[RV64_DYNAREC] config vector before VLE in sse_get_reg_vector

[zqb-all]

No description provided.

[zqb-all]

box64/src/dynarec/rv64/dynarec_rv64_660f_vector.c

Lines 508 to 521 in 24065ff

	INST_NAME("PMOVSXBW Gx, Ex");
	nextop = F8;
	if (cpuext.xtheadvector) {
	SET_ELEMENT_WIDTH(x1, VECTOR_SEW8, 1);
	GETEX_vector(q1, 0, 0, VECTOR_SEW8);
	GETGX_empty_vector(q0);
	v0 = fpu_get_scratch_lmul(dyn, VECTOR_LMUL2);
	vector_vsetvli(dyn, ninst, x1, VECTOR_SEW8, VECTOR_LMUL1, 0.5);
	VWADD_VX(v0, q1, xZR, VECTOR_UNMASKED);
	SET_ELEMENT_WIDTH(x1, VECTOR_SEW16, 1);
	VMV_V_V(q0, v0);
	} else {
	if (!MODREG) SET_ELEMENT_WIDTH(x1, VECTOR_SEW8, 1);
	GETEX_vector(q1, 0, 0, VECTOR_SEW8);

Line 520: The if condition is not met, not SET_ELEMENT_WIDTH here
Line 521: GETEX_vector

box64/src/dynarec/rv64/dynarec_rv64_helper.h

Lines 454 to 462 in 24065ff

	#define GETEX_vector(a, w, D, sew) \
	if (MODREG) { \
	a = sse_get_reg_vector(dyn, ninst, x1, (nextop & 7) + (rex.b << 3), w, sew); \
	} else { \
	SMREAD(); \
	addr = geted(dyn, addr, ninst, nextop, &ed, x3, x2, &fixedaddress, rex, NULL, 0, D); \
	a = fpu_get_scratch(dyn); \
	VLE_V(a, ed, sew, VECTOR_UNMASKED, VECTOR_NFIELD1); \
	}

Line 456: GETEX_vector calls sse_get_reg_vector.
then sse_get_reg_vector may do VLE without SET_ELEMENT_WIDTH

My initial thought was to remove the if (!MODREG) condition in instructions like PMOVSXBW, always SET_ELEMENT_WIDTH. However, it seems cleaner to add SET_ELEMENT_WIDTH directly inside sse_get_reg_vector, and it also avoids potential bugs in other places that call sse_get_reg_vector without SET_ELEMENT_WIDTH first.
If modifying sse_get_reg_vector introduces issues I'm not currently aware of now, we can fall back to the approach of removing the if (!MODREG) in PMOVSXBW.

[ksco]

My initial thought was to remove the if (!MODREG) condition in instructions like PMOVSXBW, always SET_ELEMENT_WIDTH. However, it seems cleaner to add SET_ELEMENT_WIDTH directly inside sse_get_reg_vector, and it also avoids potential bugs in other places that call sse_get_reg_vector without SET_ELEMENT_WIDTH first.

Thanks. SET_ELEMENT_WIDTH is not free, I think it's better to keep it explicit, so replacing all the if (!MODREG) SET_ELEMENT_WIDTH(x1, VECTOR_SEW8, 1); with SET_ELEMENT_WIDTH(x1, VECTOR_SEW8, 1); is preferred.

[ksco]

Sorry, wrong approval selected.

[zqb-all]

I have another concern:
PMOVSXBW is supposed to read only 8 bytes, but SET_ELEMENT_WIDTH(x1, VECTOR_SEW8, 1) will in fact cause 16 bytes to be read （as vl is 16）. A better way to control this behavior would be something similar to vector_vsetvli(dyn, ninst, x1, VECTOR_SEW8, VECTOR_LMUL1, 0.5); as done in the xtheadvector branch.

what do you think?

[ksco]

I have another concern: PMOVSXBW is supposed to read only 8 bytes, but SET_ELEMENT_WIDTH(x1, VECTOR_SEW8, 1) will in fact cause 16 bytes to be read （as vl is 16）. A better way to control this behavior would be something similar to vector_vsetvli(dyn, ninst, x1, VECTOR_SEW8, VECTOR_LMUL1, 0.5); as done in the xtheadvector branch.

what do you think?

Oh yeah, indeed. We need some GETEX64_vector/GETEX32_vector/GETEX16_vector macros.

[zqb-all]

thanks, make sense, I will add GETEX64_vector/GETEX32_vector/GETEX16_vector

[zqb-all]

GETEX_PARTIAL_vector internally contains SET_ELEMENT_WIDTH, should we also add SET_ELEMENT_WIDTH here? If so, the next step would be to remove all SET_ELEMENT_WIDTH calls that appear before any invocation of GETEX_vector