I've applied this to gomp4, which ports a bunch of cleanups and fixes from
trunk.
nathan
2015-09-09 Nathan Sidwell <nat...@codesourcery.com>
Port from trunk:
* config/nvptx/nvptx.md (allocate_stack): Emit sorry.
(atomic_compare_and_swap<mode>_1): Use sel_truesi.
* config/nvptx/nvptx.c (nvptx_write_function_decl): Reformat.
(nvptx_reorg_subreg): Pass insn pattern to asm_operands.
(walk_args_for_param): Use arg_promotion for both split and
non-split args.
(nvptx_declare_function_name): Insert formatting tabs for
consistency. Look at crtl->stack_alignment_needed to determine
alignment.
(nvptx_get_drap_rtx): New.
(TARGET_GET_DRAP_RTX): Override.
* config/nvptx/nvptx.h (MAX_STACK_ALIGNMENT): Set.
Index: gcc/config/nvptx/nvptx.c
===================================================================
--- gcc/config/nvptx/nvptx.c (revision 227598)
+++ gcc/config/nvptx/nvptx.c (working copy)
@@ -425,7 +425,8 @@ nvptx_write_function_decl (std::stringst
/* Declare argument types. */
if ((args != NULL_TREE
- && !(TREE_CODE (args) == TREE_LIST && TREE_VALUE (args) == void_type_node))
+ && !(TREE_CODE (args) == TREE_LIST
+ && TREE_VALUE (args) == void_type_node))
|| is_main
|| return_in_mem
|| DECL_STATIC_CHAIN (decl))
@@ -509,8 +510,8 @@ walk_args_for_param (FILE *file, tree ar
mode = DFmode;
}
- mode = arg_promotion (mode);
}
+ mode = arg_promotion (mode);
while (count-- > 0)
{
i++;
@@ -649,7 +650,7 @@ nvptx_declare_function_name (FILE *file,
else if (TYPE_MODE (result_type) != VOIDmode)
{
machine_mode mode = arg_promotion (TYPE_MODE (result_type));
- fprintf (file, ".reg%s %%retval;\n",
+ fprintf (file, "\t.reg%s %%retval;\n",
nvptx_ptx_type_from_mode (mode, false));
}
@@ -701,9 +702,11 @@ nvptx_declare_function_name (FILE *file,
sz = get_frame_size ();
if (sz > 0 || cfun->machine->has_call_with_sc)
{
+ int alignment = crtl->stack_alignment_needed / BITS_PER_UNIT;
+
fprintf (file, "\t.reg.u%d %%frame;\n"
- "\t.local.align 8 .b8 %%farray[" HOST_WIDE_INT_PRINT_DEC"];\n",
- BITS_PER_WORD, sz == 0 ? 1 : sz);
+ "\t.local.align %d .b8 %%farray[" HOST_WIDE_INT_PRINT_DEC"];\n",
+ BITS_PER_WORD, alignment, sz == 0 ? 1 : sz);
fprintf (file, "\tcvta.local.u%d %%frame, %%farray;\n",
BITS_PER_WORD);
}
@@ -719,10 +722,10 @@ nvptx_declare_function_name (FILE *file,
walk_args_for_param (file, TYPE_ARG_TYPES (fntype), DECL_ARGUMENTS (decl),
true, return_in_mem);
if (return_in_mem)
- fprintf (file, "ld.param.u%d %%ar1, [%%in_ar1];\n",
+ fprintf (file, "\tld.param.u%d %%ar1, [%%in_ar1];\n",
GET_MODE_BITSIZE (Pmode));
if (stdarg_p (fntype))
- fprintf (file, "ld.param.u%d %%argp, [%%in_argp];\n",
+ fprintf (file, "\tld.param.u%d %%argp, [%%in_argp];\n",
GET_MODE_BITSIZE (Pmode));
}
@@ -830,6 +833,14 @@ nvptx_function_ok_for_sibcall (tree, tre
return false;
}
+/* Return Dynamic ReAlignment Pointer RTX. For PTX there isn't any. */
+
+static rtx
+nvptx_get_drap_rtx (void)
+{
+ return NULL_RTX;
+}
+
/* Implement the TARGET_CALL_ARGS hook. Record information about one
argument to the next call. */
@@ -4330,6 +4341,8 @@ nvptx_goacc_reduction (gimple call)
#define TARGET_LIBCALL_VALUE nvptx_libcall_value
#undef TARGET_FUNCTION_OK_FOR_SIBCALL
#define TARGET_FUNCTION_OK_FOR_SIBCALL nvptx_function_ok_for_sibcall
+#undef TARGET_GET_DRAP_RTX
+#define TARGET_GET_DRAP_RTX nvptx_get_drap_rtx
#undef TARGET_SPLIT_COMPLEX_ARG
#define TARGET_SPLIT_COMPLEX_ARG hook_bool_const_tree_true
#undef TARGET_RETURN_IN_MEMORY
Index: gcc/config/nvptx/nvptx.h
===================================================================
--- gcc/config/nvptx/nvptx.h (revision 227598)
+++ gcc/config/nvptx/nvptx.h (working copy)
@@ -58,6 +58,8 @@
#define BIGGEST_ALIGNMENT 64
#define STRICT_ALIGNMENT 1
+#define MAX_STACK_ALIGNMENT (1024 * 8)
+
/* Copied from elf.h and other places. We'd otherwise use
BIGGEST_ALIGNMENT and fail a number of testcases. */
#define MAX_OFILE_ALIGNMENT (32768 * 8)
Index: gcc/config/nvptx/nvptx.md
===================================================================
--- gcc/config/nvptx/nvptx.md (revision 227598)
+++ gcc/config/nvptx/nvptx.md (working copy)
@@ -1274,6 +1274,12 @@
(match_operand 1 "nvptx_register_operand")]
""
{
+ /* The ptx documentation specifies an alloca intrinsic (for 32 bit
+ only) but notes it is not implemented. The assembler emits a
+ confused error message. Issue a blunt one now instead. */
+ sorry ("target cannot support alloca.");
+ emit_insn (gen_nop ());
+ DONE;
if (TARGET_ABI64)
emit_insn (gen_allocate_stack_di (operands[0], operands[1]));
else
@@ -1498,14 +1504,12 @@
(match_operand:SI 7 "const_int_operand")] ;; failure model
""
{
- emit_insn (gen_atomic_compare_and_swap<mode>_1 (operands[1], operands[2], operands[3],
- operands[4], operands[6]));
+ emit_insn (gen_atomic_compare_and_swap<mode>_1
+ (operands[1], operands[2], operands[3], operands[4], operands[6]));
- rtx tmp = gen_reg_rtx (GET_MODE (operands[0]));
- emit_insn (gen_cstore<mode>4 (tmp,
- gen_rtx_EQ (SImode, operands[1], operands[3]),
- operands[1], operands[3]));
- emit_insn (gen_andsi3 (operands[0], tmp, GEN_INT (1)));
+ rtx cond = gen_reg_rtx (BImode);
+ emit_move_insn (cond, gen_rtx_EQ (BImode, operands[1], operands[3]));
+ emit_insn (gen_sel_truesi (operands[0], cond, GEN_INT (1), GEN_INT (0)));
DONE;
})
