[Bug rtl-optimization/105936] [10 Regression] ICE with inline-asm and TLS on x86_64 and -O2 in move_insn
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105936 Stas Sergeev changed: What|Removed |Added Resolution|DUPLICATE |FIXED --- Comment #6 from Stas Sergeev --- But no patch is back-ported, and already closing the ticket?
[Bug rtl-optimization/105936] [10 Regression] ICE with inline-asm and TLS on x86_64 and -O2 in move_insn
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105936 Uroš Bizjak changed: What|Removed |Added Status|UNCONFIRMED |RESOLVED Resolution|--- |DUPLICATE --- Comment #5 from Uroš Bizjak --- Dup of PR104777, the patch from the PR should be backported to gcc-10. *** This bug has been marked as a duplicate of bug 104777 ***
[Bug rtl-optimization/105936] [10 Regression] ICE with inline-asm and TLS on x86_64 and -O2 in move_insn
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105936
--- Comment #4 from Uroš Bizjak ---
Digging a bit further with current gcc-10 branch...
Instrumenting a TLS address splitter in i386.md with some creative printfs:
(define_split
[(match_operand 0 "tls_address_pattern")]
"TARGET_TLS_DIRECT_SEG_REFS"
[(match_dup 0)]
{
debug_rtx (curr_insn);
debug_rtx (operands[0]);
printf ("%s\n", GET_RTX_NAME (classify_insn (operands[0])));
operands[0] = ix86_rewrite_tls_address (operands[0]);
debug_rtx (operands[0]);
printf ("%s\n", GET_RTX_NAME (classify_insn (operands[0])));
})
we get:
(insn 54 51 109 9 (parallel [
(asm_operands/v ("btrl %1,%0") ("") 0 [
(mem/c:BLK (plus:DI (plus:DI (unspec:DI [
(const_int 0 [0])
] UNSPEC_TP)
(reg:DI 100))
(const_int 96 [0x60])) [21 MEM[(struct
revectored_struct *) + 96B]+0 S32 A32])
(reg:SI 101)
]
[
(asm_input:BLK ("m") ../../../src/include/cpu.h:355)
(asm_input:SI ("r") ../../../src/include/cpu.h:355)
]
[] ../../../src/include/cpu.h:355)
(clobber (reg:CC 17 flags))
]) "../../../src/include/cpu.h":355:2 -1
(expr_list:REG_DEAD (reg:SI 101)
(expr_list:REG_DEAD (reg:DI 100)
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil)
(parallel [
(asm_operands/v ("btrl %1,%0") ("") 0 [
(mem/c:BLK (plus:DI (plus:DI (unspec:DI [
(const_int 0 [0])
] UNSPEC_TP)
(reg:DI 100))
(const_int 96 [0x60])) [21 MEM[(struct
revectored_struct *) + 96B]+0 S32 A32])
(reg:SI 101)
]
[
(asm_input:BLK ("m") ../../../src/include/cpu.h:355)
(asm_input:SI ("r") ../../../src/include/cpu.h:355)
]
[] ../../../src/include/cpu.h:355)
(clobber (reg:CC 17 flags))
])
jump_insn
(parallel [
(asm_operands/v ("btrl %1,%0") ("") 0 [
(mem/c:BLK (plus:DI (reg:DI 100)
(const_int 96 [0x60])) [21 MEM[(struct
revectored_struct *) + 96B]+0 S32 A32 AS1])
(reg:SI 101)
]
[
(asm_input:BLK ("m") ../../../src/include/cpu.h:355)
(asm_input:SI ("r") ../../../src/include/cpu.h:355)
]
[] ../../../src/include/cpu.h:355)
(clobber (reg:CC 17 flags))
])
jump_insn
Please note that classify_insn declares the pattern as JUMP_INSN due to:
if (GET_CODE (x) == ASM_OPERANDS && ASM_OPERANDS_LABEL_VEC (x))
return JUMP_INSN;
but no label is defined in asm operands.
[Bug rtl-optimization/105936] [10 Regression] ICE with inline-asm and TLS on x86_64 and -O2 in move_insn
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105936
--- Comment #3 from Uroš Bizjak ---
For some reason, split1 pass converts (insn):
(insn 54 51 109 9 (parallel [
(asm_operands/v ("btrl %1,%0") ("") 0 [
(mem/c:BLK (plus:DI (plus:DI (unspec:DI [
(const_int 0 [0])
] UNSPEC_TP)
(reg:DI 100))
(const_int 96 [0x60])) [21 MEM[(struct
revectored_struct *) + 96B]+0 S32 A32])
(reg:SI 101)
]
[
(asm_input:BLK ("m") ../../../src/include/cpu.h:355)
(asm_input:SI ("r") ../../../src/include/cpu.h:355)
]
[] ../../../src/include/cpu.h:355)
(clobber (reg:CC 17 flags))
]) "../../../src/include/cpu.h":355:2 -1
(expr_list:REG_DEAD (reg:SI 101)
(expr_list:REG_DEAD (reg:DI 100)
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil)
to (jump_insn):
(jump_insn 111 51 113 9 (parallel [
(asm_operands/v ("btrl %1,%0") ("") 0 [
(mem/c:BLK (plus:DI (reg:DI 100)
(const_int 96 [0x60])) [21 MEM[(struct
revectored_struct *) + 96B]+0 S32 A32 AS1])
(reg:SI 101)
]
[
(asm_input:BLK ("m") ../../../src/include/cpu.h:355)
(asm_input:SI ("r") ../../../src/include/cpu.h:355)
]
[] ../../../src/include/cpu.h:355)
(clobber (reg:CC 17 flags))
]) "../../../src/include/cpu.h":355:2 -1
(nil))
This confuses sched2 pass.
[Bug rtl-optimization/105936] [10 Regression] ICE with inline-asm and TLS on x86_64 and -O2 in move_insn
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105936 Andrew Pinski changed: What|Removed |Added Keywords||needs-reduction --- Comment #2 from Andrew Pinski --- Note my reduced testcase fails on 11.2.0 but passes on both 10.x and 11.3.0. So I might have reduced it to the wrong one.
[Bug rtl-optimization/105936] [10 Regression] ICE with inline-asm and TLS on x86_64 and -O2 in move_insn
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105936
Andrew Pinski changed:
What|Removed |Added
Keywords|needs-reduction |inline-asm
Summary|[10 Regression] internal|[10 Regression] ICE with
|compiler error: in |inline-asm and TLS on
|move_insn, at |x86_64 and -O2 in move_insn
|haifa-sched.c:5463 |
--- Comment #1 from Andrew Pinski ---
Reduced testcase:
typedef struct {
unsigned short offset;
unsigned short segment;
} far_t;
extern __thread int vm86u[8];
int int33_hooked;
int is_revectored(void);
void log_printf(void);
far_t int33_unrevect(unsigned short gg)
{
far_t ret = {};
if (int33_hooked)
return ret;
if (is_revectored())
{
log_printf();
__asm__ __volatile__("" : :"m" (vm86u[0]));
}
else
{
if (gg)
log_printf();
}
ret.segment = 0x0f000;
ret.offset = 0xf847;
return ret;
}
