On 8/27/23 09:36, Bastian Koppelmann wrote:
On Sun, Aug 27, 2023 at 07:49:52AM -0700, Richard Henderson wrote:
On 8/27/23 04:07, Bastian Koppelmann wrote:
On Sat, Aug 26, 2023 at 09:50:51PM -0700, Richard Henderson wrote:
On 8/26/23 09:02, Bastian Koppelmann wrote:
+uint32_t helper_ftou(CPUTriCoreState *env, uint32_t arg)
+{
+ float32 f_arg = make_float32(arg);
+ uint32_t result;
+ int32_t flags = 0;
+
+ if (float32_is_any_nan(f_arg)) {
+ result = 0;
+ flags |= float_flag_invalid;
+ } else if (float32_lt_quiet(f_arg, 0, &env->fp_status)) {
+ result = 0;
+ flags = float_flag_invalid;
+ } else {
+ result = float32_to_uint32(f_arg, &env->fp_status);
+ flags = f_get_excp_flags(env);
+ }
You should allow softfloat to diagnose the special cases, and negative -> 0
is standard behaviour. Therefore:
You're right. However, there is one special case, negative -> 0 ought to raise
float_flags_invalid.
https://gitlab.com/qemu-project/qemu/-/blob/master/fpu/softfloat-parts.c.inc?ref_type=heads#L1162
Lets say the exponent is negative and the sign is negative, then we raise
float_flag_inexact and we never reach the code you mentioned here. However,
TriCore HW raises float_flag_invalid as well in that case. This is what I'm
catching with float32_lt_quiet() in the same manner as ftouz.
Hmph. Buggy hardware. You'd better document this special case,
that less-than-zero is detected before rounding.
I presume -0.0 does not raise invalid, that the bug does not extend that far?
r~
