#include "config.h"#include "system.h"#include "coretypes.h"#include "tm.h"#include "rtl.h"#include "tree.h"#include "gimple.h"#include "flags.h"#include "expr.h"#include "optabs.h"#include "libfuncs.h"#include "function.h"#include "regs.h"#include "diagnostic-core.h"#include "output.h"#include "tm_p.h"#include "timevar.h"#include "sbitmap.h"#include "langhooks.h"#include "target.h"#include "cgraph.h"#include "except.h"#include "dbgcnt.h"
Data Structures | |
| struct | arg_data |
Macros | |
| #define | STACK_BYTES (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT) |
Variables | |
| static char * | stack_usage_map |
| static int | highest_outgoing_arg_in_use |
| static sbitmap | stored_args_map |
| static int | stack_arg_under_construction |
| struct { | |
| rtx scan_start | |
| vec< rtx > cache | |
| } | internal_arg_pointer_exp_state |
Macro Definition Documentation
| #define STACK_BYTES (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT) |
Convert function calls to rtl insns, for GNU C compiler. Copyright (C) 1989-2013 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see http://www.gnu.org/licenses/. Like PREFERRED_STACK_BOUNDARY but in units of bytes, not bits.
Function Documentation
| bool alloca_call_p | ( | ) |
Return true when exp contains alloca call.
|
static |
If X is a likely-spilled register value, copy it to a pseudo register and return that register. Return X otherwise.
Make sure that we generate a REG rather than a CONCAT. Moves into CONCATs can need nontrivial instructions, and the whole point of this function is to avoid using the hard register directly in such a situation.
| int call_expr_flags | ( | ) |
Detect flags from a CALL_EXPR.
References emit_group_load_into_temps(), int_size_in_bytes(), arg_data::parallel_value, targetm, and TREE_TYPE.
|
static |
Scan sequence after INSN if it does not dereference any argument slots we already clobbered by tail call arguments (as noted in stored_args_map bitmap). If MARK_STORED_ARGS_MAP, add stack slots for ARG to stored_args_map bitmap afterwards (when ARG is a register MARK_STORED_ARGS_MAP should be 0). Return nonzero if sequence after INSN dereferences such argument slots, zero otherwise.
|
static |
|
static |
Scan X expression if it does not dereference any argument slots we already clobbered by tail call arguments (as noted in stored_args_map bitmap). Return nonzero if X expression dereferences such argument slots, zero otherwise.
We need not check the operands of the CALL expression itself.
Scan all subexpressions.
|
static |
We need to pop PENDING_STACK_ADJUST bytes. But, if the arguments wouldn't fill up an even multiple of PREFERRED_UNIT_STACK_BOUNDARY bytes, then we would need to push some additional bytes to pad the arguments. So, we compute an adjust to the stack pointer for an amount that will leave the stack under-aligned by UNADJUSTED_ARGS_SIZE bytes. Then, when the arguments are pushed the stack will be perfectly aligned. ARGS_SIZE->CONSTANT is set to the number of bytes that should be popped after the call. Returns the adjustment.
The number of bytes to pop so that the stack will be under-aligned by UNADJUSTED_ARGS_SIZE bytes.
The alignment of the stack after the arguments are pushed, if we just pushed the arguments without adjust the stack here.
We want to get rid of as many of the PENDING_STACK_ADJUST bytes as possible – leaving just enough left to cancel out the UNADJUSTED_ALIGNMENT. In other words, we want to ensure that the PENDING_STACK_ADJUST is non-negative, and congruent to -UNADJUSTED_ALIGNMENT modulo the PREFERRED_UNIT_STACK_BOUNDARY.
Begin by trying to pop all the bytes.
Push enough additional bytes that the stack will be aligned after the arguments are pushed.
Now, sets ARGS_SIZE->CONSTANT so that we pop the right number of bytes after the call. The right number is the entire PENDING_STACK_ADJUST less our ADJUSTMENT plus the amount required by the arguments in the first place.
|
static |
If we preallocated stack space, compute the address of each argument and store it into the ARGS array.
We need not ensure it is a valid memory address here; it will be validized when it is used.
ARGBLOCK is an rtx for the address of the outgoing arguments.
Skip this parm if it will not be passed on the stack.
Only part of the parameter is being passed on the stack.
Generate a simple memory reference of the correct size. Only part of the parameter is being passed on the stack.
Generate a simple memory reference of the correct size. Function incoming arguments may overlap with sibling call
outgoing arguments and we cannot allow reordering of reads
from function arguments with stores to outgoing arguments
of sibling calls.
|
static |
Update ARGS_SIZE to contain the total size for the argument block. Return the original constant component of the argument block's size.
REG_PARM_STACK_SPACE holds the number of bytes of stack space reserved for arguments passed in registers.
For accumulate outgoing args mode we don't need to align, since the frame will be already aligned. Align to STACK_BOUNDARY in order to prevent backends from generating misaligned frame sizes.
Compute the actual size of the argument block required. The variable and constant sizes must be combined, the size may have to be rounded, and there may be a minimum required size.
We don't handle this case yet. To handle it correctly we have
to add the delta, round and subtract the delta.
Currently no machine description requires this support. The area corresponding to register parameters is not to count in
the size of the block we need. So make the adjustment.
References convert_modes(), expand_normal(), gcc_assert, gen_lowpart_SUBREG(), GET_MODE_CLASS, arg_data::initial_value, promote_mode(), REG_P, SUBREG_PROMOTED_UNSIGNED_SET, SUBREG_PROMOTED_VAR_P, TREE_ADDRESSABLE, TREE_CODE, TREE_TYPE, type(), TYPE_MODE, arg_data::unsignedp, and arg_data::value.
|
static |
Similar to special_function_p; return a set of ERF_ flags for the function FNDECL.
|
static |
Generate instructions to call function FUNEXP, and optionally pop the results. The CALL_INSN is the first insn generated.
FNDECL is the declaration node of the function. This is given to the hook TARGET_RETURN_POPS_ARGS to determine whether this function pops its own args.
FUNTYPE is the data type of the function. This is given to the hook TARGET_RETURN_POPS_ARGS to determine whether this function pops its own args. We used to allow an identifier for library functions, but that doesn't work when the return type is an aggregate type and the calling convention says that the pointer to this aggregate is to be popped by the callee.
STACK_SIZE is the number of bytes of arguments on the stack, ROUNDED_STACK_SIZE is that number rounded up to PREFERRED_STACK_BOUNDARY; zero if the size is variable. This is both to put into the call insn and to generate explicit popping code if necessary.
STRUCT_VALUE_SIZE is the number of bytes wanted in a structure value. It is zero if this call doesn't want a structure value.
NEXT_ARG_REG is the rtx that results from executing targetm.calls.function_arg (&args_so_far, VOIDmode, void_type_node, true) just after all the args have had their registers assigned. This could be whatever you like, but normally it is the first arg-register beyond those used for args in this call, or 0 if all the arg-registers are used in this call. It is passed on to `gen_call' so you can put this info in the call insn.
VALREG is a hard register in which a value is returned, or 0 if the call does not return a value.
OLD_INHIBIT_DEFER_POP is the value that `inhibit_defer_pop' had before the args to this call were processed. We restore `inhibit_defer_pop' to that value.
CALL_FUSAGE is either empty or an EXPR_LIST of USE expressions that denote registers used by the called function.
Ensure address is valid. SYMBOL_REF is already valid, so no need, and we don't want to load it into a register as an optimization, because prepare_call_address already did it if it should be done.
Although a built-in FUNCTION_DECL and its non-__builtin
counterpart compare equal and get a shared mem_attrs, they
produce different dump output in compare-debug compilations,
if an entry gets garbage collected in one compilation, then
adds a different (but equivalent) entry, while the other
doesn't run the garbage collector at the same spot and then
shares the mem_attr with the equivalent entry. Find the call we just emitted.
Some target create a fresh MEM instead of reusing the one provided above. Set its MEM_EXPR.
Put the register usage information there.
If this is a const call, then set the insn's unchanging bit.
If this is a pure call, then set the insn's unchanging bit.
If this is a const call, then set the insn's unchanging bit.
Create a nothrow REG_EH_REGION note, if needed.
Restore this now, so that we do defer pops for this call's args if the context of the call as a whole permits.
If popup is needed, stack realign must use DRAP
For noreturn calls when not accumulating outgoing args force REG_ARGS_SIZE note to prevent crossjumping of calls with different args sizes.
If returning from the subroutine does not automatically pop the args,
we need an instruction to pop them sooner or later.
Perhaps do it now; perhaps just record how much space to pop later.
If returning from the subroutine does pop the args, indicate that the
stack pointer will be changed. Just pretend we did the pop.
When we accumulate outgoing args, we must avoid any stack manipulations. Restore the stack pointer to its original value now. Usually ACCUMULATE_OUTGOING_ARGS targets don't get here, but there are exceptions. On i386 ACCUMULATE_OUTGOING_ARGS can be enabled on demand, and popping variants of functions exist as well. ??? We may optimize similar to defer_pop above, but it is probably not worthwhile. ??? It will be worthwhile to enable combine_stack_adjustments even for such machines.
References BUILT_IN_NORMAL, builtin_decl_explicit(), DECL_BUILT_IN_CLASS, DECL_FUNCTION_CODE, and set_mem_expr().
| void emit_library_call | ( | rtx | orgfun, |
| enum libcall_type | fn_type, | ||
| enum machine_mode | outmode, | ||
| int | nargs, | ||
| ... | |||
| ) |
Output a library call to function FUN (a SYMBOL_REF rtx) (emitting the queue unless NO_QUEUE is nonzero), for a value of mode OUTMODE, with NARGS different arguments, passed as alternating rtx values and machine_modes to convert them to.
FN_TYPE should be LCT_NORMAL for `normal' calls, LCT_CONST for `const' calls, LCT_PURE for `pure' calls, or other LCT_ value for other types of library calls.
Referenced by maybe_emit_sync_lock_test_and_set().
| rtx emit_library_call_value | ( | rtx | orgfun, |
| rtx | value, | ||
| enum libcall_type | fn_type, | ||
| enum machine_mode | outmode, | ||
| int | nargs, | ||
| ... | |||
| ) |
Like emit_library_call except that an extra argument, VALUE, comes second and says where to store the result. (If VALUE is zero, this function chooses a convenient way to return the value.
This function returns an rtx for where the value is to be found. If VALUE is nonzero, VALUE is returned.
|
static |
Output a library call to function FUN (a SYMBOL_REF rtx). The RETVAL parameter specifies whether return value needs to be saved, other parameters are documented in the emit_library_call function below.
Total size in bytes of all the stack-parms scanned so far.
Size of arguments before any adjustments (such as rounding).
Todo, choose the correct decl type of orgfun. Sadly this information isn't present here, so we default to native calling abi here.
Size of the stack reserved for parameter registers.
By default, library functions can not throw.
Ensure current function's preferred stack boundary is at least what we need.
If this kind of value comes back in memory, decide where in memory it should come back.
This call returns a big structure.
??? Unfinished: must pass the memory address as an argument.
Copy all the libcall-arguments out of the varargs data and into a vector ARGVEC. Compute how to pass each argument. We only support a very small subset of the full argument passing conventions to limit complexity here since library functions shouldn't have many args.
If there's a structure value address to be passed, either pass it in the special place, or pass it as an extra argument.
Make sure it is a reasonable operand for a move or push insn.
We cannot convert the arg value to the mode the library wants here;
must do it earlier where we know the signedness of the arg. Make sure it is a reasonable operand for a move or push insn.
If this was a CONST function, it is now PURE since it now
reads memory. If this machine requires an external definition for library functions, write one out.
Since the stack pointer will never be pushed, it is possible for
the evaluation of a parm to clobber something we have already
written to the stack. Since most function calls on RISC machines
do not use the stack, this is uncommon, but must work correctly.
Therefore, we save any area of the stack that was already written
and that we are using. Here we set up to do this by making a new
stack usage map from the old one.
Another approach might be to try to reorder the argument
evaluations to avoid this conflicting stack usage. Since we will be writing into the entire argument area, the
map must be allocated for its entire size, not just the part that
is the responsibility of the caller. We must be careful to use virtual regs before they're instantiated,
and real regs afterwards. Loop optimization, for example, can create
new libcalls after we've instantiated the virtual regs, and if we
use virtuals anyway, they won't match the rtl patterns. If we push args individually in reverse order, perform stack alignment before the first push (the last arg).
Push the args that need to be pushed.
ARGNUM indexes the ARGVEC array in the order in which the arguments are to be pushed.
If this is being stored into a pre-allocated, fixed-size,
stack area, save any previous data at that location. Don't worry about things in the fixed argument area;
it has already been saved. We need to make a save area.
Now mark the segment we just used.
Indicate argument access so that alias.c knows that these
values are live. When arguments are pushed, trying to tell alias.c where
exactly this argument is won't work, because the
auto-increment causes confusion. So we merely indicate
that we access something with a known mode somewhere on
the stack. If we pushed args in forward order, perform stack alignment after pushing the last arg.
Now load any reg parms into their regs.
ARGNUM indexes the ARGVEC array in the order in which the arguments are to be pushed.
Handle calls that pass values in multiple non-contiguous
locations. The PA64 has examples of this for library calls. Any regs containing parms remain in use through the call.
Pass the function the address in which to return a structure value.
Don't allow popping to be deferred, since then cse'ing of library calls could delete a call and leave the pop.
Stack must be properly aligned now.
We pass the old value of inhibit_defer_pop + 1 to emit_call_1, which will set inhibit_defer_pop to that value.
The return type is needed to decide how many bytes the function pops. Signedness plays no role in that, so for simplicity, we pretend it's always signed. We also assume that the list of arguments passed has no impact, so we pretend it is unknown.
Right-shift returned value if necessary.
For calls to `setjmp', etc., inform function.c:setjmp_warnings that it should complain if nonvolatile values are live. For functions that cannot return, inform flow that control does not fall through.
The barrier note must be emitted
immediately after the CALL_INSN. Some ports emit more than
just a CALL_INSN above, so we must search for it here. There was no CALL_INSN?
Consider that "regular" libcalls, i.e. all of them except for LCT_THROW and LCT_RETURNS_TWICE, cannot perform non-local gotos.
There was no CALL_INSN?
Now restore inhibit_defer_pop to its actual original value.
Copy the value to the right place.
Convert to the proper mode if a promotion has been active.
If we saved any argument areas, restore them.
References args_size::constant, highest_outgoing_arg_in_use, MAX, OUTGOING_REG_PARM_STACK_SPACE, plus_constant(), STACK_POINTER_OFFSET, stack_pointer_rtx, stack_usage_map, TREE_TYPE, virtual_outgoing_args_rtx, and virtuals_instantiated.
| rtx expand_call | ( | ) |
Generate all the code for a CALL_EXPR exp and return an rtx for its value. Store the value in TARGET (specified as an rtx) if convenient. If the value is stored in TARGET then TARGET is returned. If IGNORE is nonzero, then we ignore the value of the function call.
Nonzero if we are currently expanding a call.
RTX for the function to be called.
Sequence of insns to perform a normal "call".
Sequence of insns to perform a tail "call".
Data type of the function.
Declaration of the function being called, or 0 if the function is computed (not known by name).
The type of the function being called.
Register in which non-BLKmode value will be returned, or 0 if no value or if value is BLKmode.
Address where we should return a BLKmode value; 0 if value not BLKmode.
Nonzero if that address is being passed by treating it as an extra, implicit first parameter. Otherwise, it is passed by being copied directly into struct_value_rtx.
Holds the value of implicit argument for the struct value.
Size of aggregate value wanted, or zero if none wanted or if we are using the non-reentrant PCC calling convention or expecting the value in registers.
Nonzero if called function returns an aggregate in memory PCC style, by returning the address of where to find it.
Number of actual parameters in this call, including struct value addr.
Number of named args. Args after this are anonymous ones and they must all go on the stack.
Number of complex actual arguments that need to be split.
Vector of information about each argument. Arguments are numbered in the order they will be pushed, not the order they are written.
Total size in bytes of all the stack-parms scanned so far.
Size of arguments before any adjustments (such as rounding).
Data on reg parms scanned so far.
Nonzero if a reg parm has been scanned.
Nonzero if this is an indirect function call.
Nonzero if we must avoid push-insns in the args for this call. If stack space is allocated for register parameters, but not by the caller, then it is preallocated in the fixed part of the stack frame. So the entire argument block must then be preallocated (i.e., we ignore PUSH_ROUNDING in that case).
Size of the stack reserved for parameter registers.
Address of space preallocated for stack parms (on machines that lack push insns), or 0 if space not preallocated.
Mask of ECF_ and ERF_ flags.
State variables to track stack modifications.
Some stack pointer alterations we make are performed via allocate_dynamic_stack_space. This modifies the stack_pointer_delta, which we then also need to save/restore along the way.
The alignment of the stack, in bits.
The alignment of the stack, in bytes.
The static chain value to use for this call.
See if this is "nothrow" function call.
See if we can find a DECL-node for the actual function, and get the function attributes (flags) from the function decl or type node.
Warn if this value is an aggregate type, regardless of which calling convention we are using for it.
If the result of a non looping pure or const function call is ignored (or void), and none of its arguments are volatile, we can avoid expanding the call and just evaluate the arguments for side-effects.
Set up a place to return a structure.
Cater to broken compilers.
This call returns a big structure.
For variable-sized objects, we must be called with a target
specified. If we were to allocate space on the stack here,
we would have no way of knowing when to free it. Figure out the amount to which the stack should be aligned.
Without automatic stack alignment, we can't increase preferred
stack boundary. With automatic stack alignment, it is
unnecessary since unless we can guarantee that all callers will
align the outgoing stack properly, callee has to align its
stack anyway. Operand 0 is a pointer-to-function; get the type of the function.
Count whether there are actual complex arguments that need to be split into their real and imaginary parts. Munge the type_arg_types appropriately here as well.
If struct_value_rtx is 0, it means pass the address as if it were an extra parameter. Put the argument expression in structure_value_addr_value.
If structure_value_addr is a REG other than
virtual_outgoing_args_rtx, we can use always use it. If it
is not a REG, we must always copy it into a register.
If it is virtual_outgoing_args_rtx, we must copy it to another
register in some cases. Count the arguments and set NUM_ACTUALS.
Compute number of named args. First, do a raw count of the args for INIT_CUMULATIVE_ARGS.
Count the struct value address, if it is passed as a parm.
If we know nothing, treat all args as named.
Start updating where the next arg would go. On some machines (such as the PA) indirect calls have a different calling convention than normal calls. The fourth argument in INIT_CUMULATIVE_ARGS tells the backend if this is an indirect call or not.
Now possibly adjust the number of named args. Normally, don't include the last named arg if anonymous args follow. We do include the last named arg if targetm.calls.strict_argument_naming() returns nonzero. (If no anonymous args follow, the result of list_length is actually one too large. This is harmless.) If targetm.calls.pretend_outgoing_varargs_named() returns nonzero, and targetm.calls.strict_argument_naming() returns zero, this machine will be able to place unnamed args that were passed in registers into the stack. So treat all args as named. This allows the insns emitting for a specific argument list to be independent of the function declaration. If targetm.calls.pretend_outgoing_varargs_named() returns zero, we do not have any reliable way to pass unnamed args in registers, so we must force them into memory.
Don't include the last named arg.
Treat all args as named.
Make a vector to hold all the information about each arg.
Build up entries in the ARGS array, compute the size of the arguments into ARGS_SIZE, etc.
Now make final decision about preallocating stack space.
If the structure value address will reference the stack pointer, we must stabilize it. We don't need to do this if we know that we are not going to adjust the stack pointer in processing this call.
Tail calls can make things harder to debug, and we've traditionally pushed these optimizations into -O2. Don't try if we're already expanding a call, as that means we're an argument. Don't try if there's cleanups, as we know there's code to follow the call.
Rest of purposes for tail call optimizations to fail.
Doing sibling call optimization needs some work, since
structure_value_addr can be allocated on the stack.
It does not seem worth the effort since few optimizable
sibling calls will return a structure. Check whether the target is able to optimize the call
into a sibcall. Functions that do not return exactly once may not be sibcall
optimized. If the called function is nested in the current one, it might access
some of the caller's arguments, but could clobber them beforehand if
the argument areas are shared. If this function requires more stack slots than the current
function, we cannot change it into a sibling call.
crtl->args.pretend_args_size is not part of the
stack allocated by our caller. If the callee pops its own arguments, then it must pop exactly
the same number of arguments as the current function. Check if caller and callee disagree in promotion of function return value.
Ensure current function's preferred stack boundary is at least what we need. Stack alignment may also increase preferred stack boundary.
We want to make two insn chains; one for a sibling call, the other for a normal call. We will select one of the two chains after initial RTL generation is complete.
We want to emit any pending stack adjustments before the tail
recursion "call". That way we know any adjustment after the tail
recursion call can be ignored if we indeed use the tail
call expansion. State variables we need to save and restore between
iterations. Other state variables that we must reinitialize each time
through the loop (that are not initialized by the loop itself). Start a new sequence for the normal call case.
From this point on, if the sibling call fails, we want to set
sibcall_failure instead of continuing the loop. Don't let pending stack adjusts add up to too much.
Also, do all pending adjustments now if there is any chance
this might be a call to alloca or if we are expanding a sibling
call sequence.
Also do the adjustments before a throwing call, otherwise
exception handling can fail; PR 19225. Precompute any arguments as needed.
Now we are about to start emitting insns that can be deleted
if a libcall is deleted. Compute the actual size of the argument block required. The variable
and constant sizes must be combined, the size may have to be rounded,
and there may be a minimum required size. When generating a sibcall
pattern, do not round up, since we'll be re-using whatever space our
caller provided. The argument block when performing a sibling call is the
incoming argument block. If we have no actual push instructions, or shouldn't use them,
make space for all args right now. stack_arg_under_construction says whether a stack arg is
being constructed at the old stack level. Pushing the stack
gets a clean outgoing argument block. Note that we must go through the motions of allocating an argument
block even if the size is zero because we may be storing args
in the area reserved for register arguments, which may be part of
the stack frame. Store the maximum argument space used. It will be pushed by
the prologue (if ACCUMULATE_OUTGOING_ARGS, or stack overflow
checking). Since the stack pointer will never be pushed, it is
possible for the evaluation of a parm to clobber
something we have already written to the stack.
Since most function calls on RISC machines do not use
the stack, this is uncommon, but must work correctly.
Therefore, we save any area of the stack that was already
written and that we are using. Here we set up to do this
by making a new stack usage map from the old one. The
actual save will be done by store_one_arg.
Another approach might be to try to reorder the argument
evaluations to avoid this conflicting stack usage. Since we will be writing into the entire argument area,
the map must be allocated for its entire size, not just
the part that is the responsibility of the caller. The address of the outgoing argument list must not be
copied to a register here, because argblock would be left
pointing to the wrong place after the call to
allocate_dynamic_stack_space below. Try to reuse some or all of the pending_stack_adjust
to get this space. combine_pending_stack_adjustment_and_call computes
an adjustment before the arguments are allocated.
Account for them and see whether or not the stack
needs to go up or down. We're releasing stack space.
??? We can avoid any adjustment at all if we're
already aligned. FIXME. We need to allocate space. We'll do that in
push_block below. Special case this because overhead of `push_block' in
this case is non-trivial. We only really need to call `copy_to_reg' in the case
where push insns are going to be used to pass ARGBLOCK
to a function call in ARGS. In that case, the stack
pointer changes value from the allocation point to the
call point, and hence the value of
VIRTUAL_OUTGOING_ARGS_RTX changes as well. But might
as well always do it. The save/restore code in store_one_arg handles all
cases except one: a constructor call (including a C
function returning a BLKmode struct) to initialize
an argument. stack_arg_under_construction says whether a stack
arg is being constructed at the old stack level.
Pushing the stack gets a clean outgoing argument
block. Make a new map for the new argument list.
We can pass TRUE as the 4th argument because we just
saved the stack pointer and will restore it right after
the call. If argument evaluation might modify the stack pointer,
copy the address of the argument list to a register. If we push args individually in reverse order, perform stack alignment
before the first push (the last arg). When the stack adjustment is pending, we get better code
by combining the adjustments. Now that the stack is properly aligned, pops can't safely
be deferred during the evaluation of the arguments. Record the maximum pushed stack space size. We need to delay
doing it this far to take into account the optimization done
by combine_pending_stack_adjustment_and_call. Figure out the register where the value, if any, will come back.
If VALREG is a PARALLEL whose first member has a zero
offset, use that. This is for targets such as m68k that
return the same value in multiple places. Precompute all register parameters. It isn't safe to compute anything
once we have started filling any specific hard regs. Now store (and compute if necessary) all non-register parms.
These come before register parms, since they can require block-moves,
which could clobber the registers used for register parms.
Parms which have partial registers are not stored here,
but we do preallocate space here if they want that. If we have a parm that is passed in registers but not in memory
and whose alignment does not permit a direct copy into registers,
make a group of pseudos that correspond to each register that we
will later fill. Now store any partially-in-registers parm.
This is the last place a block-move can happen. If we pushed args in forward order, perform stack alignment
after pushing the last arg. If register arguments require space on the stack and stack space
was not preallocated, allocate stack space here for arguments
passed in registers. Pass the function the address in which to return a
structure value. Save a pointer to the last insn before the call, so that we can
later safely search backwards to find the CALL_INSN. Set up next argument register. For sibling calls on machines
with register windows this should be the incoming register. All arguments and registers used for the call must be set up by
now! Stack must be properly aligned now.
Generate the actual call instruction.
If the call setup or the call itself overlaps with anything
of the argument setup we probably clobbered our call address.
In that case we can't do sibcalls. If a non-BLKmode value is returned at the most significant end
of a register, shift the register right by the appropriate amount
and update VALREG accordingly. BLKmode values are handled by the
group load/store machinery below. The return value from a malloc-like function is a pointer.
The return value from a malloc-like function can not alias
anything else. Write out the sequence.
For calls to `setjmp', etc., inform
function.c:setjmp_warnings that it should complain if
nonvolatile values are live. For functions that cannot
return, inform flow that control does not fall through. The barrier must be emitted
immediately after the CALL_INSN. Some ports emit more
than just a CALL_INSN above, so we must search for it here. There was no CALL_INSN?
Stack adjustments after a noreturn call are dead code.
However when NO_DEFER_POP is in effect, we must preserve
stack_pointer_delta. If value type not void, return an rtx for the value.
This is the special C++ case where we need to
know what the true target was. We take care to
never use this value more than once in one expression. Handle calls that return values in multiple non-contiguous locations.
The Irix 6 ABI has examples of this. Handle the result of a emit_group_move_into_temps
call in the previous pass. We have to copy a return value in a CLASS_LIKELY_SPILLED hard
reg to a plain register. If TARGET is a MEM in the argument area, and we have
saved part of the argument area, then we can't store
directly into TARGET as it may get overwritten when we
restore the argument save area below. Don't work too
hard though and simply force TARGET to a register if it
is a MEM; the optimizer is quite likely to sort it out. TARGET and VALREG cannot be equal at this point
because the latter would not have
REG_FUNCTION_VALUE_P true, while the former would if
it were referring to the same register.
If they refer to the same register, this move will be
a no-op, except when function inlining is being
done. If we are setting a MEM, this code must be executed.
Since it is emitted after the call insn, sibcall
optimization cannot be performed in that case. If we promoted this return value, make the proper SUBREG.
TARGET might be const0_rtx here, so be careful. Ensure we promote as expected, and get the new unsignedness.
If size of args is variable or this was a constructor call for a stack
argument, restore saved stack-pointer value. If we saved any argument areas, restore them.
If this was alloca, record the new stack level for nonlocal gotos.
Check for the handler slots since we might not have a save area
for non-local gotos. Free up storage we no longer need.
Restore the pending stack adjustment now that we have
finished generating the sibling call sequence. Prepare arg structure for next iteration.
Verify that we've deallocated all the stack we used.
If something prevents making this a sibling call,
zero out the sequence. If tail call production succeeded, we need to remove REG_EQUIV notes on arguments too, as argument area is now clobbered by the call.
References const0_rtx, expand_expr(), EXPAND_NORMAL, FOR_EACH_CALL_EXPR_ARG, and TREE_THIS_VOLATILE.
Referenced by expand_builtin_memset_args().
|
static |
Given the current state of MUST_PREALLOCATE and information about arguments to a function call in NUM_ACTUALS, ARGS and ARGS_SIZE, compute and return the final value for MUST_PREALLOCATE.
See if we have or want to preallocate stack space.
If we would have to push a partially-in-regs parm before other stack parms, preallocate stack space instead.
If the size of some parm is not a multiple of the required stack alignment, we must preallocate.
If the total size of arguments that would otherwise create a copy in a temporary (such as a CALL) is more than half the total argument list size, preallocation is faster.
Another reason to preallocate is if we have a machine (like the m88k) where stack alignment is required to be maintained between every pair of insns, not just when the call is made. However, we assume here that such machines either do not have push insns (and hence preallocation would occur anyway) or the problem is taken care of with PUSH_ROUNDING.
| void fixup_tail_calls | ( | void | ) |
A sibling call sequence invalidates any REG_EQUIV notes made for this function's incoming arguments.
At the start of RTL generation we know the only REG_EQUIV notes in the rtl chain are those for incoming arguments, so we can look for REG_EQUIV notes between the start of the function and the NOTE_INSN_FUNCTION_BEG.
This is (slight) overkill. We could keep track of the highest argument we clobber and be more selective in removing notes, but it does not seem to be worth the effort.
There are never REG_EQUIV notes for the incoming arguments after the NOTE_INSN_FUNCTION_BEG note, so stop if we see it.
| int flags_from_decl_or_type | ( | ) |
Detect flags (function attributes) from the function decl or type node.
The function exp may have the `malloc' attribute.
The function exp may have the `returns_twice' attribute.
Process the pure and const attributes.
References ECF_LOOPING_CONST_OR_PURE, and ECF_NORETURN.
Referenced by dump_possible_polymorphic_call_targets(), ipa_reference_get_not_read_global(), and set_reference_optimization_summary().
| bool gimple_alloca_call_p | ( | ) |
Return true if STMT is an alloca call.
|
static |
Fill in ARGS_SIZE and ARGS array based on the parameters found in CALL_EXPR EXP.
NUM_ACTUALS is the total number of parameters.
N_NAMED_ARGS is the total number of named arguments.
STRUCT_VALUE_ADDR_VALUE is the implicit argument for a struct return value, or null.
FNDECL is the tree code for the target of this call (if known)
ARGS_SO_FAR holds state needed by the target to know where to place the next argument.
REG_PARM_STACK_SPACE is the number of bytes of stack space reserved for arguments which are passed in registers.
OLD_STACK_LEVEL is a pointer to an rtx which olds the old stack level and may be modified by this routine.
OLD_PENDING_ADJ, MUST_PREALLOCATE and FLAGS are pointers to integer flags which may may be modified by this routine.
MAY_TAILCALL is cleared if we encounter an invisible pass-by-reference that requires allocation of stack space.
CALL_FROM_THUNK_P is true if this call is the jump from a thunk to the thunked-to function.
1 if scanning parms front to back, -1 if scanning back to front.
Count arg position in order args appear.
In this loop, we consider args in the order they are written. We fill up ARGS from the front or from the back if necessary so that in any case the first arg to be pushed ends up at the front.
In this case, must reverse order of args
so that we compute and push the last arg first. First fill in the actual arguments in the ARGS array, splitting complex arguments if necessary.
I counts args in order (to be) pushed; ARGPOS counts in order written.
Replace erroneous argument with constant zero.
If TYPE is a transparent union or record, pass things the way
we would pass the first field of the union or record. We have
already verified that the modes are the same. Decide where to pass this arg.
args[i].reg is nonzero if all or part is passed in registers.
args[i].partial is nonzero if part but not all is passed in registers,
and the exact value says how many bytes are passed in registers.
args[i].pass_on_stack is nonzero if the argument must at least be
computed on the stack. It may then be loaded back into registers
if args[i].reg is nonzero.
These decisions are driven by the FUNCTION_... macros and must agree
with those made by function.c. See if this argument should be passed by invisible reference.
If we're compiling a thunk, pass through invisible references
instead of making a copy. We can't use sibcalls if a callee-copied argument is
stored in the current function's frame. We make a copy of the object and pass the address to the
function being called. This is a variable-sized object. Make space on the stack
for it. We can pass TRUE as the 4th argument because we just
saved the stack pointer and will restore it right after
the call. Just change the const function to pure and then let
the next test clear the pure based on
callee_copies. If this is a sibling call and the machine has register windows, the
register window has to be unwinded before calling the routine, so
arguments have to go into the incoming registers. If FUNCTION_ARG returned a (parallel [(expr_list (nil) ...) ...]),
it means that we are to pass this arg in the register(s) designated
by the PARALLEL, but also to pass it in the stack. If this is an addressable type, we must preallocate the stack
since we must evaluate the object into its final location.
If this is to be passed in both registers and the stack, it is simpler
to preallocate. Compute the stack-size of this argument.
Update ARGS_SIZE, the total stack space for args so far.
Increment ARGS_SO_FAR, which has info about which arg-registers
have been used, etc.
References ADD_PARM_SIZE, allocate_dynamic_stack_space(), assign_temp(), build_fold_addr_expr_loc(), compare_tree_int(), COMPLETE_TYPE_P, args_size::constant, copy(), DECL_P, DECL_RTL, ECF_LOOPING_CONST_OR_PURE, ECF_PURE, emit_stack_save(), error_mark_node, expr_size(), first_field(), gen_rtx_MEM(), GENERIC_STACK_CHECK, get_base_address(), GET_CODE, int_size_in_bytes(), integer_type_node, integer_zero_node, arg_data::locate, locate_and_pad_parm(), make_tree(), mark_addressable(), MEM_P, arg_data::mode, NULL_TREE, arg_data::partial, pass_by_reference(), arg_data::pass_on_stack, pending_stack_adjust, promote_function_mode(), reference_callee_copied(), arg_data::reg, SAVE_BLOCK, set_mem_attributes(), locate_and_pad_arg_data::size, store_expr(), arg_data::tail_call_reg, targetm, TREE_ADDRESSABLE, TREE_CODE, TREE_STATIC, TREE_TYPE, arg_data::tree_value, TYPE_ALIGN, TYPE_MODE, TYPE_SIZE_UNIT, TYPE_TRANSPARENT_AGGR, TYPE_UNSIGNED, arg_data::unsignedp, locate_and_pad_arg_data::where_pad, XEXP, and XVECEXP.
Referenced by rtx_for_function_call().
|
static |
Compute whether RTL is based on crtl->args.internal_arg_pointer. Return NULL_RTX if RTL isn't based on it, a CONST_INT offset if RTL is based on it with fixed offset, or PC if this is with variable or unknown offset. TOPLEVEL is true if the function is invoked at the topmost level.
When called at the topmost level, scan pseudo assignments in between the last scanned instruction in the tail call sequence and the latest insn in that sequence.
References int_size_in_bytes(), and TREE_TYPE.
|
static |
Helper function for internal_arg_pointer_based_exp, called through for_each_rtx. Return 1 if *LOC is a register based on crtl->args.internal_arg_pointer. Return -1 if *LOC is not based on it and the subexpressions need not be examined. Otherwise return 0.
|
static |
Helper function for internal_arg_pointer_based_exp. Scan insns in the tail call sequence, starting with first insn that hasn't been scanned yet, and note for each pseudo on the LHS whether it is based on crtl->args.internal_arg_pointer or not, and what offset from that that pointer it has.
Punt on pseudos set multiple times.
|
static |
Return TRUE if FNDECL is either a TM builtin or a TM cloned function. Return FALSE otherwise.
|
static |
Do the register loads required for any wholly-register parms or any parms which are passed both on the stack and in a register. Their expressions were already evaluated.
Mark all register-parms as living through the call, putting these USE insns in the CALL_INSN_FUNCTION_USAGE field.
When IS_SIBCALL, perform the check_sibcall_argument_overlap checking, setting *SIBCALL_FAILURE if appropriate.
Set non-negative if we must move a word at a time, even if
just one word (e.g, partial == 4 && mode == DFmode). Set
to -1 if we just use a normal move insn. This value can be
zero if the argument is a zero size structure.
Handle calls that pass values in multiple non-contiguous
locations. The Irix 6 ABI has examples of this. If simple case, just do move. If normal partial, store_one_arg
has already loaded the register for us. In all other cases,
load the register(s) from memory. If we have pre-computed the values to put in the registers in
the case of non-aligned structures, copy them in now. Check for overlap with already clobbered argument area,
providing that this has non-zero size. Handle a BLKmode that needs shifting.
When a parameter is a block, and perhaps in other cases, it is
possible that it did a load from an argument slot that was
already clobbered. Handle calls that pass values in multiple non-contiguous
locations. The Irix 6 ABI has examples of this.
|
static |
Return true if and only if SIZE storage units (usually bytes) starting from address ADDR overlap with already clobbered argument area. This function is used to determine if we should give up a sibcall.
References BITS_PER_UNIT, emit_move_insn(), expand_shift(), gen_reg_rtx(), gen_rtx_REG(), move_block_to_reg(), operand_subword_force(), REGNO, locate_and_pad_arg_data::size, validize_mem(), word_mode, and XEXP.
| bool must_pass_in_stack_var_size | ( | enum machine_mode | mode, |
| const_tree | type | ||
| ) |
Nonzero if we do not know how to pass TYPE solely in registers.
If the type has variable size...
If the type is marked as addressable (it is required to be constructed into the stack)...
| bool must_pass_in_stack_var_size_or_pad | ( | ) |
Another version of the TARGET_MUST_PASS_IN_STACK hook. This one takes trailing padding of a structure into account. ??? Should be able to merge these two by examining BLOCK_REG_PADDING.
If the type has variable size...
If the type is marked as addressable (it is required to be constructed into the stack)...
If the padding and mode of the type is such that a copy into a register would put it into the wrong part of the register.
|
static |
|
static |
Precompute parameters as needed for a function call.
FLAGS is mask of ECF_* constants.
NUM_ACTUALS is the number of arguments.
ARGS is an array containing information for each argument; this routine fills in the INITIAL_VALUE and VALUE fields for each precomputed argument.
If this is a libcall, then precompute all arguments so that we do not get extraneous instructions emitted as part of the libcall sequence.
If we preallocated the stack space, and some arguments must be passed on the stack, then we must precompute any parameter which contains a function call which will store arguments on the stack. Otherwise, evaluating the parameter may clobber previous parameters which have already been stored into the stack. (we have code to avoid such case by saving the outgoing stack arguments, but it results in worse code)
If this is an addressable type, we cannot pre-evaluate it.
CSE will replace this only if it contains args[i].value
pseudo, so convert it down to the declared mode using
a SUBREG.
|
static |
Precompute all register parameters as described by ARGS, storing values into fields within the ARGS array.
NUM_ACTUALS indicates the total number elements in the ARGS array.
Set REG_PARM_SEEN if we encounter a register parameter.
If we are to promote the function arg to a wider mode,
do it now.
If the value is a non-legitimate constant, force it into a
pseudo now. TLS symbols sometimes need a call to resolve. If we're going to have to load the value by parts, pull the
parts into pseudos. The part extraction process can involve
non-trivial computation. If the value is expensive, and we are inside an appropriately
short loop, put the value into a pseudo and then put the pseudo
into the hard reg.
For small register classes, also do this if this call uses
register parameters. This is to avoid reload conflicts while
loading the parameters registers.
References targetm.
| rtx prepare_call_address | ( | tree | fndecl, |
| rtx | funexp, | ||
| rtx | static_chain_value, | ||
| rtx * | call_fusage, | ||
| int | reg_parm_seen, | ||
| int | sibcallp | ||
| ) |
Force FUNEXP into a form suitable for the address of a CALL, and return that as an rtx. Also load the static chain register if FNDECL is a nested function.
CALL_FUSAGE points to a variable holding the prospective CALL_INSN_FUNCTION_USAGE information.
Make a valid memory address and copy constants through pseudo-regs, but not for a constant address if -fno-function-cse.
If we are using registers for parameters, force the function address into a register now.
References targetm.
|
static |
Given a FNDECL and EXP, return an rtx suitable for use as a target address in a call instruction.
FNDECL is the tree node for the target function. For an indirect call FNDECL will be NULL_TREE.
ADDR is the operand 0 of CALL_EXPR for this call.
Get the function to call, in the form of RTL.
Get a SYMBOL_REF rtx for the function address.
Generate an rtx (probably a pseudo-register) for the address.
References internal_arg_pointer_based_exp(), INTVAL, NULL_RTX, pc_rtx, plus_constant(), and XEXP.
| int setjmp_call_p | ( | ) |
Return nonzero when FNDECL represents a call to setjmp.
References DECL_FUNCTION_CODE.
Referenced by check_call().
| bool shift_return_value | ( | ) |
Given that a function returns a value of mode MODE at the most significant end of hard register VALUE, shift VALUE left or right as specified by LEFT_P. Return true if some action was needed.
Use ashr rather than lshr for right shifts. This is for the benefit of the MIPS port, which requires SImode values to be sign-extended when stored in 64-bit registers.
|
static |
|
static |
Determine if the function identified by NAME and FNDECL is one with special properties we wish to know about.
For example, if the function might return more than one time (setjmp), then set RETURNS_TWICE to a nonzero value.
Similarly set NORETURN if the function is in the longjmp family.
Set MAY_BE_ALLOCA for any memory allocation function that might allocate space from the stack such as alloca.
Exclude functions not at the file scope, or not `extern',
since they are not the magic functions we would otherwise
think they are.
FIXME: this should be handled with attributes, not with this
hacky imitation of DECL_ASSEMBLER_NAME. It's (also) wrong
because you can declare fork() inside a function if you
wish.
We assume that alloca will always be called by name. It
makes no sense to pass it as a pointer-to-function to
anything that does not understand its behavior. Disregard prefix _, __, __x or __builtin_.
References ECF_LEAF, and ECF_NORETURN.
|
static |
Traverse a list of TYPES and expand all complex types into their components.
Before allocating memory, check for the common case of no complex.
Rewrite complex type with component type.
Add another component type for the imaginary part.
Skip the newly created node.
References args_size::constant, count, gcc_assert, locate_and_pad_parm(), and NULL_TREE.
|
static |
Store a single argument for a function call into the register or memory area where it must be passed. *ARG describes the argument value and where to pass it.
ARGBLOCK is the address of the stack-block for all the arguments, or 0 on a machine where arguments are pushed individually.
MAY_BE_ALLOCA nonzero says this could be a call to `alloca' so must be careful about how the stack is used.
VARIABLE_SIZE nonzero says that this was a variable-sized outgoing argument stack. This is used if ACCUMULATE_OUTGOING_ARGS to indicate that we need not worry about saving and restoring the stack.
FNDECL is the declaration of the function we are calling.
Return nonzero if this arg should cause sibcall failure, zero otherwise.
Push a new temporary level for any temporaries we make for this argument.
If this is being stored into a pre-allocated, fixed-size, stack area,
save any previous data at that location. Don't worry about things in the fixed argument area;
it has already been saved. We need to make a save area.
If this isn't going to be placed on both the stack and in registers, set up the register and number of words.
Being passed entirely in a register. We shouldn't be called in this case.
If this arg needs special alignment, don't load the registers here.
If this is being passed partially in a register, we can't evaluate it directly into its stack slot. Otherwise, we can.
stack_arg_under_construction is nonzero if a function argument is
being evaluated directly into the outgoing argument list and
expand_call must take special action to preserve the argument list
if it is called recursively.
For scalar function arguments stack_usage_map is sufficient to
determine which stack slots must be saved and restored. Scalar
arguments in general have pass_on_stack == 0.
If this argument is initialized by a function which takes the
address of the argument (a C++ constructor or a C function
returning a BLKmode structure), then stack_usage_map is
insufficient and expand_call must push the stack around the
function call. Such arguments have pass_on_stack == 1.
Note that it is always safe to set stack_arg_under_construction,
but this generates suboptimal code if set when not needed. If we are promoting object (or for any other reason) the mode
doesn't agree, convert the mode. Check for overlap with already clobbered argument area.
Don't allow anything left on stack from computation of argument to alloca.
If the value is already in the stack slot, we are done.
Argument is a scalar, not entirely passed in registers.
(If part is passed in registers, arg->partial says how much
and emit_push_insn will take care of putting it there.)
Push it, and if its size is less than the
amount of space allocated to it,
also bump stack pointer by the additional space.
Note that in C the default argument promotions
will prevent such mismatches. Compute how much space the push instruction will push.
On many machines, pushing a byte will advance the stack
pointer by a halfword. Compute how much space the argument should get:
round up to a multiple of the alignment for arguments. Compute the alignment of the pushed argument.
This isn't already where we want it on the stack, so put it there.
This can either be done with push or copy insns. Unless this is a partially-in-register argument, the argument is now
in the stack. BLKmode, at least partly to be pushed.
Pushing a nonscalar.
If part is passed in registers, PARTIAL says how much
and emit_push_insn will take care of putting it there. Round its size up to a multiple
of the allocation unit for arguments. PUSH_ROUNDING has no effect on us, because emit_push_insn
for BLKmode is careful to avoid it. When an argument is padded down, the block is aligned to
PARM_BOUNDARY, but the actual argument isn't. emit_push_insn might not work properly if arg->value and
argblock + arg->locate.offset areas overlap. expand_call should ensure this.
Use arg->locate.size.constant instead of size_rtx
because we only care about the part of the argument
on the stack. Even though they appear to be at the same location,
if part of the outgoing argument is in registers,
they aren't really at the same location. Check for
this by making sure that the incoming size is the
same as the outgoing size. Unless this is a partially-in-register argument, the argument is now
in the stack.
??? Unlike the case above, in which we want the actual
address of the data, so that we can load it directly into a
register, here we want the address of the stack slot, so that
it's properly aligned for word-by-word copying or something
like that. It's not clear that this is always correct. Mark all slots this store used.
Once we have pushed something, pops can't safely be deferred during the rest of the arguments.
Free any temporary slots made in processing this argument.
|
static |
|
static |
If any elements in ARGS refer to parameters that are to be passed in registers, but not in memory, and whose alignment does not permit a direct copy into registers. Copy the values into a group of pseudos which we will later copy into the appropriate hard registers.
Pseudos for each unaligned argument will be stored into the array args[argnum].aligned_regs. The caller is responsible for deallocating the aligned_regs array if it is nonzero.
Structures smaller than a word are normally aligned to the least significant byte. On a BYTES_BIG_ENDIAN machine, this means we must skip the empty high order bytes when calculating the bit offset.
There is no need to restrict this code to loading items
in TYPE_ALIGN sized hunks. The bitfield instructions can
load up entire word sized registers efficiently.
??? This may not be needed anymore.
We use to emit a clobber here but that doesn't let later
passes optimize the instructions we emit. By storing 0 into
the register later passes know the first AND to zero out the
bitfield being set in the register is unnecessary. The store
of 0 will be deleted as will at least the first AND.
Variable Documentation
Vector indexed by REGNO - FIRST_PSEUDO_REGISTER, recording if a pseudo is based on crtl->args.internal_arg_pointer. The element is NULL_RTX if the pseudo isn't based on it, a CONST_INT offset if the pseudo is based on it with fixed offset, or PC if this is with variable or unknown offset.
Referenced by preload_common_nodes().
|
static |
Size of STACK_USAGE_MAP.
Referenced by emit_library_call_value_1().
| struct { ... } internal_arg_pointer_exp_state |
Internal state for internal_arg_pointer_based_exp and its helpers.
| rtx scan_start |
Last insn that has been scanned by internal_arg_pointer_based_exp_scan, or NULL_RTX if none has been scanned yet.
|
static |
stack_arg_under_construction is nonzero when an argument may be initialized with a constructor call (including a C function that returns a BLKmode struct) and expand_call must take special action to make sure the object being constructed does not overlap the argument list for the constructor call.
|
static |
A vector of one char per byte of stack space. A byte if nonzero if the corresponding stack location has been used. This vector is used to prevent a function call within an argument from clobbering any stack already set up.
Referenced by emit_library_call_value_1().
|
static |
A bitmap of virtual-incoming stack space. Bit is set if the corresponding stack location's tail call argument has been already stored into the stack. This bitmap is used to prevent sibling call optimization if function tries to use parent's incoming argument slots when they have been already overwritten with tail call arguments.
