[BUG F2I] FP64 to signed INT32 conversion generates erroneous NV flag

[IhsaneTahir]

Hello @pascalgouedo,

I am running a test that converts an input in FP64 format into signed INT32.

operands_i  = '{64'hc1e0000000000020, 64'h0, 64'h0}
rnd_mode_i = RNE
op_i = F2I
op_mode_i  = 0
src_fmt_i = FP64
int_fmt_i = INT32

The result computed by the cvfpu is 64'hffffffff80000000 and the NV flag is set. However, after rounding, the converted result is equal to 64'hffffffff80000000 , ie -2^31 which is representable in INT32 format. Therefore, I expect the NX to be generated instead.

A similar issue was reported here #83 and fixed with release v0.8.1 #101.

The difference with that issue is that the input value was an integer equal to -2^31 (the smallest representable value). In this case, the input value is a double that is rounded into -2^31, which should not trigger an overflow.

@khandelwaltanuj

[IhsaneTahir]

It seems that the proposed bug fix in #101 does not handle handle this case.

cvfpu/src/fpnew_cast_multi.sv

Lines 447 to 454 in 2c79477

	// overflow: when converting to unsigned the range is larger by one
	if ((input_exp_q >= signed'(fpnew_pkg::int_width(int_fmt_q2) - 1 + op_mod_q2)) // Exponent larger than max int range,
	&& !(!op_mod_q2 // unless cast to signed int
	&& input_sign_q // and input value is larges negative int value
	&& (input_exp_q == signed'(fpnew_pkg::int_width(int_fmt_q2) - 1))
	&& (input_mant_q == {1'b1, {INT_MAN_WIDTH-1{1'b0}}}))) begin
	denorm_shamt = '0; // prevent shifting
	of_before_round = 1'b1;

[pascalgouedo]

#97 was addressing FP32 only format, not FP64 one as it was not used/verified for CV32E40Pv2.
So probably it was not generic enough to handle both FP64 and FP32 to INT32 conversion.

@lucabertaccini could help you on this one...