コグノスケ

16文字で SEGV

$ echo -n '*a;main(){*a=0;}' > a.c && gcc a.c

a.c:1:1: warning: data definition has no type or storage class
    1 | *a;main(){*a=0;}
      | ^
a.c:1:2: warning: type defaults to 'int' in declaration of 'a' [-Wimplicit-int]
    1 | *a;main(){*a=0;}
      |  ^
a.c:1:4: warning: return type defaults to 'int' [-Wimplicit-int]
    1 | *a;main(){*a=0;}
      |    ^~~~

$ ./a.out

Segmentation fault

5文字で SEGV

$ echo -n 'main;' > a.c && gcc a.c

a.c:1:1: warning: data definition has no type or storage class
    1 | main;
      | ^~~~
a.c:1:1: warning: type defaults to 'int' in declaration of 'main' [-Wimplicit-int]

$ ./a.out

Segmentation fault

SEGV したくない場合

$ echo -n 'main=0xc3;' > a.c && gcc -z execstack a.c

a.c:1:1: warning: data definition has no type or storage class
    1 | main=0xc3;
      | ^~~~
a.c:1:1: warning: type defaults to 'int' in declaration of 'main' [-Wimplicit-int]

$ ./a.out

（SEGV しない）

0 文字で SEGV

$ echo -n > a.c && gcc -nostdlib a.c

/usr/bin/ld: warning: cannot find entry symbol _start; defaulting to 0000000000001000

$ ./a.out

Segmentation fault

0 文字で SEGV のバイナリサイズ

$ gcc -nostdlib a.c

/usr/bin/ld: warning: cannot find entry symbol _start; defaulting to 0000000000001000

$ ls -la a.out

-rwxr-xr-x 1 katsuhiro katsuhiro 9528 Jul  3 04:11 a.out

0 文字で SEGV のバイナリサイズ、static 版

$ echo -n  > a.c && gcc -nostdlib -static a.c

/usr/bin/ld: warning: cannot find entry symbol _start; defaulting to 0000000000401000

$ ls -la a.out

-rwxr-xr-x 1 katsuhiro katsuhiro 968 Jul  3 04:12 a.out

$ strace ./a.out

execve("./a.out", ["./a.out"], 0x7ffc92c2b390 /* 46 vars */) = 0
--- SIGSEGV {si_signo=SIGSEGV, si_code=SEGV_MAPERR, si_addr=0x401000} ---
+++ killed by SIGSEGV +++
Segmentation fault

SEGV するバイナリ、strip する

$ strip -R ".note.gnu.build-id" -R ".comment" a.out

$ ls -la a.out
-rwxr-xr-x 1 katsuhiro katsuhiro 376 Jul  3 04:18 a.out

SEGV するバイナリ、手で削る

$ ls -la a.out

-rwxrwxr-x+ 1 katsuhiro katsuhiro 92 Jul  3 05:43 a.out

SEGV する 92 バイトバイナリ、実行

$ strace ./a.out

execve("./a.out", ["./a.out"], 0x7fff2b4f74a0 /* 47 vars */) = 0
--- SIGSEGV {si_signo=SIGSEGV, si_code=SEGV_MAPERR, si_addr=0x400000001} ---
+++ killed by SIGSEGV +++
Segmentation fault

SEGV する 92 バイトバイナリ、readelf

$ readelf -a a.out

ELF Header:
  Magic:   7f 45 4c 46 02 01 01 00 00 00 00 00 00 00 00 00
  Class:                             ELF64
  Data:                              2's complement, little endian
  Version:                           1 (current)
  OS/ABI:                            UNIX - System V
  ABI Version:                       0
  Type:                              EXEC (Executable file)
  Machine:                           Advanced Micro Devices X86-64
  Version:                           0x1
  Entry point address:               0x400000001
  Start of program headers:          36 (bytes into file)
  Start of section headers:          0 (bytes into file)
  Flags:                             0x0
  Size of this header:               64 (bytes)
  Size of program headers:           56 (bytes)
  Number of program headers:         1
  Size of section headers:           0 (bytes)
  Number of section headers:         0
  Section header string table index: 0
There are no sections in this file.
There are no sections to group in this file.
Program Headers:
  Type           Offset             VirtAddr           PhysAddr
                 FileSiz            MemSiz              Flags  Align
  NULL           0x0000000000000000 0x0000000100380040 0x0000000000000000
                 0x0000000000000000 0x0000000000000000         0x1000
There is no dynamic section in this file.
There are no relocations in this file.
The decoding of unwind sections for machine type Advanced Micro Devices X86-64 is not currently supported.
Dynamic symbol information is not available for displaying symbols.
No version information found in this file.

SEGV するバイナリ、64バイト版

$ ls -la a.out

-rwxrwxr-x+ 1 katsuhiro katsuhiro 64 Jul  3 06:39 a.out

SEGV する 64 バイトバイナリ、実行

$ strace ./a.out

execve("./a.out", ["./a.out"], 0x7ffedf7dfa90 /* 47 vars */) = 0
--- SIGSEGV {si_signo=SIGSEGV, si_code=SEGV_MAPERR, si_addr=0x400000001} ---
+++ killed by SIGSEGV +++
Segmentation fault

SEGV する 64 バイトバイナリ、readelf

$ readelf -a a.out
ELF Header:
  Magic:   7f 45 4c 46 02 01 01 00 00 00 00 00 00 00 00 00 
  Class:                             ELF64
  Data:                              2's complement, little endian
  Version:                           1 (current)
  OS/ABI:                            UNIX - System V
  ABI Version:                       0
  Type:                              EXEC (Executable file)
  Machine:                           Advanced Micro Devices X86-64
  Version:                           0x1
  Entry point address:               0x400000001
  Start of program headers:          8 (bytes into file)
  Start of section headers:          0 (bytes into file)
  Flags:                             0x0
  Size of this header:               64 (bytes)
  Size of program headers:           56 (bytes)
  Number of program headers:         1
  Size of section headers:           0 (bytes)
  Number of section headers:         0
  Section header string table index: 0
There are no sections in this file.
There are no sections to group in this file.
Program Headers:
  Type           Offset             VirtAddr           PhysAddr
                 FileSiz            MemSiz              Flags  Align
  NULL           0x00000001003e0002 0x0000000400000001 0x0000000000000008
                 0x0000000000000000 0x0038004000000000         0x1
There is no dynamic section in this file.
There are no relocations in this file.
The decoding of unwind sections for machine type Advanced Micro Devices X86-64 is not curr
ently supported.
Dynamic symbol information is not available for displaying symbols.
No version information found in this file.

浮動小数点を 60個使うプログラム

void _start(void)
{
	static volatile float gf[60];
	float f00, f01, f02, f03, f04, f05, f06, f07, f08, f09;
	float f10, f11, f12, f13, f14, f15, f16, f17, f18, f19;
	float f20, f21, f22, f23, f24, f25, f26, f27, f28, f29;
	float f30, f31, f32, f33, f34, f35, f36, f37, f38, f39;
	float f40, f41, f42, f43, f44, f45, f46, f47, f48, f49;
	float f50, f51, f52, f53, f54, f55, f56, f57, f58, f59;

	__asm__ volatile("nop;\n");

	f00 = gf[ 0]; f01 = gf[ 1]; f02 = gf[ 2]; f03 = gf[ 3]; f04 = gf[ 4];
	f05 = gf[ 5]; f06 = gf[ 6]; f07 = gf[ 7]; f08 = gf[ 8]; f09 = gf[ 9];
	f10 = gf[10]; f11 = gf[11]; f12 = gf[12]; f13 = gf[13]; f14 = gf[14];
	f15 = gf[15]; f16 = gf[16]; f17 = gf[17]; f18 = gf[18]; f19 = gf[19];
	f20 = gf[20]; f21 = gf[21]; f22 = gf[22]; f23 = gf[23]; f24 = gf[24];
	f25 = gf[25]; f26 = gf[26]; f27 = gf[27]; f28 = gf[28]; f29 = gf[29];
	f30 = gf[30]; f31 = gf[31]; f32 = gf[32]; f33 = gf[33]; f34 = gf[34];
	f35 = gf[35]; f36 = gf[36]; f37 = gf[37]; f38 = gf[38]; f39 = gf[39];
	f40 = gf[40]; f41 = gf[41]; f42 = gf[42]; f43 = gf[43]; f44 = gf[44];
	f45 = gf[45]; f46 = gf[46]; f47 = gf[47]; f48 = gf[48]; f49 = gf[49];
	f50 = gf[59]; f51 = gf[51]; f52 = gf[52]; f53 = gf[53]; f54 = gf[54];
	f55 = gf[55]; f56 = gf[56]; f57 = gf[57]; f58 = gf[58]; f59 = gf[59];

	__asm__ volatile("nop;\n");

	gf[ 0] = f00; gf[ 1] = f01; gf[ 2] = f02; gf[ 3] = f03; gf[ 4] = f04;
	gf[ 5] = f05; gf[ 6] = f06; gf[ 7] = f07; gf[ 8] = f08; gf[ 9] = f09;
	gf[10] = f10; gf[11] = f11; gf[12] = f12; gf[13] = f13; gf[14] = f14;
	gf[15] = f15; gf[16] = f16; gf[17] = f17; gf[18] = f18; gf[19] = f19;
	gf[20] = f20; gf[21] = f21; gf[22] = f22; gf[23] = f23; gf[24] = f24;
	gf[25] = f25; gf[26] = f26; gf[27] = f27; gf[28] = f28; gf[29] = f29;
	gf[30] = f30; gf[31] = f31; gf[32] = f32; gf[33] = f33; gf[34] = f34;
	gf[35] = f35; gf[36] = f36; gf[37] = f37; gf[38] = f38; gf[39] = f39;
	gf[40] = f40; gf[41] = f41; gf[42] = f42; gf[43] = f43; gf[44] = f44;
	gf[45] = f45; gf[46] = f46; gf[47] = f47; gf[48] = f48; gf[49] = f49;
	gf[50] = f50; gf[51] = f51; gf[52] = f52; gf[53] = f53; gf[54] = f54;
	gf[55] = f55; gf[56] = f56; gf[57] = f57; gf[58] = f58; gf[59] = f59;

	__asm__ volatile("nop;\n");
}

reload でコンパイルエラー

$ riscv32-unknown-elf-gcc b.c -march=rv32imafcv -mabi=ilp32f -O3 -Wall -fno-builtin -fno-inline -nostdlib -g

b.c: In function '_start':
b.c:52:1: error: insn does not satisfy its constraints:
   52 | }
      | ^
(insn 562 17 18 2 (set (reg/v:SF 65 v1 [orig:104 f00 ] [104])
        (reg/v:SF 47 fa5 [orig:104 f00 ] [104])) "b.c":23:6 153 {*movsf_hardfloat}
     (nil))

during RTL pass: reload
dump file: b.c.282r.reload
b.c:52:1: internal compiler error: in extract_constrain_insn, at recog.c:2195
0x11b9e4d _fatal_insn(char const*, rtx_def const*, char const*, int, char const*)
        gcc/gcc/rtl-error.c:108
0x11b9ed1 _fatal_insn_not_found(rtx_def const*, char const*, int, char const*)
        gcc/gcc/rtl-error.c:118
0x1130298 extract_constrain_insn(rtx_insn*)
        gcc/gcc/recog.c:2195
0xeabdd8 check_rtl
        gcc/gcc/lra.c:2167
0xead5c0 lra(_IO_FILE*)
        gcc/gcc/lra.c:2604
0xe1a63b do_reload
        gcc/gcc/ira.c:5523
0xe1b078 execute
        gcc/gcc/ira.c:5709

エラーの出力箇所を追う

// gcc/lra.c

/* Function checks RTL for correctness.	 If FINAL_P is true, it is
   done at the end of LRA and the check is more rigorous.  */
static void
check_rtl (bool final_p)
{
  basic_block bb;
  rtx_insn *insn;

  lra_assert (! final_p || reload_completed);
  FOR_EACH_BB_FN (bb, cfun)
    FOR_BB_INSNS (bb, insn)
    if (NONDEBUG_INSN_P (insn)
	&& GET_CODE (PATTERN (insn)) != USE
	&& GET_CODE (PATTERN (insn)) != CLOBBER
	&& GET_CODE (PATTERN (insn)) != ASM_INPUT)
      {
	if (final_p)
	  {
	    extract_constrain_insn (insn);    //★★これ
	    continue;
	  }


// gcc/recog.c

/* Do uncached extract_insn, constrain_operands and complain about failures.
   This should be used when extracting a pre-existing constrained instruction
   if the caller wants to know which alternative was chosen.  */
void
extract_constrain_insn (rtx_insn *insn)
{
  extract_insn (insn);    //★★ INSN RTL から recog_data を作成する
  if (!constrain_operands (reload_completed, get_enabled_alternatives (insn)))    //★★ここで false が返る
    fatal_insn_not_found (insn);    //★★internal compile error が出力される
}

エラーが発生する RTL

(insn 562 17 18 2 (set (reg/v:SF 65 v1 [orig:104 f00 ] [104])
        (reg/v:SF 47 fa5 [orig:104 f00 ] [104])) "b.c":23:6 145 {*movsf_hardfloat}
     (nil))

ブレークするために少し改造

// gcc/lra.c

/* Function checks RTL for correctness.	 If FINAL_P is true, it is
   done at the end of LRA and the check is more rigorous.  */
static void
check_rtl (bool final_p)
{
  basic_block bb;
  rtx_insn *insn;

  lra_assert (! final_p || reload_completed);
  FOR_EACH_BB_FN (bb, cfun)
    FOR_BB_INSNS (bb, insn)
    if (NONDEBUG_INSN_P (insn)
	&& GET_CODE (PATTERN (insn)) != USE
	&& GET_CODE (PATTERN (insn)) != CLOBBER
	&& GET_CODE (PATTERN (insn)) != ASM_INPUT)
      {
        if (INSN_UID(insn) == 562)      //★★ブレークポイントを仕掛けるために追加した部分
          printf("check point!!\n");    //★★ブレークポイントを仕掛けるために追加した部分

	if (final_p)
	  {
	    extract_constrain_insn (insn);
	    continue;
	  }

エラーが発生する constraints

(gdb) p recog_data.constraints
$6 = {
  0x2d717f9 "=f,f,f,m,m,*f,*r,*r,*r,*m",
  0x2d71813 "f,G,m,f,G,*r,*f,*G*r,*m,*r",
...

エラーが起きているときの constraints

// gcc/config/riscv/riscv.md

(define_insn "*movsf_hardfloat"
  [(set (match_operand:SF 0 "nonimmediate_operand" "=f,f,f,m,m,*f,*r,  *r,*r,*m")
	(match_operand:SF 1 "move_operand"         " f,G,m,f,G,*r,*f,*G*r,*m,*r"))]

...

エラーが起きたときの 281r.ira

...

      Popping a57(r107,l0)  -- assign reg 26
      Popping a58(r106,l0)  -- assign reg 27
      Popping a59(r105,l0)  -- assign reg 64    ★64は v0 レジスタの番号
      Popping a60(r104,l0)  -- assign reg 65    ★65は v1 レジスタの番号
      Popping a61(r165,l0)  -- assign reg 5

...

281r.ira で Popping ... を出力しているコード

// gcc/ira-color.c

/* Pop the coloring stack and assign hard registers to the popped
   allocnos.  */
static void
pop_allocnos_from_stack (void)
{
  ira_allocno_t allocno;
  enum reg_class aclass;

  for (;allocno_stack_vec.length () != 0;)
    {
      allocno = allocno_stack_vec.pop ();
      aclass = ALLOCNO_CLASS (allocno);
      if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
	{
	  fprintf (ira_dump_file, "      Popping");  //★★Popping はここで出力
	  ira_print_expanded_allocno (allocno);
	  fprintf (ira_dump_file, "  -- ");
	}
      if (aclass == NO_REGS)
	{
	  ALLOCNO_HARD_REGNO (allocno) = -1;
	  ALLOCNO_ASSIGNED_P (allocno) = true;
	  ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (allocno) == NULL);
	  ira_assert
	    (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno) == NULL);
	  if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
	    fprintf (ira_dump_file, "assign memory\n");
	}
      else if (assign_hard_reg (allocno, false))    //★★レジスタの割当をする
	{
	  if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
	    fprintf (ira_dump_file, "assign reg %d\n",
		     ALLOCNO_HARD_REGNO (allocno));    //★★ assign reg 64 はここで出力
	}


// gcc/ira-build.c

/* This recursive function outputs allocno A and if it is a cap the
   function outputs its members.  */
void
ira_print_expanded_allocno (ira_allocno_t a)
{
  basic_block bb;

  //★★Popping の後ろの axx(rxx, lxx) を出力している

  fprintf (ira_dump_file, " a%d(r%d", ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
  if ((bb = ALLOCNO_LOOP_TREE_NODE (a)->bb) != NULL)
    fprintf (ira_dump_file, ",b%d", bb->index);
  else
    fprintf (ira_dump_file, ",l%d", ALLOCNO_LOOP_TREE_NODE (a)->loop_num);
  if (ALLOCNO_CAP_MEMBER (a) != NULL)
    {
      fprintf (ira_dump_file, ":");
      ira_print_expanded_allocno (ALLOCNO_CAP_MEMBER (a));
    }
  fprintf (ira_dump_file, ")");
}


// gcc/ira-int.h

#define ALLOCNO_NUM(A) ((A)->num)
#define ALLOCNO_REGNO(A) ((A)->regno)
...
#define ALLOCNO_HARD_REGNO(A) ((A)->hard_regno)
...

281r.ira の Popping の値

いずれも ira_allocno_t のメンバを参照する。

（例）
Popping a59(r105,l0)  -- assign reg 64

  - a59   : num
  - r105  : regno
  - l0    : loop_num
  - reg 64: hard_regno

281r.ira でレジスタ割当を決めるコード

// gcc/ira-color.c

static bool
assign_hard_reg (ira_allocno_t a, bool retry_p)
{
  HARD_REG_SET conflicting_regs[2], profitable_hard_regs;
  int i, j, hard_regno, best_hard_regno, class_size;

...

  best_hard_regno = -1;

...

  /* We don't care about giving callee saved registers to allocnos no
     living through calls because call clobbered registers are
     allocated first (it is usual practice to put them first in
     REG_ALLOC_ORDER).  */
  mode = ALLOCNO_MODE (a);
  for (i = 0; i < class_size; i++)
    {
      hard_regno = ira_class_hard_regs[aclass][i];
#ifdef STACK_REGS
      if (no_stack_reg_p
	  && FIRST_STACK_REG <= hard_regno && hard_regno <= LAST_STACK_REG)
	continue;
#endif
      if (! check_hard_reg_p (a, hard_regno,
			      conflicting_regs, profitable_hard_regs))    //★★このチェックを通過すると、割り当てる候補になる
	continue;
      cost = costs[i];
      full_cost = full_costs[i];
      if (!HONOR_REG_ALLOC_ORDER)
	{

...

	}
      if (min_cost > cost)
	min_cost = cost;
      if (min_full_cost > full_cost
	  || (!HONOR_REG_ALLOC_ORDER && min_full_cost == full_cost
	      && best_hard_regno > hard_regno))
	{
	  min_full_cost = full_cost;
	  best_hard_regno = hard_regno;    //★★どのハードウェアレジスタに割り当てるか決まる
	  ira_assert (hard_regno >= 0);
	}
      if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL)
	fprintf (ira_dump_file, "(%d=%d,%d) ", hard_regno, cost, full_cost);
    }

...

  if (! retry_p)
    restore_costs_from_copies (a);
  ALLOCNO_HARD_REGNO (a) = best_hard_regno;    //★★allocno->hard_regno に格納
  ALLOCNO_ASSIGNED_P (a) = true;
  if (best_hard_regno >= 0)
    update_costs_from_copies (a, true, ! retry_p);
  ira_assert (ALLOCNO_CLASS (a) == aclass);
  /* We don't need updated costs anymore.  */
  ira_free_allocno_updated_costs (a);
  return best_hard_regno >= 0;    //★★best_hard_regno に何か値が入っていれば成功
}

281r.ira でレジスタ割当候補を判断するコード

// gcc/ira-color.c

/* Return true if HARD_REGNO is ok for assigning to allocno A with
   PROFITABLE_REGS and whose objects have CONFLICT_REGS.  */
static inline bool
check_hard_reg_p (ira_allocno_t a, int hard_regno,
		  HARD_REG_SET *conflict_regs, HARD_REG_SET profitable_regs)
{
  int j, nwords, nregs;
  enum reg_class aclass;
  machine_mode mode;

  aclass = ALLOCNO_CLASS (a);
  mode = ALLOCNO_MODE (a);
  if (TEST_HARD_REG_BIT (ira_prohibited_class_mode_regs[aclass][mode],
			 hard_regno))    //★★このチェックを通過したら割当候補
    return false;
  /* Checking only profitable hard regs.  */
  if (! TEST_HARD_REG_BIT (profitable_regs, hard_regno))    //★★このチェックは基本的に通過するはず
    return false;

...


// gcc/ira-int.h

#define ALLOCNO_MODE(A) ((A)->mode)
...
#define ALLOCNO_CLASS(A) ((A)->aclass)
...

aclass, mode の値

ブレークに設定する条件は if hard_regno == 64

(gdb) p aclass
$4 = ALL_REGS

(gdb) p mode
$5 = E_SFmode

281r.ira でレジスタ割当候補を判断に使う ira_prohibited_class_mode_regs を初期化するコード

// gcc/ira.c

static void
setup_prohibited_class_mode_regs (void)
ira_prohibited_class_mode_regs
{
  int j, k, hard_regno, cl, last_hard_regno, count;

  for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--)
    {
      temp_hard_regset = reg_class_contents[cl] & ~no_unit_alloc_regs;
      for (j = 0; j < NUM_MACHINE_MODES; j++)
	{
	  count = 0;
	  last_hard_regno = -1;
	  CLEAR_HARD_REG_SET (ira_prohibited_class_mode_regs[cl][j]);
	  for (k = ira_class_hard_regs_num[cl] - 1; k >= 0; k--)
	    {
	      hard_regno = ira_class_hard_regs[cl][k];
	      if (!targetm.hard_regno_mode_ok (hard_regno, (machine_mode) j))    //★★hard_regno_mode_ok() が true であれば割当候補になる
		SET_HARD_REG_BIT (ira_prohibited_class_mode_regs[cl][j],
				  hard_regno);
	      else if (in_hard_reg_set_p (temp_hard_regset,
					  (machine_mode) j, hard_regno))
		{
		  last_hard_regno = hard_regno;
		  count++;
		}
	    }
	  ira_class_singleton[cl][j] = (count == 1 ? last_hard_regno : -1);
	}
    }
}


// gcc/config/riscv/riscv.h

enum reg_class
{
  NO_REGS,			/* no registers in set */
  SIBCALL_REGS,			/* registers used by indirect sibcalls */
  JALR_REGS,			/* registers used by indirect calls */
  GR_REGS,			/* integer registers */
  FP_REGS,			/* floating-point registers */
  VP_REGS,			/* vector registers */
  FRAME_REGS,			/* arg pointer and frame pointer */
  ALL_REGS,			/* all registers */
  LIM_REG_CLASSES		/* max value + 1 */
};

#define N_REG_CLASSES (int) LIM_REG_CLASSES

targetm.hard_regno_mode_ok() の RISC-V 向け実装

// gcc/config/riscv/riscv.c

/* Implement TARGET_HARD_REGNO_MODE_OK.  */

static bool
riscv_hard_regno_mode_ok (unsigned int regno, machine_mode mode)
{
  unsigned int nregs = riscv_hard_regno_nregs (regno, mode);

  if (GP_REG_P (regno))
    {
      if (!GP_REG_P (regno + nregs - 1))
	return false;
    }
  else if (FP_REG_P (regno))
    {
      if (!FP_REG_P (regno + nregs - 1))
	return false;

      if (GET_MODE_CLASS (mode) != MODE_FLOAT
	  && GET_MODE_CLASS (mode) != MODE_COMPLEX_FLOAT)
	return false;

      /* Only use callee-saved registers if a potential callee is guaranteed
	 to spill the requisite width.  */
      if (GET_MODE_UNIT_SIZE (mode) > UNITS_PER_FP_REG
	  || (!call_used_or_fixed_reg_p (regno)
	      && GET_MODE_UNIT_SIZE (mode) > UNITS_PER_FP_ARG))
	return false;
    }
  else if (VP_REG_P (regno))    //★★前回足した実装
    {
      return true;
    }
  else
    return false;

...

machine mode クラスの定義

// gcc/machmode.def

...

/* Basic integer modes.  We go up to TI in generic code (128 bits).
   TImode is needed here because the some front ends now genericly
   support __int128.  If the front ends decide to generically support
   larger types, then corresponding modes must be added here.  The
   name OI is reserved for a 256-bit type (needed by some back ends).
    */

//★★MODE_INT クラスになる

INT_MODE (QI, 1);
INT_MODE (HI, 2);
INT_MODE (SI, 4);
INT_MODE (DI, 8);
INT_MODE (TI, 16);

...


// gcc/config/riscv/riscv-modes.def

//★★MODE_FLOAT クラスになる

FLOAT_MODE (TF, 16, ieee_quad_format);

//★★以前、追加した実装
//★★VECTOR_MODE の場合は少し特殊で、MODE_VECTOR + 最初の引数 クラスになる
//★★この例だと MODE_VECTOR_INT になる

VECTOR_MODE (INT, SI, 32);
VECTOR_MODE (INT, SI, 64);

machine mode とクラスの対応、早見方法

// build_gcc/insn-modes.c

const unsigned char mode_class[NUM_MACHINE_MODES] =
{
  MODE_RANDOM,             /* VOID */
  MODE_RANDOM,             /* BLK */
  MODE_CC,                 /* CC */
  MODE_INT,                /* BI */
  MODE_INT,                /* QI */
  MODE_INT,                /* HI */
  MODE_INT,                /* SI */
  MODE_INT,                /* DI */
  MODE_INT,                /* TI */
  MODE_FRACT,              /* QQ */
  MODE_FRACT,              /* HQ */
  MODE_FRACT,              /* SQ */
  MODE_FRACT,              /* DQ */
  MODE_FRACT,              /* TQ */
  MODE_UFRACT,             /* UQQ */
  MODE_UFRACT,             /* UHQ */
  MODE_UFRACT,             /* USQ */
  MODE_UFRACT,             /* UDQ */
  MODE_UFRACT,             /* UTQ */
  MODE_ACCUM,              /* HA */
  MODE_ACCUM,              /* SA */
  MODE_ACCUM,              /* DA */
  MODE_ACCUM,              /* TA */
  MODE_UACCUM,             /* UHA */
  MODE_UACCUM,             /* USA */
  MODE_UACCUM,             /* UDA */
  MODE_UACCUM,             /* UTA */
  MODE_FLOAT,              /* SF */
  MODE_FLOAT,              /* DF */
  MODE_FLOAT,              /* TF */

...

  MODE_COMPLEX_FLOAT,      /* SC */
  MODE_COMPLEX_FLOAT,      /* DC */
  MODE_COMPLEX_FLOAT,      /* TC */
  MODE_VECTOR_INT,         /* V32SI */
  MODE_VECTOR_INT,         /* V64SI */
};

targetm.hard_regno_mode_ok() の RISC-V 向け実装、修正版

// gcc/config/riscv/riscv.c

/* Implement TARGET_HARD_REGNO_MODE_OK.  */

static bool
riscv_hard_regno_mode_ok (unsigned int regno, machine_mode mode)
{
  unsigned int nregs = riscv_hard_regno_nregs (regno, mode);

  if (GP_REG_P (regno))
    {
      if (!GP_REG_P (regno + nregs - 1))
	return false;
    }
  else if (FP_REG_P (regno))
    {
      if (!FP_REG_P (regno + nregs - 1))
	return false;

      if (GET_MODE_CLASS (mode) != MODE_FLOAT
	  && GET_MODE_CLASS (mode) != MODE_COMPLEX_FLOAT)
	return false;

      /* Only use callee-saved registers if a potential callee is guaranteed
	 to spill the requisite width.  */
      if (GET_MODE_UNIT_SIZE (mode) > UNITS_PER_FP_REG
	  || (!call_used_or_fixed_reg_p (regno)
	      && GET_MODE_UNIT_SIZE (mode) > UNITS_PER_FP_ARG))
	return false;
    }
  else if (VP_REG_P (regno))    //★★前回足した実装
    {
      if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT
	  && GET_MODE_CLASS (mode) != MODE_VECTOR_FLOAT)    //★★今回足した実装
	return false;

      return true;
    }
  else
    return false;

...

1993                                          2000
OPeNBooK  ---,    1996                   ,--> オープンブック（The Tower の版権を持つ）
             +--> オープンブック 9003 ---+--> シノミクス
9003, inc ---'
1990

Debian Testing の ICD loader

$ apt-cache search ocl-icd-libopencl1

ocl-icd-libopencl1 - Generic OpenCL ICD Loader

cl_platform_id と dispatch

// ocl-icd/ocl_icd_loader.c

static inline void _find_and_check_platforms(cl_uint num_icds) {
  cl_uint i;

...

    cl_platform_id *platforms = (cl_platform_id *) malloc( sizeof(cl_platform_id) * num_platforms);
    error = (*plt_fn_ptr)(num_platforms, platforms, NULL);

...

    for(j=0; j<num_platforms; j++) {
      debug(D_LOG, "Checking platform %i", j);
      struct platform_icd *p=&_picds[_num_picds];
      char *param_value=NULL;
      p->extension_suffix=NULL;
      p->vicd=&_icds[i];
      p->pid=platforms[j];    //★★pid = platform ID のことらしい

      /* If clGetPlatformInfo is not exported and we are here, it
       * means that OCL_ICD_ASSUME_ICD_EXTENSION. Si we try to take it
       * from the dispatch * table. If that fails too, we have to
       * bail.
       */
      if (plt_info_ptr == NULL) {
        plt_info_ptr = p->pid->dispatch->clGetPlatformInfo;    //★★dispatch メンバが存在することを前提としている

      
// ocl-icd/khronos-headers/CL/cl.h

typedef struct _cl_platform_id *    cl_platform_id;


// ocl-icd/(build-dir)/ocl_icd_loader_gen.h

struct _cl_platform_id { struct _cl_icd_dispatch *dispatch; };    //★★dispatch 以外は特に規定がなさそう


// ocl-icd/(build-dir)/ocl_icd.h

struct _cl_icd_dispatch {
#ifdef CL_VERSION_1_0
  CL_API_ENTRY cl_int (CL_API_CALL*clGetPlatformIDs)(
    cl_uint          /* num_entries */,
    cl_platform_id * /* platforms */,
    cl_uint *        /* num_platforms */
  ) CL_API_SUFFIX__VERSION_1_0;
#else
  CL_API_ENTRY cl_int (CL_API_CALL* clUnknown0)(void);
#endif

#ifdef CL_VERSION_1_0
  CL_API_ENTRY cl_int (CL_API_CALL*
  clGetPlatformInfo)(
    cl_platform_id   /* platform */,
    cl_platform_info /* param_name */,
    size_t           /* param_value_size */,
    void *           /* param_value */,
    size_t *         /* param_value_size_ret */
  ) CL_API_SUFFIX__VERSION_1_0;
#else
  CL_API_ENTRY cl_int (CL_API_CALL* clUnknown1)(void);
#endif

...

//★★こんな調子で関数ポインタの定義が延々と続く

自動ベクトル化を有効にするコード

// gcc/config/riscv/riscv.c

/* Implement TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES.  */

static unsigned int
riscv_autovectorize_vector_modes (vector_modes *modes, bool)
{
  if (TARGET_VECTOR)
    {
      modes->safe_push (V64SImode);
      modes->safe_push (V32SImode);
    }

  return 0;
}

...

#undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES
#define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES riscv_autovectorize_vector_modes

テスト用の int 配列をコピーする関数

void *cpy(void *dst, const void *src, int n)
{
	int *d = dst;
	const int *s = src;
	int i;

	for (i = 0; i < n / sizeof(*d); i++) {
		d[i] = s[i];
	}

	return dst;
}

自動ベクトル化の結果

//★★自動ベクトル化、あり

                d[i] = s[i];
   100b8:       1202e007                vlw.v   v0,(t0)
   100bc:       10038393                addi    t2,t2,256
   100c0:       f0038793                addi    a5,t2,-256
   100c4:       10028293                addi    t0,t0,256
   100c8:       0207e027                vsw.v   v0,(a5)
        for (i = 0; i < n / sizeof(*d); i++) {
   100cc:       fee296e3                bne     t0,a4,100b8 <cpy+0x44>
   100d0:       00661293                slli    t0,a2,0x6
   100d4:       02568963                beq     a3,t0,10106 <cpy+0x92>
   100d8:       959a                    add     a1,a1,t1
   100da:       932a                    add     t1,t1,a0
                d[i] = s[i];
   100dc:       0005a383                lw      t2,0(a1)


//★★自動ベクトル化、なし

                d[i] = s[i];
   10080:       0005a303                lw      t1,0(a1)
        for (i = 0; i < n / sizeof(*d); i++) {
   10084:       0591                    addi    a1,a1,4
   10086:       0291                    addi    t0,t0,4
                d[i] = s[i];
   10088:       fe62ae23                sw      t1,-4(t0)
        for (i = 0; i < n / sizeof(*d); i++) {
   1008c:       fe759ae3                bne     a1,t2,10080 <cpy+0xc>
        }

        return dst;
}
   10090:       8082                    ret

ベクトル演算の加算、減算、乗算、除算、論理演算（and, or, xor）の定義

;; gcc/config/riscv/riscv.md

(define_attr "vecmode" "unknown,V32SI,V64SI"
  (const_string "unknown"))

...

;; Iterator for hardware supported vector modes.
(define_mode_iterator ANYV [(V32SI "TARGET_VECTOR")
			    (V64SI "TARGET_VECTOR")])

...


;;★★加算

(define_insn "add<mode>3"
  [(set (match_operand:ANYV	       0 "register_operand" "=v")
	(plus:ANYV (match_operand:ANYV 1 "register_operand" " v")
		   (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TARGET_VECTOR"
  "vadd.vvt%0,%1,%2"
  [(set_attr "type" "arith")
   (set_attr "vecmode" "<MODE>")])


;;★★減算

(define_insn "sub<mode>3"
  [(set (match_operand:ANYV		0 "register_operand" "=v")
	(minus:ANYV (match_operand:ANYV 1 "register_operand" " v")
		    (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TARGET_VECTOR"
  "vsub.vvt%0,%1,%2"
  [(set_attr "type" "arith")
   (set_attr "vecmode" "<MODE>")])


;;★★乗算

(define_insn "mul<mode>3"
  [(set (match_operand:ANYV	       0 "register_operand" "=v")
	(mult:ANYV (match_operand:ANYV 1 "register_operand" " v")
		   (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TARGET_VECTOR"
  "vmul.vvt%0,%1,%2"
  [(set_attr "type" "arith")
   (set_attr "vecmode" "<MODE>")])


;;★★除算

;; This code iterator allows unsigned and signed division to be generated
;; from the same template.
(define_code_iterator any_div [div udiv mod umod])

(define_insn "<optab><mode>3"
  [(set (match_operand:ANYV		  0 "register_operand" "=v")
	(any_div:ANYV (match_operand:ANYV 1 "register_operand" " v")
		      (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TARGET_VECTOR"
  "v<insn>.vvt%0,%1,%2"
  [(set_attr "type" "arith")
   (set_attr "vecmode" "<MODE>")])


;;★★論理演算

;; This code iterator allows the three bitwise instructions to be generated
;; from the same template.
(define_code_iterator any_bitwise [and ior xor])

...

(define_insn "<optab><mode>3"
  [(set (match_operand:ANYV		      0 "register_operand" "=v")
	(any_bitwise:ANYV (match_operand:ANYV 1 "register_operand" "%v")
			  (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TAREGET_VECTOR"
  "v<insn>.vvt%0,%1,%2"
  [(set_attr "type" "logical")
   (set_attr "vecmode" "<MODE>")])

ベクトル四則演算、論理演算のサンプルプログラム

typedef int __v64si __attribute__((__vector_size__(256)));

void test()
{
	__v64si v10, v11, v12, v13;0;

	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v10) : "A"(b[10]));
	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v11) : "A"(b[20]));
	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v12) : "A"(b[30]));
	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v13) : "A"(b[40]));

	v10 = v11 + v12;
	v11 &= v12 - v13;
	v12 |= v13 * v10;
	v13 ^= v10 / v11;

	__asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[40]) : "v"(v10));
	__asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[50]) : "v"(v11));
	__asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[60]) : "v"(v12));
	__asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[70]) : "v"(v13));
}

ベクトル四則演算、論理演算のサンプルプログラム（逆アセンブル）

$ riscv32-unknown-elf-gcc b.c -nostdlib -g -Og -march=rv32gcv -mabi=ilp32f

$ riscv32-unknown-elf-objdump -dS a.out

...

        __asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v13) : "A"(b[40]));
   10092:       0a028793                addi    a5,t0,160
   10096:       1207e207                vlw.v   v4,(a5)

        v10 = v11 + v12;
   1009a:       022081d7                vadd.vv v3,v2,v1
        v11 &= v12 - v13;
   1009e:       0a120057                vsub.vv v0,v1,v4
   100a2:       26010057                vand.vv v0,v0,v2
        v12 |= v13 * v10;
   100a6:       9641a157                vmul.vv v2,v4,v3
   100aa:       2a208157                vor.vv  v2,v2,v1
        v13 ^= v10 / v11;
   100ae:       863020d7                vdiv.vv v1,v3,v0
   100b2:       2e1200d7                vxor.vv v1,v1,v4

        __asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[40]) : "v"(v10));
   100b6:       0207e1a7                vsw.v   v3,(a5)

...

ベクトル Not 演算の定義

;; gcc/config/riscv/riscv.md

(define_insn "one_cmpl<mode>2"
  [(set (match_operand:ANYV           0 "register_operand" "=v")
	(not:ANYV (match_operand:ANYV 1 "register_operand" "%v")))]
  "TARGET_VECTOR"
  "vnot.vt%0,%1"
  [(set_attr "type" "logical")
   (set_attr "vecmode" "<MODE>")])

ベクトル Not 演算のサンプルプログラム

typedef int __v64si __attribute__((__vector_size__(256)));

void test()
{
	__v64si v10, v11;
	static int b[1024 * 1024] = {0};

	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v10) : "A"(b[10]));
	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v11) : "A"(b[20]));

	v10 = ~v11;

	__asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[40]) : "v"(v10));
}

コンパイルエラー

$ riscv32-unknown-elf-gcc b.c -nostdlib -Og -march=rv32gcv -mabi=ilp32f

during RTL pass: reload
dump file: b.c.282r.reload
b.c: In function 'test':
b.c:14:1: internal compiler error: in setup_operand_alternative, at lra.c:814
   14 | }
      | ^
0xea5c36 setup_operand_alternative
        gcc/gcc/lra.c:814
0xea70c2 lra_set_insn_recog_data(rtx_insn*)
        gcc/gcc/lra.c:1073
0xea30f9 lra_get_insn_recog_data
        gcc/gcc/lra-int.h:488
0xeab0b9 remove_scratches_1
        gcc/gcc/lra.c:2058
0xeab4c4 remove_scratches
        gcc/gcc/lra.c:2094
0xeac629 lra(_IO_FILE*)
        gcc/gcc/lra.c:2396
0xe1a41f do_reload
        gcc/gcc/ira.c:5523
0xe1ae5c execute
        gcc/gcc/ira.c:5709

エラーを出している箇所

// gcc/lra.c

/* Setup info about operands in alternatives of LRA DATA of insn.  */
static void
setup_operand_alternative (lra_insn_recog_data_t data,
			   const operand_alternative *op_alt)
{
  int i, j, nop, nalt;
  int icode = data->icode;
  struct lra_static_insn_data *static_data = data->insn_static_data;

  static_data->commutative = -1;
  nop = static_data->n_operands;
  nalt = static_data->n_alternatives;
  static_data->operand_alternative = op_alt;
  for (i = 0; i < nop; i++)
    {
      static_data->operand[i].early_clobber_alts = 0;
      static_data->operand[i].is_address = false;
      if (static_data->operand[i].constraint[0] == '%')    //★★% が付いていれば commutative
	{
	  /* We currently only support one commutative pair of operands.  */
	  if (static_data->commutative < 0)
	    static_data->commutative = i;
	  else
	    lra_assert (icode < 0); /* Asm  */
	  /* The last operand should not be marked commutative.  */
	  lra_assert (i != nop - 1);    //★★このアサートに引っかかる
	}
    }

...

ベクトル演算の Not の定義（修正後）

(define_insn "one_cmpl<mode>2"
  [(set (match_operand:ANYV           0 "register_operand" "=v")
	(not:ANYV (match_operand:ANYV 1 "register_operand" " v")))]    ★★% を消す
  "TARGET_VECTOR"
  "vnot.vt%0,%1"
  [(set_attr "type" "logical")
   (set_attr "vecmode" "<MODE>")])

RISC-V スカラ論理演算の定義

;; gcc/config/riscv/riscv.md

;; This code iterator allows the three bitwise instructions to be generated
;; from the same template.
(define_code_iterator any_bitwise [and ior xor])

...

(define_insn "<optab><mode>3"
  [(set (match_operand:X                0 "register_operand" "=r,r")
	(any_bitwise:X (match_operand:X 1 "register_operand" "%r,r")
		       (match_operand:X 2 "arith_operand"    " r,I")))]
  ""
  "<insn>%i2t%0,%1,%2"
  [(set_attr "type" "logical")
   (set_attr "mode" "<MODE>")])