On 6/2/21 1:36 PM, Alex Bennée wrote:
Richard Henderson <richard.hender...@linaro.org> writes:
Rename to parts$N_minmax. Combine 3 bool arguments to a bitmask,
return a tri-state value to indicate nan vs unchanged operand.
Introduce ftype_minmax functions as a common optimization point.
Fold bfloat16 expansions into the same macro as the other types.
Signed-off-by: Richard Henderson <richard.hender...@linaro.org>
---
<snip>
- switch (a.cls) {
- case float_class_normal:
- a_exp = a.exp;
- break;
- case float_class_inf:
- a_exp = INT_MAX;
- break;
- case float_class_zero:
- a_exp = INT_MIN;
- break;
- default:
- g_assert_not_reached();
- break;
- }
- switch (b.cls) {
- case float_class_normal:
- b_exp = b.exp;
- break;
- case float_class_inf:
- b_exp = INT_MAX;
- break;
- case float_class_zero:
- b_exp = INT_MIN;
- break;
- default:
- g_assert_not_reached();
- break;
- }
-
<snip>
+ if (unlikely(ab_mask != float_cmask_normal)) {
+ switch (a->cls) {
+ case float_class_normal:
+ break;
+ case float_class_inf:
+ a_exp = INT16_MAX;
We've gone from min/max of our internal representation to INT16_MIN/MAX
- I guess that doesn't break until we get to Float256 but why the change?
Ah, because I've changed from "bool a_less" to "int cmp". I subtract the two
exponents and I don't want that subtraction to overflow.
r~
