurs. outer_code is
the code of the immediately enclosing expression (MEM for the top level
of an address, ADDRESS for something that occurs in an
address_operand). index_code is the code of the corresponding
index expression if outer_code is PLUS; SCRATCH otherwise.
A macro whose definition is the name of the class to which a valid index register must belong. An index register is one used in an address where its value is either multiplied by a scale factor or added to another register (as well as added to a displacement).
A C expression which is nonzero if register number num is suitable for use as a base register in operand addresses.
A C expression that is just like REGNO_OK_FOR_BASE_P, except that
that expression may examine the mode of the memory reference in
mode. You should define this macro if the mode of the memory
reference affects whether a register may be used as a base register. If
you define this macro, the compiler will use it instead of
REGNO_OK_FOR_BASE_P. The mode may be VOIDmode for
addresses that appear outside a MEM, i.e., as an
address_operand.
A C expression which is nonzero if register number num is suitable for use as a base register in base plus index operand addresses, accessing memory in mode mode. It may be either a suitable hard register or a pseudo register that has been allocated such a hard register. You should define this macro if base plus index addresses have different requirements than other base register uses.
Use of this macro is deprecated; please use the more general
REGNO_MODE_CODE_OK_FOR_BASE_P.
A C expression which is nonzero if register number num is
suitable for use as a base register in operand addresses, accessing
memory in mode mode in address space address_space.
This is similar to REGNO_MODE_OK_FOR_BASE_P, except
that that expression may examine the context in which the register
appears in the memory reference. outer_code is the code of the
immediately enclosing expression (MEM if at the top level of the
address, ADDRESS for something that occurs in an
address_operand). index_code is the code of the
corresponding index expression if outer_code is PLUS;
SCRATCH otherwise. The mode may be VOIDmode for addresses
that appear outside a MEM, i.e., as an address_operand.
A C expression which is nonzero if register number num is suitable for use as an index register in operand addresses. It may be either a suitable hard register or a pseudo register that has been allocated such a hard register.
The difference between an index register and a base register is that the index register may be scaled. If an address involves the sum of two registers, neither one of them scaled, then either one may be labeled the “base” and the other the “index”; but whichever labeling is used must fit the machine’s constraints of which registers may serve in each capacity. The compiler will try both labelings, looking for one that is valid, and will reload one or both registers only if neither labeling works.
A target hook that places additional preference on the register class to use when it is necessary to rename a register in class rclass to another class, or perhaps NO_REGS, if no preferred register class is found or hook preferred_rename_class is not implemented. Sometimes returning a more restrictive class makes better code. For example, on ARM, thumb-2 instructions using LO_REGS may be smaller than instructions using GENERIC_REGS. By returning LO_REGS from preferred_rename_class, code size can be reduced.
A target hook that places additional restrictions on the register class to use when it is necessary to copy value x into a register in class rclass. The value is a register class; perhaps rclass, or perhaps another, smaller class.
The default version of this hook always returns value of rclass argument.
Sometimes returning a more restrictive class makes better code. For
example, on the 68000, when x is an integer constant that is in range
for a ‘moveq’ instruction, the value of this macro is always
DATA_REGS as long as rclass includes the data registers.
Requiring a data register guarantees that a ‘moveq’ will be used.
One case where TARGET_PREFERRED_RELOAD_CLASS must not return
rclass is if x is a legitimate constant which cannot be
loaded into some register class. By returning NO_REGS you can
force x into a memory location. For example, rs6000 can load
immediate values into general-purpose registers, but does not have an
instruction for loading an immediate value into a floating-point
register, so TARGET_PREFERRED_RELOAD_CLASS returns NO_REGS when
x is a floating-point constant. If the constant can’t be loaded
into any kind of register, code generation will be better if
TARGET_LEGITIMATE_CONSTANT_P makes the constant illegitimate instead
of using TARGET_PREFERRED_RELOAD_CLASS.
If an insn has pseudos in it after register allocation, reload will go
through the alternatives and call repeatedly TARGET_PREFERRED_RELOAD_CLASS
to find the best one. Returning NO_REGS, in this case, makes
reload add a ! in front of the constraint: the x86 back-end uses
this feature to discourage usage of 387 registers when math is done in
the SSE registers (and vice versa).
A C expression that places additional restrictions on the register class to use when it is necessary to copy value x into a register in class class. The value is a register class; perhaps class, or perhaps another, smaller class. On many machines, the following definition is safe:
#define PREFERRED_RELOAD_CLASS(X,CLASS) CLASS
Sometimes returning a more restrictive class makes better code. For
example, on the 68000, when x is an integer constant that is in range
for a ‘moveq’ instruction, the value of this macro is always
DATA_REGS as long as class includes the data registers.
Requiring a data register guarantees that a ‘moveq’ will be used.
One case where PREFERRED_RELOAD_CLASS must not return
class is if x is a legitimate constant which cannot be
loaded into some register class. By returning NO_REGS you can
force x into a memory location. For example, rs6000 can load
immediate values into general-purpose registers, but does not have an
instruction for loading an immediate value into a floating-point
register, so PREFERRED_RELOAD_CLASS returns NO_REGS when
x is a floating-point constant. If the constant cannot be loaded
into any kind of register, code generation will be better if
TARGET_LEGITIMATE_CONSTANT_P makes the constant illegitimate instead
of using TARGET_PREFERRED_RELOAD_CLASS.
If an insn has pseudos in it after register allocation, reload will go
through the alternatives and call repeatedly PREFERRED_RELOAD_CLASS
to find the best one. Returning NO_REGS, in this case, makes
reload add a ! in front of the constraint: the x86 back-end uses
this feature to discourage usage of 387 registers when math is done in
the SSE registers (and vice versa).
Like TARGET_PREFERRED_RELOAD_CLASS, but for output reloads instead of
input reloads.
The default version of this hook always returns value of rclass
argument.
You can also use TARGET_PREFERRED_OUTPUT_RELOAD_CLASS to discourage
reload from using some alternatives, like TARGET_PREFERRED_RELOAD_CLASS.
A C expression that places additional restrictions on the register class to use when it is necessary to be able to hold a value of mode mode in a reload register for which class class would ordinarily be used.
Unlike PREFERRED_RELOAD_CLASS, this macro should be used when
there are certain modes that simply cannot go in certain reload classes.
The value is a register class; perhaps class, or perhaps another, smaller class.
Don’t define this macro unless the target machine has limitations which require the macro to do something nontrivial.
Many machines have some registers that cannot be copied directly to or from memory or even from other types of registers. An example is the ‘MQ’ register, which on most machines, can only be copied to or from general registers, but not memory. Below, we shall be using the term ’intermediate register’ when a move operation cannot be performed directly, but has to be done by copying the source into the intermediate register first, and then copying the intermediate register to the destination. An intermediate register always has the same mode as source and destination. Since it holds the actual value being copied, reload might apply optimizations to re-use an intermediate register and eliding the copy from the source when it can determine that the intermediate register still holds the required value.
Another kind of secondary reload is required on some machines which allow copying all registers to and from memory, but require a scratch register for stores to some memory locations (e.g., those with symbolic address on the RT, and those with certain symbolic address on the SPARC when compiling PIC). Scratch registers need not have the same mode as the value being copied, and usually hold a different value than that being copied. Special patterns in the md file are needed to describe how the copy is performed with the help of the scratch register; these patterns also describe the number, register class(es) and mode(s) of the scratch register(s).
In some cases, both an intermediate and a scratch register are required.
For input reloads, this target hook is called with nonzero in_p, and x is an rtx that needs to be copied to a register of class reload_class in reload_mode. For output reloads, this target hook is called with zero in_p, and a register of class reload_class needs to be copied to rtx x in reload_mode.
If copying a register of reload_class from/to x requires
an intermediate register, the hook secondary_reload should
return the register class required for this intermediate register.
If no intermediate register is required, it should return NO_REGS.
If more than one intermediate register is required, describe the one
that is closest in the copy chain to the reload register.
If scratch registers are needed, you also have to describe how to perform the copy from/to the reload register to/from this closest intermediate register. Or if no intermediate register is required, but still a scratch register is needed, describe the copy from/to the reload register to/from the reload operand x.
You do this by setting sri->icode to the instruction code of a pattern
in the md file which performs the move. Operands 0 and 1 are the output
and input of this copy, respectively. Operands from operand 2 onward are
for scratch operands. These scratch operands must have a mode, and a
single-register-class
output constraint.
When an intermediate register is used, the secondary_reload
hook will be called again to determine how to copy the intermediate
register to/from the reload operand x, so your hook must also
have code to handle the register class of the intermediate operand.
x might be a pseudo-register or a subreg of a
pseudo-register, which could either be in a hard register or in memory.
Use true_regnum to find out; it will return -1 if the pseudo is
in memory and the hard register number if it is in a register.
Scratch operands in memory (constraint "=m" / "=&m") are
currently not supported. For the time being, you will have to continue
to use SECONDARY_MEMORY_NEEDED for that purpose.
copy_cost also uses this target hook to find out how values are
copied. If you want it to include some extra cost for the need to allocate
(a) scratch register(s), set sri->extra_cost to the additional cost.
Or if two dependent moves are supposed to have a lower cost than the sum
of the individual moves due to expected fortuitous scheduling and/or special
forwarding logic, you can set sri->extra_cost to a negative amount.
These macros are obsolete, new ports should use the target hook
TARGET_SECONDARY_RELOAD instead.
These are obsolete macros, replaced by the TARGET_SECONDARY_RELOAD
target hook. Older ports still define these macros to indicate to the
reload phase that it may
need to allocate at least one register for a reload in addition to the
register to contain the data. Specifically, if copying x to a
register class in mode requires an intermediate register,
you were supposed to define SECONDARY_INPUT_RELOAD_CLASS to return the
largest register class all of whose registers can be used as
intermediate registers or scratch registers.
If copying a register class in mode to x requires an
intermediate or scratch register, SECONDARY_OUTPUT_RELOAD_CLASS
was supposed to be defined be defined to return the largest register
class required. If the
requirements for input and output reloads were the same, the macro
SECONDARY_RELOAD_CLASS should have been used instead of defining both
macros identically.
The values returned by these macros are often GENERAL_REGS.
Return NO_REGS if no spare register is needed; i.e., if x
can be directly copied to or from a register of class in
mode without requiring a scratch register. Do not define this
macro if it would always return NO_REGS.
If a scratch register is required (either with or without an
intermediate register), you were supposed to define patterns for
‘reload_inm
’ or ‘reload_outm
’, as required
(see Standard Names. These patterns, which were normally
implemented with a define_expand, should be similar to the
‘movm
’ patterns, except that operand 2 is the scratch
register.
These patterns need constraints for the reload register and scratch register that contain a single register class. If the original reload register (whose class is class) can meet the constraint given in the pattern, the value returned by these macros is used for the class of the scratch register. Otherwise, two additional reload registers are required. Their classes are obtained from the constraints in the insn pattern.
x might be a pseudo-register or a subreg of a
pseudo-register, which could either be in a hard register or in memory.
Use true_regnum to find out; it will return -1 if the pseudo is
in memory and the hard register number if it is in a register.
These macros should not be used in the case where a particular class of
registers can only be copied to memory and not to another class of
registers. In that case, secondary reload registers are not needed and
would not be helpful. Instead, a stack location must be used to perform
the copy and the movm
pattern should use memory as an
intermediate storage. This case often occurs between floating-point and
general registers.
Certain machines have the property that some registers cannot be copied to some other registers without using memory. Define this macro on those machines to be a C expression that is nonzero if objects of mode m in registers of class1 can only be copied to registers of class class2 by storing a register of class1 into memory and loading that memory location into a register of class2.
Do not define this macro if its value would always be zero.
Normally when SECONDARY_MEMORY_NEEDED is defined, the compiler
allocates a stack slot for a memory location needed for register copies.
If this macro is defined, the compiler instead uses the memory location
defined by this macro.
Do not define this macro if you do not define
SECONDARY_MEMORY_NEEDED.
When the compiler needs a secondary memory location to copy between two
registers of mode mode, it normally allocates sufficient memory to
hold a quantity of BITS_PER_WORD bits and performs the store and
load operations in a mode that many bits wide and whose class is the
same as that of mode.
This is right thing to do on most machines because it ensures that all bits of the register are copied and prevents accesses to the registers in a narrower mode, which some machines prohibit for floating-point registers.
However, this default behavior is not correct on some machines, such as the DEC Alpha, that store short integers in floating-point registers differently than in integer registers. On those machines, the default widening will not work correctly and you must define this macro to suppress that widening in some cases. See the file alpha.h for details.
Do not define this macro if you do not define
SECONDARY_MEMORY_NEEDED or if widening mode to a mode that
is BITS_PER_WORD bits wide is correct for your machine.
A target hook which returns true if pseudos that have been assigned
to registers of class rclass would likely be spilled because
registers of rclass are needed for spill registers.
The default version of this target hook returns true if rclass
has exactly one register and false otherwise. On most machines, this
default should be used. For generally register-starved machines, such as
i386, or machines with right register constraints, such as SH, this hook
can be used to avoid excessive spilling.
This hook is also used by some of the global intra-procedural code transformations to throtle code motion, to avoid increasing register pressure.
A target hook returns the maximum number of consecutive registers of class rclass needed to hold a value of mode mode.
This is closely related to the macro HARD_REGNO_NREGS. In fact,
the value returned by TARGET_CLASS_MAX_NREGS (rclass,
mode) target hook should be the maximum value of
HARD_REGNO_NREGS (regno, mode) for all regno
values in the class rclass.
This target hook helps control the handling of multiple-word values in the reload pass.
The default version of this target hook returns the size of mode in words.
A C expression for the maximum number of consecutive registers of class class needed to hold a value of mode mode.
This is closely related to the macro HARD_REGNO_NREGS. In fact,
the value of the macro CLASS_MAX_NREGS (class, mode)
should be the maximum value of HARD_REGNO_NREGS (regno,
mode) for all regno values in the class class.
This macro helps control the handling of multiple-word values in the reload pass.
If defined, a C expression that returns nonzero for a class for which a change from mode from to mode to is invalid.
For example, loading 32-bit integer or floating-point objects into
floating-point registers on Alpha extends them to 64 bits.
Therefore loading a 64-bit object and then storing it as a 32-bit object
does not store the low-order 32 bits, as would be the case for a normal
register. Therefore, alpha.h defines CANNOT_CHANGE_MODE_CLASS
as below:
#define CANNOT_CHANGE_MODE_CLASS(FROM, TO, CLASS) \ (GET_MODE_SIZE (FROM) != GET_MODE_SIZE (TO) \ ? reg_classes_intersect_p (FLOAT_REGS, (CLASS)) : 0)
Even if storing from a register in mode to would be valid,
if both from and raw_reg_mode for class are wider
than word_mode, then we must prevent to narrowing the
mode. This happens when the middle-end assumes that it can load
or store pieces of an N-word pseudo, and that the pseudo will
eventually be allocated to N word_mode hard registers.
Failure to prevent this kind of mode change will result in the
entire raw_reg_mode being modified instead of the partial
value that the middle-end intended.
A target hook which can change allocno class for given pseudo from allocno and best class calculated by IRA.
The default version of this target hook always returns given class.
A target hook which returns true if we use LRA instead of reload pass. The default version of this target hook returns true. New ports should use LRA, and existing ports are encouraged to convert.
A target hook which returns the register priority number to which the register hard_regno belongs to. The bigger the number, the more preferable the hard register usage (when all other conditions are the same). This hook can be used to prefer some hard register over others in LRA. For example, some x86-64 register usage needs additional prefix which makes instructions longer. The hook can return lower priority number for such registers make them less favorable and as result making the generated code smaller. The default version of this target hook returns always zero.
A target hook which returns true if we need register usage leveling. That means if a few hard registers are equally good for the assignment, we choose the least used hard register. The register usage leveling may be profitable for some targets. Don’t use the usage leveling for targets with conditional execution or targets with big register files as it hurts if-conversion and cross-jumping optimizations. The default version of this target hook returns always false.
A target hook which returns true if an address with the same structure can have different maximal legitimate displacement. For example, the displacement can depend on memory mode or on operand combinations in the insn. The default version of this target hook returns always false.
A target hook which returns true if subst can’t
substitute safely pseudos with equivalent memory values during
register allocation.
The default version of this target hook returns false.
On most machines, this default should be used. For generally
machines with non orthogonal register usage for addressing, such
as SH, this hook can be used to avoid excessive spilling.
A target hook which returns true if *disp is
legitimezed to valid address displacement with subtracting *offset
at memory mode mode.
The default version of this target hook returns false.
This hook will benefit machines with limited base plus displacement
addressing.
This hook defines a class of registers which could be used for spilling pseudos of the given mode and class, or NO_REGS if only memory should be used. Not defining this hook is equivalent to returning NO_REGS for all inputs.
This hook should return true if given class of registers should be an allocno class in any way. Usually RA uses only one register class from all classes containing the same register set. In some complicated cases, you need to have two or more such classes as allocno ones for RA correct work. Not defining this hook is equivalent to returning false for all inputs.
This hook defines the machine mode to use for the boolean result of conditional store patterns. The ICODE argument is the instruction code for the cstore being performed. Not definiting this hook is the same as accepting the mode encoded into operand 0 of the cstore expander patterns.
A target hook which lets a backend compute the set of pressure classes to be used by those optimization passes which take register pressure into account, as opposed to letting IRA compute them. It returns the number of register classes stored in the array pressure_classes.
Next: Stack and Calling, Previous: Registers, Up: Target Macros [Contents][Index]