path: root/0031-AutoBOLT-Support-saving-feedback-count-info-to-ELF-s.patch

From 72531376df5ed93c2d945469368ba5514eca8407 Mon Sep 17 00:00:00 2001
From: zhenyu--zhao_admin <zhaozhenyu17@huawei.com>
Date: Tue, 5 Dec 2023 15:33:08 +0800
Subject: [PATCH] [AutoBOLT] Support saving feedback count info to ELF segment
 1/3

---
 gcc/common.opt |   8 +
 gcc/final.cc   | 405 ++++++++++++++++++++++++++++++++++++++++++++++++-
 gcc/opts.cc    |  61 ++++++++
 3 files changed, 473 insertions(+), 1 deletion(-)

diff --git a/gcc/common.opt b/gcc/common.opt
index b01df919e..e69947fc2 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -2546,6 +2546,14 @@ freorder-functions
 Common Var(flag_reorder_functions) Optimization
 Reorder functions to improve code placement.
 
+fauto-bolt
+Common Var(flag_auto_bolt)
+Generate profile from AutoFDO or PGO and do BOLT optimization after linkage.
+
+fauto-bolt=
+Common Joined RejectNegative
+Specify the feedback data directory required by BOLT-plugin.  The default is the current directory.
+
 frerun-cse-after-loop
 Common Var(flag_rerun_cse_after_loop) Optimization
 Add a common subexpression elimination pass after loop optimizations.
diff --git a/gcc/final.cc b/gcc/final.cc
index a9868861b..d4c4fa08f 100644
--- a/gcc/final.cc
+++ b/gcc/final.cc
@@ -81,6 +81,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "rtl-iter.h"
 #include "print-rtl.h"
 #include "function-abi.h"
+#include "insn-codes.h"
 #include "common/common-target.h"
 
 #ifdef XCOFF_DEBUGGING_INFO
@@ -4266,7 +4267,403 @@ leaf_renumber_regs_insn (rtx in_rtx)
       }
 }
 #endif
-
+
+#define ASM_FDO_SECTION_PREFIX ".text.fdo."
+
+#define ASM_FDO_CALLER_FLAG ".fdo.caller "
+#define ASM_FDO_CALLER_SIZE_FLAG ".fdo.caller.size "
+#define ASM_FDO_CALLER_BIND_FLAG ".fdo.caller.bind"
+
+#define ASM_FDO_CALLEE_FLAG ".fdo.callee"
+
+/* Return the relative offset address of the start instruction of BB,
+   return -1 if it is empty instruction.    */
+
+static int 
+get_bb_start_addr (basic_block bb)
+{
+  rtx_insn *insn;
+  FOR_BB_INSNS (bb, insn)
+    {
+      if (!INSN_P (insn))
+	{
+	  continue;
+	}
+      /* The jump target of call is not in this function, so
+	 it should be excluded.    */
+      if (CALL_P (insn))
+        {
+	  return -1;
+	}
+
+      int insn_code = recog_memoized (insn);
+
+      /* The instruction NOP in llvm-bolt belongs to the previous
+	 BB, so it needs to be skipped.   */
+      if (insn_code != CODE_FOR_nop)
+        {
+	  return INSN_ADDRESSES (INSN_UID (insn));
+	}
+    }
+  return -1;
+}
+
+/* Return the relative offet address of the end instruction of BB,
+   return -1 if it is empty or call instruction.    */
+
+static int
+get_bb_end_addr (basic_block bb)
+{
+  rtx_insn *insn;
+  int num_succs = EDGE_COUNT (bb->succs);
+  FOR_BB_INSNS_REVERSE (bb, insn)
+    {
+      if (!INSN_P (insn))
+        {
+	  continue;
+	}
+      /* The jump target of call is not in this function, so
+	 it should be excluded.     */
+      if (CALL_P (insn))
+        {
+	  return -1;
+	}
+      if ((num_succs == 1)
+	   || ((num_succs == 2) && any_condjump_p (insn)))
+	{
+	  return INSN_ADDRESSES (INSN_UID (insn));
+	}
+      else
+        {
+	  return -1;
+	}
+    }
+  return -1;
+}
+
+/* Return the end address of cfun.    */
+
+static int 
+get_function_end_addr ()
+{
+  rtx_insn *insn = get_last_insn ();
+  for (; insn != get_insns (); insn = PREV_INSN (insn))
+    {
+      if (!INSN_P (insn))
+        {
+	  continue;
+	}
+      return INSN_ADDRESSES (INSN_UID (insn));
+    }
+	  
+  return -1;
+} 
+
+/* Return the function profile status string.    */
+
+static const char * 
+get_function_profile_status () 
+{
+  const char *profile_status[] = {
+    "PROFILE_ABSENT",
+    "PROFILE_GUESSED",
+    "PROFILE_READ",
+    "PROFILE_LAST"     /* Last value, used by profile streaming.    */
+  };
+
+  return profile_status[profile_status_for_fn (cfun)];
+}
+
+/* Return the count from the feedback data, such as PGO or ADDO.    */
+
+inline static gcov_type 
+get_fdo_count (profile_count count)
+{
+  return count.quality () >= GUESSED 
+         ? count.to_gcov_type () : 0;
+}
+
+/* Return the profile quality string.    */
+
+static const char *
+get_fdo_count_quality (profile_count count)
+{
+  const char *profile_quality[] = {
+    "UNINITIALIZED_PROFILE",
+    "GUESSED_LOCAL",
+    "GUESSED_GLOBAL0",
+    "GUESSED_GLOBAL0_ADJUSTED",
+    "GUESSED",
+    "AFDO",
+    "ADJUSTED",
+    "PRECISE"
+  };
+
+  return profile_quality[count.quality ()];
+}
+
+static const char *
+alias_local_functions (const char *fnname)
+{
+  if (TREE_PUBLIC (cfun->decl))
+    {
+      return fnname;
+    }
+  return concat (fnname, "/", lbasename (dump_base_name), NULL);
+}
+
+/* Return function bind type string.    */
+
+static const char * 
+simple_get_function_bind ()
+{
+  const char *function_bind[] = {
+    "GLOBAL",
+    "WEAK",
+    "LOCAL",
+    "UNKNOWN"
+  };
+
+  if (TREE_PUBLIC (cfun->decl))
+    {
+      if (!(DECL_WEAK (cfun->decl)))
+        {
+	  return function_bind[0];
+	}
+      else
+        {
+	  return function_bind[1];
+	}
+    }
+  else  
+    {
+      return function_bind[2];
+    }
+		
+  return function_bind[3];
+}
+
+/* Dumo the callee functions insn in bb by CALL_P (insn).   */
+
+static void 
+dump_direct_callee_info_to_asm (basic_block bb, gcov_type call_count)
+{
+  rtx_insn *insn;
+  FOR_BB_INSNS (bb, insn)
+    {
+      if (insn && CALL_P (insn))
+        {
+	  tree callee = get_call_fndecl (insn);
+
+	  if (callee)
+	    {
+	      fprintf (asm_out_file, "\t.string \"%x\"\n",
+		       INSN_ADDRESSES (INSN_UID (insn)));
+
+	      fprintf (asm_out_file, "\t.string \"%s%s\"\n",
+		       ASM_FDO_CALLEE_FLAG,
+                       alias_local_functions (get_fnname_from_decl (callee)));
+
+              fprintf (asm_out_file,
+                       "\t.string \"" HOST_WIDE_INT_PRINT_DEC "\"\n",
+                       call_count);
+
+              if (dump_file)
+                {
+                  fprintf (dump_file, "call: %x --> %s \n",
+                           INSN_ADDRESSES (INSN_UID (insn)),
+                           alias_local_functions
+                           (get_fnname_from_decl (callee)));
+                }
+            }
+        }
+     } 
+}
+
+/* Dump the edge info into asm.    */
+static int
+dump_edge_jump_info_to_asm (basic_block bb, gcov_type bb_count)
+{
+  edge e;
+  edge_iterator ei;
+  gcov_type edge_total_count = 0;
+
+  FOR_EACH_EDGE (e, ei, bb->succs)
+    {
+      gcov_type edge_count = get_fdo_count (e->count ());
+      edge_total_count += edge_count;
+
+      int edge_start_addr = get_bb_end_addr (e->src);
+      int edge_end_addr = get_bb_start_addr(e->dest);
+
+      if (edge_start_addr == -1 || edge_end_addr == -1)
+        {
+          continue;
+        }
+      
+      /* This is a reserved assert for the original design.    If this
+         assert is found, use the address of the previous instruction
+         as edge_start_addr.   */
+      gcc_assert (edge_start_addr != edge_end_addr);
+
+      if (dump_file)
+        {
+          fprintf (dump_file, "edge: %x --> %x = (%ld)\n",
+                   edge_start_addr, edge_end_addr, edge_count);
+        }
+
+      if (edge_count > 0)
+        {
+          fprintf(asm_out_file, "\t.string \"%x\"\n", edge_start_addr);
+          fprintf(asm_out_file, "\t.string \"%x\"\n", edge_end_addr);
+          fprintf(asm_out_file, "\t.string \"" HOST_WIDE_INT_PRINT_DEC "\"\n",
+                  edge_count);
+        }
+    }
+
+    gcov_type call_count = MAX (edge_total_count, bb_count);
+    if (call_count > 0)
+      {
+        dump_direct_callee_info_to_asm (bb, call_count);
+      }
+}
+
+/* Dump the bb info into asm.    */
+
+static void 
+dump_bb_info_to_asm (basic_block bb, gcov_type bb_count)
+{
+  int bb_start_addr = get_bb_start_addr (bb);
+  if (bb_start_addr != -1)
+    {
+      fprintf (asm_out_file, "\t.string \"%x\"\n", bb_start_addr);
+      fprintf (asm_out_file, "\t.string \"" HOST_WIDE_INT_PRINT_DEC "\"\n",
+               bb_count);
+    }
+}
+
+/* Dump the function info into asm.    */
+
+static void 
+dump_function_info_to_asm (const char *fnname)
+{
+  fprintf (asm_out_file, "\t.string \"%s%s\"\n",
+           ASM_FDO_CALLER_FLAG, alias_local_functions (fnname));
+  fprintf (asm_out_file, "\t.string \"%s%d\"\n",
+           ASM_FDO_CALLER_SIZE_FLAG, get_function_end_addr ());
+  fprintf (asm_out_file, "\t.string \"%s%s\"\n",
+           ASM_FDO_CALLER_BIND_FLAG, simple_get_function_bind ());
+
+  if (dump_file)
+    {
+      fprintf (dump_file, "\n FUNC_NAME: %s\n",
+               alias_local_functions (fnname));
+      fprintf (dump_file, " file: %s\n",
+               dump_base_name);
+      fprintf (dump_file, "profile_status: %s\n",
+               get_function_profile_status ());
+      fprintf (dump_file, " size: %x\n",
+               get_function_end_addr ());
+      fprintf (dump_file, " function_bind: %s\n",
+               simple_get_function_bind ());
+    }
+}
+
+/* Dump function profile into form AutoFDO or PGO to asm.    */
+
+static void
+dump_fdo_info_to_asm (const char *fnname)
+{
+  basic_block bb;
+
+  dump_function_info_to_asm (fnname);
+
+  FOR_EACH_BB_FN (bb, cfun)
+    {
+      gcov_type bb_count = get_fdo_count (bb->count);
+      if (bb_count == 0)
+        {
+          continue;
+        }
+      
+      if (dump_file)
+        {
+          fprintf (dump_file, "BB: %x --> %x = (%ld) [%s]\n",
+                   get_bb_start_addr (bb), get_bb_end_addr (bb),
+                   bb_count, get_fdo_count_quality (bb->count));
+        }
+
+      if (flag_profile_use) 
+        {
+          dump_edge_jump_info_to_asm (bb, bb_count);
+        }
+      else if (flag_auto_profile)
+        {
+          dump_bb_info_to_asm (bb, bb_count);
+        }
+    }
+}
+
+/* When -fauto-bolt option is turnded on, the .text.fdo section 
+   will be generated in the *.s file if there is feedback information
+   from PGO or AutoFDO. This section will parserd in BOLT-plugin.    */
+
+static void 
+dump_profile_to_elf_sections ()
+{
+  if (!flag_function_sections)
+    {
+      error ("-fauto-bolt should work with -ffunction-section");
+      return;
+    }
+  if (!flag_ipa_ra)
+    {
+      error ("-fauto-bolt should work with -fipa-ra");
+      return;
+    }
+  if (flag_align_jumps)
+    {
+      error ("-fauto-bolt is not supported with -falign-jumps");
+      return;
+    }
+  if (flag_align_labels)
+    {
+      error ("-fauto-bolt is not spported with -falign-loops");
+      return;
+    }
+  if (flag_align_loops)
+    {
+      error ("-fauto-bolt is not supported with -falign-loops");
+      return;
+    }
+  
+  /* Return if no feedback data.    */
+  if (!flag_profile_use && !flag_auto_profile)
+    {
+      error ("-fauto-bolt should use with -profile-use or -fauto-profile");
+      return;
+    }
+  
+  /* Avoid empty functions.    */
+  if (TREE_CODE (cfun->decl) != FUNCTION_DECL)
+    {
+      return;
+    }
+  int flags = SECTION_DEBUG | SECTION_EXCLUDE;
+  const char *fnname = get_fnname_from_decl (current_function_decl);
+  char *profile_fnname = NULL;
+
+  asprintf (&profile_fnname, "%s%s", ASM_FDO_SECTION_PREFIX, fnname);
+  switch_to_section (get_section (profile_fnname, flags, NULL));
+  dump_fdo_info_to_asm (fnname);
+
+  if (profile_fnname)
+    {
+      free (profile_fnname);
+      profile_fnname = NULL;
+    }
+}
+
 /* Turn the RTL into assembly.  */
 static unsigned int
 rest_of_handle_final (void)
@@ -4334,6 +4731,12 @@ rest_of_handle_final (void)
     targetm.asm_out.destructor (XEXP (DECL_RTL (current_function_decl), 0),
 				decl_fini_priority_lookup
 				  (current_function_decl));
+    
+    if (flag_auto_bolt)
+      {
+        dump_profile_to_elf_sections ();
+      }
+
   return 0;
 }
 
diff --git a/gcc/opts.cc b/gcc/opts.cc
index b868d189e..6d57e7d69 100644
--- a/gcc/opts.cc
+++ b/gcc/opts.cc
@@ -1279,6 +1279,10 @@ finish_options (struct gcc_options *opts, struct gcc_options *opts_set,
   if (opts->x_flag_vtable_verify && opts->x_flag_lto)
     sorry ("vtable verification is not supported with LTO");
 
+  /* Currently -fauto-bolt is not supported for LTO.    */
+  if (opts->x_flag_auto_bolt && opts->x_flag_lto)
+    sorry ("%<-fauto-bolt%> is not supported with LTO");
+
   /* Control IPA optimizations based on different -flive-patching level.  */
   if (opts->x_flag_live_patching)
     control_options_for_live_patching (opts, opts_set,
@@ -1291,6 +1295,58 @@ finish_options (struct gcc_options *opts, struct gcc_options *opts_set,
       = (opts->x_flag_unroll_loops
          || opts->x_flag_peel_loops
          || opts->x_optimize >= 3);
+  
+  if (opts->x_flag_auto_bolt)
+    {
+      /* Record the function section to facilitate the feedback
+	 data storage.  */
+      if (!opts->x_flag_function_sections)
+        {
+	  inform (loc,
+	          "%<-fauto-bolt%> should work with %<-ffunction-sections%>,"
+		  " enabling %<-ffunction-sections%>");
+	  opts->x_flag_function_sections = true;
+	}
+
+      /* Cancel the internal alignment of the function.  The binary
+	 optimizer bolt will cancel the internal alignment optimization
+	 of the function, so the alignment is meaningless at this time,
+	 and if not, it will bring trouble to the calculation of the
+	 offset address of the instruction.  */
+      if (opts->x_flag_align_jumps)
+        {
+	  inform (loc,
+		  "%<-fauto-bolt%> should not work with %<-falign-jumps%>,"
+		  " disabling %<-falign-jumps%>");
+	  opts->x_flag_align_jumps = false;
+	}
+
+      if (opts->x_flag_align_labels)
+        {
+	  inform (loc,
+		  "%<-fauto-bolt%> should not work with %<-falign-labels%>,"
+		  " disabling %<-falign-labels%>");
+	          opts->x_flag_align_labels = false;
+	}
+
+      if (opts->x_flag_align_loops)
+        {
+	  inform (loc,
+		  "%<-fauto-bolt%> should not work with %<-falign-loops%>,"
+		  " disabling %<-falign-loops%>");
+	  opts->x_flag_align_loops = false;
+	}
+
+      /* When parsing instructions in RTL phase, we need to know
+	 the call information of instructions to avoid being optimized.  */
+      if (!opts->x_flag_ipa_ra)
+        {
+	  inform (loc,
+		  "%<-fauto-bolt%> should work with %<-fipa-ra%>,"
+		  " enabling %<-fipa-ra%>");
+	  opts->x_flag_ipa_ra = true;
+	}
+    }
 
   /* With -fcx-limited-range, we do cheap and quick complex arithmetic.  */
   if (opts->x_flag_cx_limited_range)
@@ -3226,6 +3282,11 @@ common_handle_option (struct gcc_options *opts,
 				&opts->x_flag_align_functions,
 				&opts->x_str_align_functions);
       break;
+    
+    case OPT_fauto_bolt_:
+    case OPT_fauto_bolt:
+      /* Deferred.  */  
+      break;  
 
     case OPT_ftabstop_:
       /* It is documented that we silently ignore silly values.  */
-- 
2.33.0