Performance improvement when using "QEMU_FLATTEN" with softfloat type conversions

This bug has been copied automatically from: https://bugs.launchpad.net/qemu/+bug/1892081
Reported by 'Ahmed Karaman (ahmedkrmn)' on 2020-08-18 :

Attached below is a matrix multiplication program for double data
types. The program performs the casting operation "(double)rand()"
when generating random numbers.

This operation calls the integer to float softfloat conversion
function "int32_to_float_64".

Adding the "QEMU_FLATTEN" attribute to the function definition
decreases the instructions per call of the function by about 63%.

Attached are before and after performance screenshots from
KCachegrind.