On 8/31/21 9:39 AM, Luis Pires wrote:
These will be used to implement new decimal floating point
instructions from Power ISA 3.1.
A new argument, prem, was added to divu128/divs128 to receive the
remainder, freeing up phigh to receive the high 64 bits of the
quotient.
For scenarios supported by the previous implementation
(<= 64-bit quotient) with large (> 64-bit) dividends, testing showed
that:
- when dividend >> divisor, the performance of the new implementation
is equivalent to the old one.
- as the dividend and the divisor get closer (e.g. 65-bit dividend and
64-bit divisor), the performance is significantly improved, due to the
smaller number of shift-subtract iterations.
Hmm. I'll note that we have a better divmod primitive in tree, but we aren't using it
here: udiv_qrnnd in include/fpu/softfloat-macros.h.
Given that none of the existing uses require the high part, should we be creating a new
interface? The bug you highlight wrt truncation could be fixed separately.
-void divs128(int64_t *plow, int64_t *phigh, int64_t divisor)
+void divs128(uint64_t *plow, int64_t *phigh, int64_t *prem, int64_t divisor)
{
- int sgn_dvdnd = *phigh < 0;
- int sgn_divsr = divisor < 0;
+ int neg_quotient = 0, neg_remainder = 0;
You might as well use bool.
r~
