Hi all,

after I wrote about a problem about the variadic function handling on the mailing list (http://lists.openrisc.net/pipermail/openrisc/2013-July/001776.html) a few month ago I realized that a lot of programs do not follow the C specification including uClibc, gcc (Stefan Kristiansson send a patch to correct this issue) and DirectFB.

For me there are two solutions available
1. We can fight the endless war against the whole C-community, tell them that they are not following the C specification and patch every single program out there to be able to run it on the or1k processor.
2. Change our ABI.

I would suggest number 2 :). Attached to this email is a patch which alters the gcc code (including the ABI) to be able to handle function calls correctly in which stdargs and varargs are exchanged. The varargs are treated like stdargs and saved in registers and on the stack.

The patch uses the method provided by gcc to save the register arguments on the stack in the prologue of the function (target hook: TARGET_SETUP_INCOMING_VARARGS) . Then the standard implementation of varargs can be used. This implementation makes it necessary, that the arguments are split across register and stack.

Stefan Kristiansson already agreed to alter to LLVM code accordingly if this ABI changes are accepted. No other changes of the current code base are necessary. But a full recompilation of all programs is required.

One sentence in chapter "16.2.3 Parameter Passing" of the or1k-specification has to be changed from

"Individual arguments are not split across registers and stack, and variadic arguments are always put on the stack."

to

"Individual arguments are split across registers and stack, and variadic arguments are passed like standard arguments."

So far I was not able to run the whole gcc testsuite because of random crashes of the kernel during that process. But I can tell you that the C-torture test is running without errors. This part of the suite is testing the varargs behavior very accurately. Additionally I could compile and run a lot of programs and libraries successfully. Take a look at https://github.com/s-macke/jor1k-toolchain-builder for a list of programs I tested.

Best wishes,

Sebastian




*
*








>From 5aedd041d184887975a073acd611ef7e373c67b1 Mon Sep 17 00:00:00 2001
From: Sebastian Macke <[email protected]>
Date: Thu, 12 Sep 2013 23:36:08 +0000
Subject: [PATCH] Treat varargs as stdargs and split arguments between register
 and stack

This patch changes the ABI of or1k to be able to handle function calls
correctly in which stdargs and varargs are exchanged.
The varargs are treated like stdargs and saved in registers and on the stack.
The patch uses the method provided by gcc to save the register arguments
on the stack in the prologue of the function. Then the standard
implementation of varargs can be used.
This implementation makes it necessary, that the arguments are splitted
between register and stack.
---
 gcc/config/or1k/or1k.c | 93 +++++++++++++++++++++++++++++++++++++++-----------
 gcc/config/or1k/or1k.h | 13 +++----
 2 files changed, 78 insertions(+), 28 deletions(-)

diff --git a/gcc/config/or1k/or1k.c b/gcc/config/or1k/or1k.c
index 0070d6e..9695e5a 100644
--- a/gcc/config/or1k/or1k.c
+++ b/gcc/config/or1k/or1k.c
@@ -92,6 +92,7 @@ static struct
   int total_size;
   int vars_size;
   int args_size;
+  int pretend_size;
   int gpr_frame;
   int late_frame;
   HOST_WIDE_INT mask;
@@ -174,6 +175,7 @@ or1k_compute_frame_size (HOST_WIDE_INT size)
 {
   HOST_WIDE_INT args_size;
   HOST_WIDE_INT vars_size;
+  HOST_WIDE_INT pretend_size;
   HOST_WIDE_INT stack_offset;
   HOST_WIDE_INT save_size;
   bool interrupt_p = false;
@@ -181,9 +183,12 @@ or1k_compute_frame_size (HOST_WIDE_INT size)
 
   args_size = crtl->outgoing_args_size;
   vars_size = OR1K_ALIGN (size, 4);
+  /* pretend_size was set in or1k_setup_incoming_varargs */
+  pretend_size = crtl->args.pretend_args_size; 
 
   frame_info.args_size = args_size;
   frame_info.vars_size = vars_size;
+  frame_info.pretend_size = pretend_size; 
   frame_info.gpr_frame = interrupt_p ? or1k_redzone : 0;
 
   /* If the function has local variables, we're committed to
@@ -251,7 +256,7 @@ or1k_compute_frame_size (HOST_WIDE_INT size)
   save_size = (frame_info.gpr_size 
               + (frame_info.save_fp_p ? UNITS_PER_WORD : 0)
               + (frame_info.save_lr_p || flag_pic ? UNITS_PER_WORD : 0));
-  frame_info.total_size = save_size + vars_size + args_size;
+  frame_info.total_size = save_size + vars_size + args_size + pretend_size;
   gcc_assert (PROLOGUE_TMP != STATIC_CHAIN_REGNUM);
   if (frame_info.total_size > 32767 && interrupt_p)
     {
@@ -921,6 +926,12 @@ or1k_expand_prologue (void)
     /* No frame needed.  */
     return;
 
+  /* Reserve stack space before the std arguments for the varargs.
+     This space is filled in or1k_setup_incoming_varargs */
+  if (frame_info.pretend_size)
+    emit_frame_insn (gen_add2_insn (stack_pointer_rtx,
+                                   GEN_INT (-frame_info.pretend_size)));
+
   gcc_assert (!frame_info.save_lr_p || !frame_info.save_fp_p
              || frame_info.lr_save_offset != frame_info.fp_save_offset);
 
@@ -1072,6 +1083,11 @@ or1k_expand_epilogue (void)
   if (frame_info.gpr_frame)
     emit_insn (gen_add2_insn (stack_pointer_rtx,
                              GEN_INT (frame_info.gpr_frame)));
+
+  if (frame_info.pretend_size)
+    emit_frame_insn (gen_add2_insn (stack_pointer_rtx,
+                                   GEN_INT (crtl->args.pretend_args_size)));
+
   emit_jump_insn (gen_return_internal (gen_rtx_REG (Pmode, 9)));
 
 }      /* or1k_expand_epilogue () */
@@ -1569,26 +1585,39 @@ or1k_initial_elimination_offset(int from, int to)
    used by the caller for this argument; likewise FUNCTION_INCOMING_ARG, for
    the called function.
 
-   On the OR1K we never split argumetns between registers and memory.
-
-   JPB 30-Aug-10: Is this correct? Surely we should allow this. The ABI spec
-                  is incomplete on this point.
+   On the OR1K we split arguments between registers and memory.
 
    @param[in] cum    Position of argument under consideration.
    @param[in[ mode   Not sure what this relates to.
    @param[in] type   Type of the argument.
-   @param[in] named  Not sure what this relates to.
+   @param[in] named  stdarg or vararg
 
    @return  The number of bytes of the argument to go into registers          
*/
 /* -------------------------------------------------------------------------- 
*/
 static int
-or1k_arg_partial_bytes (cumulative_args_t cum ATTRIBUTE_UNUSED,
-                        enum machine_mode  mode ATTRIBUTE_UNUSED,
-                        tree               type ATTRIBUTE_UNUSED,
+or1k_arg_partial_bytes (cumulative_args_t cum_t,
+                        enum machine_mode  mode,
+                        tree               type,
                         bool               named ATTRIBUTE_UNUSED)
 {
-  return 0;
+  CUMULATIVE_ARGS *cum = get_cumulative_args (cum_t);
+
+  int words;
+  /* Determine the size of the argument in bytes */
+  unsigned int size =
+    (((mode == BLKmode && type)
+      ? (unsigned int) int_size_in_bytes (type)
+      : GET_MODE_SIZE (mode)) + UNITS_PER_WORD - 1)
+    / UNITS_PER_WORD;
+
+  if (*cum >= GP_ARG_NUM_REG)
+    words = 0;
+  else if (*cum + size > GP_ARG_NUM_REG)
+    words = (*cum + size) - GP_ARG_NUM_REG;
+  else
+    words = 0;
 
+  return words * UNITS_PER_WORD;
 }      /* or1k_arg_partial_bytes () */
 
 
@@ -2012,14 +2041,6 @@ or1k_dwarf_calling_convention (const_tree  function 
ATTRIBUTE_UNUSED)
 #undef TARGET_DWARF_CALLING_CONVENTION
 #define TARGET_DWARF_CALLING_CONVENTION  or1k_dwarf_calling_convention
 
-/* uClibc has some instances where (non-coforming to ISO C) a non-varargs
-   prototype is in scope when calling that function which is implemented
-   as varargs.  We want this to work at least where none of the anonymous
-   arguments are used.  I.e. we want the last named argument to be known
-   as named so it can be passed in a register, varars funtion or not.  */
-#undef TARGET_STRICT_ARGUMENT_NAMING
-#define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
-
 /* Is this suitable for an immediate operand.
 
    JPB 1-Sep-10: Is this correct. We can only do 16-bit immediates directly. */
@@ -2072,6 +2093,7 @@ or1k_function_arg (cumulative_args_t cum, enum 
machine_mode mode,
 }
 #undef TARGET_FUNCTION_ARG
 #define TARGET_FUNCTION_ARG or1k_function_arg
+
 /* Update the data in "cum" to advance over an argument of mode "mode" and
    data type "type".  ("type" is null for libcalls where that information may
    not be available.)  */
@@ -2084,10 +2106,43 @@ or1k_function_arg_advance (cumulative_args_t cum, enum 
machine_mode mode,
   *cum_pnt = OR1K_ROUND_ADVANCE_CUM (*cum_pnt, mode, type)
     + OR1K_ROUND_ADVANCE_ARG (mode, type);
 }
-
 #undef TARGET_FUNCTION_ARG_ADVANCE
 #define TARGET_FUNCTION_ARG_ADVANCE or1k_function_arg_advance
 
+/* Worker function for TARGET_SETUP_INCOMING_VARARGS.
+   This function pushes the required number of 
+   argument registers on the stack.  */
+static void
+or1k_setup_incoming_varargs (cumulative_args_t cum,
+                            enum machine_mode mode,
+                            tree type,
+                            int *pretend_sizep,
+                            int second_time)
+{
+  int i;
+
+  if (second_time)
+    return;
+
+  CUMULATIVE_ARGS *cum_pnt = get_cumulative_args (cum);
+
+  /*  Get the first variadic argument */
+  int next_arg = OR1K_ROUND_ADVANCE_CUM(*cum_pnt, mode, type) + 
OR1K_ROUND_ADVANCE_ARG (mode, type);
+  /*  Determine the number of arguments to push on the stack */
+  *pretend_sizep = (GP_ARG_NUM_REG - next_arg) * UNITS_PER_WORD;
+  
+  /* loop over all arguments and emit move instruction to push the arguments 
on the stack*/
+  for (i = next_arg; i < GP_ARG_NUM_REG; i++)
+    {
+      rtx reg = gen_rtx_REG (Pmode, i + GP_ARG_MIN_REG);
+      rtx slot = plus_constant (Pmode, arg_pointer_rtx, (i-next_arg) * 
UNITS_PER_WORD);
+      rtx mem = gen_rtx_MEM (Pmode, slot);
+      emit_move_insn (mem, reg);  
+    }
+}
+#undef TARGET_SETUP_INCOMING_VARARGS
+#define TARGET_SETUP_INCOMING_VARARGS or1k_setup_incoming_varargs
+
 #undef TARGET_PRINT_OPERAND_ADDRESS
 #define TARGET_PRINT_OPERAND_ADDRESS or1k_print_operand_address
 
diff --git a/gcc/config/or1k/or1k.h b/gcc/config/or1k/or1k.h
index cbfac53..09c176c 100644
--- a/gcc/config/or1k/or1k.h
+++ b/gcc/config/or1k/or1k.h
@@ -656,18 +656,13 @@ enum reg_class
    reference as the pointer to them is passed in a reg if one is available
    (and that is what we're given).
 
-   When passing arguments "named" is always 1.  When receiving arguments
-   "named" is 1 for each argument except the last in a stdarg/varargs
-   function.  In a stdarg function we want to treat the last named arg as
-   named.  In a varargs function we want to treat the last named arg (which is
-   `__builtin_va_alist') as unnamed.
+   Stdargs and varargs defined by the argument named are treated 
+   in the same way.
 
    This macro is only used in this file.  */
 #define OR1K_PASS_IN_REG_P(cum, mode, type, named)                     \
-  ((named)                                                             \
-   && ((OR1K_ROUND_ADVANCE_CUM ((cum), (mode), (type))                 \
-       + OR1K_ROUND_ADVANCE_ARG ((mode), (type))                       \
-       <= GP_ARG_NUM_REG)))
+  ( (OR1K_ROUND_ADVANCE_CUM ((cum), (mode), (type))                    \
+    < GP_ARG_NUM_REG))
 
 /* Output assembler code to FILE to increment profiler label # LABELNO
    for profiling a function entry. */
-- 
1.8.3.2

_______________________________________________
OpenRISC mailing list
[email protected]
http://lists.openrisc.net/listinfo/openrisc

Reply via email to