[X86] fdata-sections not work

[LukeSTM]

https://godbolt.org/z/357ro6MK3

demo:

#include <string.h>                                                                                                                                             
 
struct ss_t {
	int s_id;
	char msg[];
};

struct ss_t *mst_1;

void do_some_test(void)
{
	struct ss_t *mst_2;
	char buf[100];
	int fill_num=4;
	char tmp_str[20]="test string!";
	memcpy(buf,&fill_num,sizeof(int));
	memcpy(buf+sizeof(int), tmp_str, strlen(tmp_str) + 1); 
	mst_1=(struct ss_t *)buf;
 
	if (mst_1->s_id==4)
		printf("mst_1->s_id=%d,mst_1->msg=%s\n",mst_1->s_id,mst_1->msg);
}```

Here is a use case where it was found that the fdata-sections option is not worked for the x86 when using clang, but it works fine for aarch64. The gcc compiler also works fine for the x86 architecture.

gcc readelf:
Relocation section '.rela.text.do_some_test' at offset 0x908 contains 4 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
0000000000000012  000000050000000a R_X86_64_32            0000000000000000 .rodata.do_some_test.str1.1 + 0
000000000000001a  0000001100000002 R_X86_64_PC32          0000000000000000 .LC1 - 4
000000000000003d  0000001400000002 R_X86_64_PC32          0000000000000008 mst_1 - 4
000000000000005f  0000001500000002 R_X86_64_PC32          0000000000000000 printf - 4

llvm aarch64 readelf:
Relocation section '.rela.text.do_some_test' at offset 0x6d0 contains 7 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
000000000000000c  0000000400000113 R_AARCH64_ADR_PREL_PG_HI21 0000000000000000 .rodata..L__const.do_some_test.tmp_str + 0
0000000000000010  0000000400000115 R_AARCH64_ADD_ABS_LO12_NC 0000000000000000 .rodata..L__const.do_some_test.tmp_str + 0
000000000000001c  0000001500000113 R_AARCH64_ADR_PREL_PG_HI21 0000000000000000 mst_1 + 0
0000000000000028  0000000700000113 R_AARCH64_ADR_PREL_PG_HI21 0000000000000000 .rodata.str1.1 + 0
000000000000002c  0000000700000115 R_AARCH64_ADD_ABS_LO12_NC 0000000000000000 .rodata.str1.1 + 0
0000000000000044  000000150000011e R_AARCH64_LDST64_ABS_LO12_NC 0000000000000000 mst_1 + 0
0000000000000048  000000160000011b R_AARCH64_CALL26       0000000000000000 printf + 0

llvm x86 readelf:
Relocation section '.rela.text.do_some_test' at offset 0x600 contains 3 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
0000000000000034  0000000900000002 R_X86_64_PC32          0000000000000000 mst_1 - 4
000000000000003b  0000000300000002 R_X86_64_PC32          0000000000000000 .L.str - 4
0000000000000047  0000000a00000004 R_X86_64_PLT32         0000000000000000 printf - 4

[LukeSTM]

CC @topperc @pranavk @compnerd

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Luke (LukeSTM)

https://godbolt.org/z/357ro6MK3

demo:

#include <string.h>                                                                                                                                             
 
struct ss_t {
	int s_id;
	char msg[];
};

struct ss_t *mst_1;

void do_some_test(void)
{
	struct ss_t *mst_2;
	char buf[100];
	int fill_num=4;
	char tmp_str[20]="test string!";
	memcpy(buf,&fill_num,sizeof(int));
	memcpy(buf+sizeof(int), tmp_str, strlen(tmp_str) + 1); 
	mst_1=(struct ss_t *)buf;
 
	if (mst_1->s_id==4)
		printf("mst_1->s_id=%d,mst_1->msg=%s\n",mst_1->s_id,mst_1->msg);
}```

Here is a use case where it was found that the fdata-sections option is not worked for the x86 when using clang, but it works fine for aarch64. The gcc compiler also works fine for the x86 architecture.

gcc readelf:
Relocation section '.rela.text.do_some_test' at offset 0x908 contains 4 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
0000000000000012  000000050000000a R_X86_64_32            0000000000000000 .rodata.do_some_test.str1.1 + 0
000000000000001a  0000001100000002 R_X86_64_PC32          0000000000000000 .LC1 - 4
000000000000003d  0000001400000002 R_X86_64_PC32          0000000000000008 mst_1 - 4
000000000000005f  0000001500000002 R_X86_64_PC32          0000000000000000 printf - 4

llvm aarch64 readelf:
Relocation section '.rela.text.do_some_test' at offset 0x6d0 contains 7 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
000000000000000c  0000000400000113 R_AARCH64_ADR_PREL_PG_HI21 0000000000000000 .rodata..L__const.do_some_test.tmp_str + 0
0000000000000010  0000000400000115 R_AARCH64_ADD_ABS_LO12_NC 0000000000000000 .rodata..L__const.do_some_test.tmp_str + 0
000000000000001c  0000001500000113 R_AARCH64_ADR_PREL_PG_HI21 0000000000000000 mst_1 + 0
0000000000000028  0000000700000113 R_AARCH64_ADR_PREL_PG_HI21 0000000000000000 .rodata.str1.1 + 0
000000000000002c  0000000700000115 R_AARCH64_ADD_ABS_LO12_NC 0000000000000000 .rodata.str1.1 + 0
0000000000000044  000000150000011e R_AARCH64_LDST64_ABS_LO12_NC 0000000000000000 mst_1 + 0
0000000000000048  000000160000011b R_AARCH64_CALL26       0000000000000000 printf + 0

llvm x86 readelf:
Relocation section '.rela.text.do_some_test' at offset 0x600 contains 3 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
0000000000000034  0000000900000002 R_X86_64_PC32          0000000000000000 mst_1 - 4
000000000000003b  0000000300000002 R_X86_64_PC32          0000000000000000 .L.str - 4
0000000000000047  0000000a00000004 R_X86_64_PLT32         0000000000000000 printf - 4

</details>

[topperc]

I'm not sure what I'm supposed to be seeing in the readelf dumps? Can you explain what you're not seeing that you expect?

[Zhenhang1213]

I'm not sure what I'm supposed to be seeing in the readelf dumps? Can you explain what you're not seeing that you expect?

I think he means the difference of rodata

[compnerd]

I think that @topperc is right - the relocation list doesn't really give much. The symbol list is going to be more useful to see what is being emitted where.

[LukeSTM]

compile with -O2 -ffunction-sections -fdata-sections， i think in llvm x86 readelf dump should have a symbol .rodata..L__const.do_some_test.tmp_str like llvm aarch64 readelf dump

llvm x86
Num: Value Size Type Bind Vis Ndx Name
0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND
1: 0000000000000000 0 FILE LOCAL DEFAULT ABS do_something.c
2: 0000000000000000 0 SECTION LOCAL DEFAULT 3 .text.do_some_test
3: 0000000000000000 30 OBJECT LOCAL DEFAULT 7 .L.str
4: 0000000000000000 80 FUNC GLOBAL DEFAULT 3 do_some_test
5: 0000000000000000 8 OBJECT GLOBAL DEFAULT 6 mst_1
6: 0000000000000000 0 NOTYPE GLOBAL DEFAULT UND printf

llvm aarch64
Num: Value Size Type Bind Vis Ndx Name
0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND
1: 0000000000000000 0 FILE LOCAL DEFAULT ABS do_something.c
2: 0000000000000000 0 SECTION LOCAL DEFAULT 3 .text.do_some_test
3: 0000000000000000 0 NOTYPE LOCAL DEFAULT 3 $x.0
4: 0000000000000000 0 SECTION LOCAL DEFAULT 5 .rodata..L__const.do_some_test.tmp_str
5: 0000000000000000 0 NOTYPE LOCAL DEFAULT 5 $d.1
6: 0000000000000000 0 NOTYPE LOCAL DEFAULT 6 $d.2
7: 0000000000000000 0 SECTION LOCAL DEFAULT 7 .rodata.str1.1
8: 0000000000000000 0 NOTYPE LOCAL DEFAULT 7 $d.3
9: 0000000000000000 0 NOTYPE LOCAL DEFAULT 8 $d.4
10: 0000000000000000 0 NOTYPE LOCAL DEFAULT 10 $d.5
11: 0000000000000000 88 FUNC GLOBAL DEFAULT 3 do_some_test
12: 0000000000000000 8 OBJECT GLOBAL DEFAULT 6 mst_1
13: 0000000000000000 0 NOTYPE GLOBAL DEFAULT UND printf

gcc x86
Num: Value Size Type Bind Vis Ndx Name
0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND
1: 0000000000000000 0 FILE LOCAL DEFAULT ABS do_something.c
2: 0000000000000000 0 SECTION LOCAL DEFAULT 1 .text
3: 0000000000000000 0 SECTION LOCAL DEFAULT 2 .data
4: 0000000000000000 0 SECTION LOCAL DEFAULT 3 .bss
5: 0000000000000000 0 SECTION LOCAL DEFAULT 4 .rodata.do_some_test.str1.1
6: 0000000000000000 0 SECTION LOCAL DEFAULT 5 .text.do_some_test
7: 0000000000000000 0 SECTION LOCAL DEFAULT 7 .rodata.cst16
8: 0000000000000000 0 SECTION LOCAL DEFAULT 9 .note.GNU-stack
9: 0000000000000000 0 SECTION LOCAL DEFAULT 11 .eh_frame
10: 0000000000000000 0 NOTYPE LOCAL DEFAULT 7 .LC1
11: 0000000000000000 0 SECTION LOCAL DEFAULT 8 .comment
12: 0000000000000000 0 SECTION LOCAL DEFAULT 10 .note.gnu.property
13: 0000000000000000 109 FUNC GLOBAL DEFAULT 5 do_some_test
14: 0000000000000008 8 OBJECT GLOBAL DEFAULT COM mst_1
15: 0000000000000000 0 NOTYPE GLOBAL DEFAULT UND printf

[MaskRay]

The question is why the declaration char tmp_str[20] = "test string!"; doesn't result in a .rodata section in x86. The constant string is relatively short, yet in x86, it gets expanded into movabsq instructions (as seen in SelectionDAG::getMemcpy). On AArch64, however, this expansion doesn't occur, and the string is placed in a .rodata section instead.

[LukeSTM]

I have a question. Since constant strings can be treated as instructions, why is the rodata section of constant strings still retained when using the fdata-section option?

[LukeSTM]

CC @MaskRay