gcc/0019-fp-model-Enable-fp-model-on-kunpeng.patch

From 8cdb316a3fe205a3089b9c17aec0442f4d5f75be Mon Sep 17 00:00:00 2001
From: bule <bule1@huawei.com>
Date: Sun, 27 Aug 2023 16:49:04 +0800
Subject: [PATCH 19/22] [fp-model] Enable fp-model on kunpeng

Enable fp-model options on kunpeng for precision control.
---
 gcc/common.opt                     |  26 +++++
 gcc/config/aarch64/aarch64-linux.h |   3 +-
 gcc/flag-types.h                   |   9 ++
 gcc/fortran/options.cc             |   8 ++
 gcc/opts-common.cc                 | 146 ++++++++++++++++++++++++++++-
 gcc/opts.cc                        |  68 ++++++++++++++
 6 files changed, 256 insertions(+), 4 deletions(-)

diff --git a/gcc/common.opt b/gcc/common.opt
index 8a0dafc52..f5eef8a45 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -1642,6 +1642,32 @@ ffp-int-builtin-inexact
 Common Var(flag_fp_int_builtin_inexact) Init(1) Optimization
 Allow built-in functions ceil, floor, round, trunc to raise \"inexact\" exceptions.
 
+fftz
+Common Var(flag_ftz) Optimization
+Control fpcr register for flush to zero.
+
+fp-model=
+Common Joined RejectNegative Enum(fp_model) Var(flag_fp_model) Init(FP_MODEL_NORMAL) Optimization
+-fp-model=[normal|fast|precise|except|strict] Perform floating-point precision control.
+
+Enum
+Name(fp_model) Type(enum fp_model) UnknownError(unknown floating point precision model %qs)
+
+EnumValue
+Enum(fp_model) String(normal) Value(FP_MODEL_NORMAL)
+
+EnumValue
+Enum(fp_model) String(fast) Value(FP_MODEL_FAST)
+
+EnumValue
+Enum(fp_model) String(precise) Value(FP_MODEL_PRECISE)
+
+EnumValue
+Enum(fp_model) String(except) Value(FP_MODEL_EXCEPT)
+
+EnumValue
+Enum(fp_model) String(strict) Value(FP_MODEL_STRICT)
+
 ; Nonzero means don't put addresses of constant functions in registers.
 ; Used for compiling the Unix kernel, where strange substitutions are
 ; done on the assembly output.
diff --git a/gcc/config/aarch64/aarch64-linux.h b/gcc/config/aarch64/aarch64-linux.h
index 5e4553d79..a5cba6391 100644
--- a/gcc/config/aarch64/aarch64-linux.h
+++ b/gcc/config/aarch64/aarch64-linux.h
@@ -50,7 +50,8 @@
 #define LINK_SPEC LINUX_TARGET_LINK_SPEC AARCH64_ERRATA_LINK_SPEC
 
 #define GNU_USER_TARGET_MATHFILE_SPEC \
-  "%{Ofast|ffast-math|funsafe-math-optimizations:crtfastmath.o%s}"
+  "%{Ofast|ffast-math|funsafe-math-optimizations|fp-model=fast|fftz:\
+  %{!fno-ftz:crtfastmath.o%s}}"
 
 #undef ENDFILE_SPEC
 #define ENDFILE_SPEC   \
diff --git a/gcc/flag-types.h b/gcc/flag-types.h
index 2c8498169..64c64eb32 100644
--- a/gcc/flag-types.h
+++ b/gcc/flag-types.h
@@ -260,6 +260,15 @@ enum fp_contract_mode {
   FP_CONTRACT_FAST = 2
 };
 
+/* Floating-point precision mode.  */
+enum fp_model {
+  FP_MODEL_NORMAL = 0,
+  FP_MODEL_FAST = 1,
+  FP_MODEL_PRECISE = 2,
+  FP_MODEL_EXCEPT = 3,
+  FP_MODEL_STRICT = 4
+};
+
 /* Scalar storage order kind.  */
 enum scalar_storage_order_kind {
   SSO_NATIVE = 0,
diff --git a/gcc/fortran/options.cc b/gcc/fortran/options.cc
index d0fa634f1..3eb99a84a 100644
--- a/gcc/fortran/options.cc
+++ b/gcc/fortran/options.cc
@@ -243,6 +243,7 @@ form_from_filename (const char *filename)
   return f_form;
 }
 
+static void gfc_handle_fpe_option (const char *arg, bool trap);
 
 /* Finalize commandline options.  */
 
@@ -286,6 +287,13 @@ gfc_post_options (const char **pfilename)
   if (flag_protect_parens == -1)
     flag_protect_parens = !optimize_fast;
 
+  /* If fp-model=precise/strict, turn on all ffpe-trap and ffpe-summary.  */
+  if (flag_fp_model == FP_MODEL_EXCEPT || flag_fp_model == FP_MODEL_STRICT)
+    {
+      gfc_handle_fpe_option ("all", false);
+      gfc_handle_fpe_option ("invalid,zero,overflow,underflow", true);
+    }
+
   /* -Ofast sets implies -fstack-arrays unless an explicit size is set for
      stack arrays.  */
   if (flag_stack_arrays == -1 && flag_max_stack_var_size == -2)
diff --git a/gcc/opts-common.cc b/gcc/opts-common.cc
index 7c07d5046..489a6e02a 100644
--- a/gcc/opts-common.cc
+++ b/gcc/opts-common.cc
@@ -28,7 +28,8 @@ along with GCC; see the file COPYING3.  If not see
 #include "spellcheck.h"
 #include "opts-jobserver.h"
 
-static void prune_options (struct cl_decoded_option **, unsigned int *);
+static void prune_options (struct cl_decoded_option **, unsigned int *,
+			   unsigned int);
 
 /* An option that is undocumented, that takes a joined argument, and
    that doesn't fit any of the classes of uses (language/common,
@@ -1091,7 +1092,7 @@ decode_cmdline_options_to_array (unsigned int argc, const char **argv,
 
   *decoded_options = opt_array;
   *decoded_options_count = num_decoded_options;
-  prune_options (decoded_options, decoded_options_count);
+  prune_options (decoded_options, decoded_options_count, lang_mask);
 }
 
 /* Return true if NEXT_OPT_IDX cancels OPT_IDX.  Return false if the
@@ -1112,11 +1113,109 @@ cancel_option (int opt_idx, int next_opt_idx, int orig_next_opt_idx)
   return false;
 }
 
+/* Check whether opt_idx exists in decoded_options array between index
+   start and end.  If found, return its index in decoded_options,
+   else return end.  */
+static unsigned int
+find_opt_idx (const struct cl_decoded_option *decoded_options,
+	       unsigned int decoded_options_count,
+	       unsigned int start, unsigned int end, unsigned int opt_idx)
+{
+  gcc_assert (end <= decoded_options_count);
+  gcc_assert (opt_idx < cl_options_count);
+  unsigned int k;
+  for (k = start; k < end; k++)
+    {
+      if (decoded_options[k].opt_index == opt_idx)
+	{
+	  return k;
+	}
+    }
+  return k;
+}
+
+/* remove the opt_index element from decoded_options array.  */
+static unsigned int
+remove_option (struct cl_decoded_option *decoded_options,
+	       unsigned int decoded_options_count,
+	       unsigned int opt_index)
+{
+  gcc_assert (opt_index < decoded_options_count);
+  unsigned int i;
+  for (i = opt_index; i < decoded_options_count - 1; i++)
+    {
+      decoded_options[i] = decoded_options[i + 1];
+    }
+  return decoded_options_count - 1;
+}
+
+/* Handle the priority between fp-model, Ofast, and
+   ffast-math.  */
+static unsigned int
+handle_fp_model_driver (struct cl_decoded_option *decoded_options,
+			unsigned int decoded_options_count,
+			unsigned int fp_model_index,
+			unsigned int lang_mask)
+{
+  struct cl_decoded_option fp_model_opt = decoded_options[fp_model_index];
+  enum fp_model model = (enum fp_model) fp_model_opt.value;
+  if (model == FP_MODEL_PRECISE || model == FP_MODEL_STRICT)
+    {
+      /* If found Ofast, override Ofast with O3.  */
+      unsigned int Ofast_index;
+      Ofast_index = find_opt_idx (decoded_options, decoded_options_count,
+				  0, decoded_options_count, OPT_Ofast);
+      while (Ofast_index != decoded_options_count)
+	{
+	  const char *tmp_argv = "-O3";
+	  decode_cmdline_option (&tmp_argv, lang_mask,
+				 &decoded_options[Ofast_index]);
+	  warning (0, "%<-Ofast%> is degraded to %<-O3%> due to %qs",
+		   fp_model_opt.orig_option_with_args_text);
+	  Ofast_index = find_opt_idx (decoded_options, decoded_options_count,
+				      0, decoded_options_count, OPT_Ofast);
+	}
+      /* If found ffast-math before fp-model=precise/strict
+	 it, cancel it.  */
+      unsigned int ffast_math_index;
+      ffast_math_index
+	= find_opt_idx (decoded_options, decoded_options_count, 0,
+			fp_model_index, OPT_ffast_math);
+      if (ffast_math_index != fp_model_index)
+	{
+	  decoded_options_count
+	    = remove_option (decoded_options, decoded_options_count,
+			     ffast_math_index);
+	  warning (0, "%<-ffast-math%> before %qs is canceled",
+		   fp_model_opt.orig_option_with_args_text);
+	}
+    }
+  if (model == FP_MODEL_FAST)
+    {
+      /* If found -fno-fast-math after fp-model=fast, cancel this one.  */
+      unsigned int fno_fast_math_index;
+      fno_fast_math_index
+	= find_opt_idx (decoded_options, decoded_options_count, fp_model_index,
+			decoded_options_count, OPT_ffast_math);
+      if (fno_fast_math_index != decoded_options_count
+	  && decoded_options[fno_fast_math_index].value == 0)
+	{
+	  decoded_options_count
+	    = remove_option (decoded_options, decoded_options_count,
+			     fp_model_index);
+	  warning (0,
+		   "%<-fp-model=fast%> before %<-fno-fast-math%> is canceled");
+	}
+    }
+  return decoded_options_count;
+}
+
 /* Filter out options canceled by the ones after them.  */
 
 static void
 prune_options (struct cl_decoded_option **decoded_options,
-	       unsigned int *decoded_options_count)
+	       unsigned int *decoded_options_count,
+	       unsigned int lang_mask)
 {
   unsigned int old_decoded_options_count = *decoded_options_count;
   struct cl_decoded_option *old_decoded_options = *decoded_options;
@@ -1127,7 +1226,12 @@ prune_options (struct cl_decoded_option **decoded_options,
   const struct cl_option *option;
   unsigned int fdiagnostics_color_idx = 0;
 
+  if (!diagnostic_ready_p ())
+    diagnostic_initialize (global_dc, 0);
+
   /* Remove arguments which are negated by others after them.  */
+
+  unsigned int fp_model_index = old_decoded_options_count;
   new_decoded_options_count = 0;
   for (i = 0; i < old_decoded_options_count; i++)
     {
@@ -1151,6 +1255,34 @@ prune_options (struct cl_decoded_option **decoded_options,
 	  fdiagnostics_color_idx = i;
 	  continue;
 
+	case OPT_fp_model_:
+	  /* Only the last fp-model option will take effect.  */
+	  unsigned int next_fp_model_idx;
+	  next_fp_model_idx = find_opt_idx (old_decoded_options,
+					    old_decoded_options_count,
+					    i + 1,
+					    old_decoded_options_count,
+					    OPT_fp_model_);
+	  if (next_fp_model_idx != old_decoded_options_count)
+	    {
+	      /* Found more than one fp-model, cancel this one.  */
+	      if (old_decoded_options[i].value
+		  != old_decoded_options[next_fp_model_idx].value)
+		{
+		  warning (0, "%qs is overrided by %qs",
+			   old_decoded_options[i].
+			   orig_option_with_args_text,
+			   old_decoded_options[next_fp_model_idx].
+			   orig_option_with_args_text);
+		}
+	      break;
+	    }
+	  else
+	    {
+	      /* Found the last fp-model option.  */
+	      fp_model_index = new_decoded_options_count;
+	    }
+	  /* FALLTHRU.  */
 	default:
 	  gcc_assert (opt_idx < cl_options_count);
 	  option = &cl_options[opt_idx];
@@ -1190,6 +1322,14 @@ keep:
 	  break;
 	}
     }
+  if (fp_model_index < new_decoded_options_count)
+    {
+      new_decoded_options_count
+	= handle_fp_model_driver (new_decoded_options,
+				  new_decoded_options_count,
+				  fp_model_index,
+				  lang_mask);
+    }
 
   if (fdiagnostics_color_idx >= 1)
     {
diff --git a/gcc/opts.cc b/gcc/opts.cc
index a97630d1c..b522ed7e2 100644
--- a/gcc/opts.cc
+++ b/gcc/opts.cc
@@ -328,6 +328,7 @@ static void set_debug_level (uint32_t dinfo, int extended,
 			     struct gcc_options *opts_set,
 			     location_t loc);
 static void set_fast_math_flags (struct gcc_options *opts, int set);
+static void set_fp_model_flags (struct gcc_options *opts, int set);
 static void decode_d_option (const char *arg, struct gcc_options *opts,
 			     location_t loc, diagnostic_context *dc);
 static void set_unsafe_math_optimizations_flags (struct gcc_options *opts,
@@ -2857,6 +2858,10 @@ common_handle_option (struct gcc_options *opts,
       set_fast_math_flags (opts, value);
       break;
 
+    case OPT_fp_model_:
+      set_fp_model_flags (opts, value);
+      break;
+
     case OPT_funsafe_math_optimizations:
       set_unsafe_math_optimizations_flags (opts, value);
       break;
@@ -3266,6 +3271,69 @@ set_fast_math_flags (struct gcc_options *opts, int set)
     }
 }
 
+/* Handle fp-model options.  */
+static void
+set_fp_model_flags (struct gcc_options *opts, int set)
+{
+  enum fp_model model = (enum fp_model) set;
+  switch (model)
+    {
+      case FP_MODEL_FAST:
+	/* Equivalent to open ffast-math.  */
+	set_fast_math_flags (opts, 1);
+	break;
+
+      case FP_MODEL_PRECISE:
+	/* Equivalent to close ffast-math.  */
+	set_fast_math_flags (opts, 0);
+	/* Turn on -frounding-math -fsignaling-nans.  */
+	if (!opts->frontend_set_flag_signaling_nans)
+	  opts->x_flag_signaling_nans = 1;
+	if (!opts->frontend_set_flag_rounding_math)
+	  opts->x_flag_rounding_math = 1;
+	opts->x_flag_expensive_optimizations = 0;
+	opts->x_flag_code_hoisting = 0;
+	opts->x_flag_predictive_commoning = 0;
+	opts->x_flag_fp_contract_mode = FP_CONTRACT_OFF;
+	break;
+
+      case FP_MODEL_EXCEPT:
+	if (!opts->frontend_set_flag_signaling_nans)
+	  opts->x_flag_signaling_nans = 1;
+	if (!opts->frontend_set_flag_errno_math)
+	  opts->x_flag_errno_math = 1;
+	if (!opts->frontend_set_flag_trapping_math)
+	  opts->x_flag_trapping_math = 1;
+	opts->x_flag_fp_int_builtin_inexact = 1;
+	/* Also turn on ffpe-trap in fortran.  */
+	break;
+
+      case FP_MODEL_STRICT:
+	/* Turn on both precise and except.  */
+	if (!opts->frontend_set_flag_signaling_nans)
+	  opts->x_flag_signaling_nans = 1;
+	if (!opts->frontend_set_flag_rounding_math)
+	  opts->x_flag_rounding_math = 1;
+	opts->x_flag_expensive_optimizations = 0;
+	opts->x_flag_code_hoisting = 0;
+	opts->x_flag_predictive_commoning = 0;
+	if (!opts->frontend_set_flag_errno_math)
+	  opts->x_flag_errno_math = 1;
+	if (!opts->frontend_set_flag_trapping_math)
+	  opts->x_flag_trapping_math = 1;
+	opts->x_flag_fp_int_builtin_inexact = 1;
+	opts->x_flag_fp_contract_mode = FP_CONTRACT_OFF;
+	break;
+
+      case FP_MODEL_NORMAL:
+	/* Do nothing.  */
+	break;
+
+      default:
+	gcc_unreachable ();
+    }
+}
+
 /* When -funsafe-math-optimizations is set the following
    flags are set as well.  */
 static void
-- 
2.33.0