Fix up call_site_parameter (PR debug/50299)

Hi!

CALL_INSN_FUNCTION_USAGE contains for each (use (reg:M)) or (use (mem:M))
the mode of the promoted argument, not the original mode, which is needed
for RTL optimizations, but prepare_call_arguments would like to know
the original non-promoted argument mode (i.e. TYPE_MODE of the argument's
type).  While sometimes prepare_call_arguments can get at the tree type
of the callee, for e.g. indirect calls that information is simply not
available.  And without that information, e.g. if on a big-endian
64-bit hosts an argument is passed on the stack promoted from SImode to
DImode, we store the address of the whole 64-bit slot instead of adding
+4 to it to get at the 32-bit slot where the actual int argument is passed.

The following patch fixes it by not storing just one mode - the promoted
mode - in C_I_F_U, but also storing there the original mode as well.
The advantage of that is that various non-argument USEs in C_I_F_U will
no longer be tracked as DW_TAG_call_site_parameter - e.g. x86_64 %al
for var-arg functions which is passed the number of SSE arguments.
The original mode is stored in EXPR_LIST's mode.

I've adjusted just the single register arguments (which on various targets
includes even multi-word arguments, as long as the registers are consecutive
etc.) and memory slots, not sure what to do with multi-register arguments,
so I've left them with VOIDmode (this could be changed by adjusting
use_group_regs and/or use_regs).  We do a terribly bad job debug-info wise
for testcases like
struct S { long a, b; };
void baz (void);

__attribute__((noinline, noclone))
int foo (struct S s)
{
  long a = s.a;
  long b = s.b;
  baz ();
  return 2;
}

int bar (void)
{
  struct S s = { 1, 2 };
  foo (s);
  return 0;
}
on x86_64-linux anyway - var-tracking doesn't track those, RTL DSE removes
the argument stores (so debug_insns for e.g. a and b above reference
uninitialized memory) and DECL_RTL of the parameter is used too (so again,
the same uninitialized memory).  Maybe we could for small arguments (say up
to 4 or 8 words) attempt to track them by pieces during var-tracking, as
if SRA split them up, if their DECL_RTL isn't the memory slot in which they
have been passed.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2011-09-06  Jakub Jelinek  <jakub@redhat.com>

	PR debug/50299
	* calls.c (load_register_parameters): Use use_reg_mode instead
	of use_reg when adding a single register CALL_INSN_FUNCTION_USAGE
	entry.
	(expand_call): Set EXPR_LIST mode to TYPE_MODE of the argument
	for stack CALL_INSN_FUNCTION_USAGE uses.
	* expr.h (use_reg_mode): New prototype.
	(use_reg): Changed into inline around use_reg_mode.
	* expr.c (use_reg): Renamed to...
	(use_reg_mode): ... this.  Added MODE argument, set EXPR_LIST
	mode to that mode instead of VOIDmode.
	* var-tracking.c (prepare_call_arguments): Don't track parameters
	whose EXPR_LIST mode is VOIDmode, BLKmode or X mode isn't convertible
	to it using lowpart_subreg.  Convert VALUE and REG/MEM to the
	EXPR_LIST mode.

	Jakub

Fix up call_site_parameter (PR debug/50299)

Commit Message

Comments

Patch