https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127246
Bug ID: 127246
Summary: [17 Regression] Vectorisation vs computed goto
unfactoring
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
Target: aarch64
We've noticed a large ~30% regression on Cpython performance on aarch64 with
GCC trunk. The problem seems to be with the hot interpreter switch not getting
the pass_duplicate_computed_gotos treatment. A reduced testcase is:
struct frame
{
const unsigned short *instr_ptr;
long *stackpointer;
};
/* Opcode number and stack delta. The deltas must not all be equal, otherwise
the handler tails are identical and cross-jumping merges them anyway. */
#define OPS(X) \
X (0, -1) X (1, 0) X (2, -1) X (3, 0) X (4, -1) X (5, 0) \
X (6, -1) X (7, 0) X (8, -1) X (9, 0) X (10, -1) X (11, 0) \
X (12, -1) X (13, 0) X (14, -1) X (15, 0) X (16, -1) X (17, 0) \
X (18, -1) X (19, 0)
extern long op (long);
#define DISPATCH() \
{ \
unsigned int word = *next_instr; \
goto *targets[word & 0xff]; \
}
#define HANDLER(N, SD) \
L_##N: \
frame->instr_ptr = next_instr; \
next_instr += 1; \
frame->stackpointer = stack_pointer; \
stack_pointer[-1] = op (N); \
stack_pointer += (SD); \
DISPATCH ();
#define ENTRY(N, SD) &&L_##N,
long
interp (struct frame *frame)
{
static const void *const targets[] = { OPS (ENTRY) };
const unsigned short *next_instr = frame->instr_ptr;
long *stack_pointer = frame->stackpointer;
DISPATCH ();
OPS (HANDLER)
}
https://godbolt.org/z/Ee57vv5o9
