patch to handle i386 stack alignment using the stack pointer

[Jeffrey A Law]

  In message <199901270648.BAA16929@jwlab.FEITH.COM>you write:
  > This patch causes gcc to align the i386 stack when necessary.  It works
  > by using reload to eliminate the frame pointer from local variables in
  > favor of the stack pointer and by aligning the stack pointer in the
  > prologue.
  > 
  > Notes:
  > 
  >   1) In order to be useful "recommended alignment for local i386 FP
  >      variables patch (version 2)" and "patch to handle i386 stack
  >      alignment using PREFERRED_STACK_BOUNDARY" should be installed.
  >      However, it is * not * necessary to actually define
  >      PREFERRED_STACK_BOUNDARY.
  > 
  >   2) The BRL-CAD benchmark has a total of 443 functions which desire the
  >      stack to be aligned on a 64 bit boundary.  When compiling with a frame
  >      pointer 160 functions were successfully aligned.  When compiling
  >      without a frame pointer 89 functions were successfully aligned.
  > 
  >      The crafty benchmark has a total of 2 functions which desire the
  >      stack to be aligned on a 64 bit boundary.  When compiling with a frame
  >      1 function was successfully aligned.  When compiling without a frame
  >      0 functions were successfully aligned.
  > 
  >      A typical reason that compiling with a frame pointer failed to align
  >      the stack is due to the EXIT_IGNORE_STACK optimization preventing
  >      reload from replacing the frame pointer with the stack pointer.  A
  >      typical reason that compiling without a frame pointer failed to align
  >      the stack is due to the frame pointer register being already used
  >      (I don't force it to be spilled).
  > 
  >   3) Currently aligning the stack is disabled if flag_exceptions is set
  >      since it potentally interacts with the ability to unwind the stack.
  > 
  >   4) There are no regressions with the testsuite on FreeBSD 3.0.
  > 
  >   5) This version of gdb-4.17-PatchJLW02 has been submitted to the
  >      gdb maintainers.  It is included here only as a convenience.
I'd like to try this patch, but I get a variety of conflicts when I apply it
to the current sources.  Any chance you could update and resubmit diffs
against the current tree?

[ ... ]
  > Wed Jan 27 01:42:53 EST 1999  John Wehle  (john@feith.com)
  > 
  > 	* rtl.h (desired_stack_local_alignment): Prototype.
  > 	* function.c (desired_stack_local_alignment): New function.
  > 	(current_function_desired_stack_alignment): Define.
  > 	(init_function_start): Initialize it.
  > 	* output.h Declare it.
  > 	* reload1.c (reload, alter_reg): Set it.
  > 	(reload): Assign stack slots to pseudos that lack hard regs or
  > 	equivalents before calling init_elim_table.
  > 	(init_elim_table): Disable frame (or argument when
  > 	-fomit-frame-pointer) pointer elimination unless stack alignment is
  > 	desired and the hard frame pointer is available.
  > 	(reload): Enable the proper frame (or argument when
  > 	-fomit-frame-pointer) pointer elimination once it's known if
  > 	stack alignment is desired and if the hard frame pointer is available.
  > 	(update_eliminables): Don't clear frame_pointer_needed if there is
  > 	an elimination which depends on it.
  > 
  > 	* i386.h (HARD_FRAME_POINTER_REGNUM): Define.
  > 	(FRAME_POINTER_REGNUM): Change to 17.
  > 	(ELIMINABLE_REGS): Add frame pointer eliminations.
  > 	(FIRST_PSEUDO_REGISTER, FIXED_REGISTERS, CALL_USED_REGISTERS,
  > 	REG_ALLOC_ORDER, REG_CLASS_CONTENTS, INITIAL_ELIMINATION_OFFSET,
  > 	REGNO_OK_FOR_BASE_P, REGNO_OK_FOR_BASE_NONSTRICT_P, HI_REGISTER_NAMES,
  > 	PRINT_REG, DEBUG_PRINT_REG): Match HARD_FRAME_POINTER_REGNUM and
  > 	FRAME_POINTER_REGNUM changes.
  > 	* i386.c (AT_BP, regclass_map, ix86_prologue,
  > 	ix86_can_use_return_insn_p, ix86_epilogue): Likewise.
  > 	* i386.md (epilogue_set_stack_ptr): Likewise.
  > 
  > 	* i386.c (ix86_prologue, ix86_epilogue): Align the stack.
  > 	(ix86_compute_frame_size): Account for the stack being aligned.
  > 	* i386.h (INITIAL_ELIMINATION_OFFSET): Likewise.
  > 
  > 	* gdb-4.17-PatchJLW02: New file containing gdb support for
  > 	debugging i386 code where the frame has been aligned.
  > 
jeff