Workaround intel bug #3652
Workaround intel bug #3652
Conversation
Potential workaround / restructure for the intel bug that is the cause of fmtlib#3645. Make the variable in the external struct instead an embedded static constexpr variable in the only function that uses the variable.
|
Pull request not correct for certain c++ std levels. Will let someone with more knowledge propose a solution. |
include/fmt/format.h
Outdated
| // For checking rounding thresholds. | ||
| // The kth entry is chosen to be the smallest integer such that the | ||
| // upper 32-bits of 10^(k+1) times it is strictly bigger than 5 * 10^k. | ||
| static constexpr uint32_t fractional_part_rounding_thresholds[8] = { |
There was a problem hiding this comment.
You need to drop static here:
/Users/runner/work/fmt/fmt/include/fmt/format.h:3268:29: error: definition of a static variable in a constexpr function is a C++2b extension [-Werror,-Wc++2b-extensions]
static constexpr uint32_t fractional_part_rounding_thresholds[8] = {
^
Also please move to a separate function to avoid parametrization on Float, something like
FMT_CONSTEXPR uint32_t fractional_part_rounding_threshold(int index) {
constexpr uint32_t thresholds[] = {...};
return thresholds[index];
}
There was a problem hiding this comment.
@vitaut
Will a non-static constexpr array be optimized by the compiler?
There was a problem hiding this comment.
There was a problem hiding this comment.
There was a problem hiding this comment.
I don't think we should revert this PR but ideas/PRs to optimize it would be welcome.
There was a problem hiding this comment.
@vitaut
Solution for GCC, Clang: https://godbolt.org/z/ezj9GhnYW
But not for MSVC (current code: https://godbolt.org/z/5Pxn3zMfT)
There was a problem hiding this comment.
I have another idea: 06b2038.
There was a problem hiding this comment.
@vitaut Solution for GCC, Clang: https://godbolt.org/z/ezj9GhnYW But not for MSVC (current code: https://godbolt.org/z/5Pxn3zMfT)
That looks good. I had something similar but didn't think of the unreachable to quiet the warnings and keep the optimization.
Moved variable out of function to avoid specialization on Float. Made it a separate function that is called from format_float.
|
Thanks |
* Workaround intel bug Potential workaround / restructure for the intel bug that is the cause of fmtlib#3645. Make the variable in the external struct instead an embedded static constexpr variable in the only function that uses the variable. * Finish the proposed change -- remove struct accessor * Refactor proposed intel fix. Moved variable out of function to avoid specialization on Float. Made it a separate function that is called from format_float. * Fix incorrect function name. * Add missing inline.
Potential workaround / restructure for the intel bug that is the cause of #3645.
Make the variable in the external struct instead an embedded static constexpr variable in the only function that uses the variable.