[libc-commits] [PATCH] D117590: [libc] Implement correct rounding with all rounding modes for hypot functions.

[Tue Ly via Phabricator via libc-commits]

lntue added a comment.

In D117590#3258279 <https://reviews.llvm.org/D117590#3258279>, @zimmermann6 wrote:

> I'm ok with the new revision. However I see there are still some calls to get_round(). Did you try to replace them by floating-point operations?
>
> You might also want to add the following hard-to-round cases (for binary32) in your test cases:
>
>   /* the following are hard-to-round cases with many identical bits after       
>      the round bit */
>   {0x1.900004p+34,0x1.400002p+23}, /* 45 identical bits */
>   {0x1.05555p+34,0x1.bffffep+23},  /* 44 identical bits */
>   {0x1.e5fffap+34,0x1.affffep+23}, /* 45 identical bits */
>   {0x1.260002p+34,0x1.500002p+23}, /* 45 identical bits */
>   {0x1.fffffap+34,0x1.fffffep+23}, /* 45 identical bits */
>   {0x1.8ffffap+34,0x1.3ffffep+23}, /* 45 identical bits */
>   {0x1.87fffcp+35,0x1.bffffep+23}, /* 47 identical bits */
>
> By the way, none of the other libraries is correctly rounded for binary32, here are the corresponding worst cases:
>
>   /* hypot(x,y) */
>   {0x1.b8e50ap-52,-0x1.db1e78p-64},   /* GNU libc       0.500001 */
>   {0x1.03b54cp-33,0x1.6ca6bep-45},    /* icc            0.500001 */
>   {0x1.e2eff6p+97,-0x1.044cb2p+108},  /* AMD LibM       0.500001 */
>   {-0x1.6b05c4p-127,0x1.6b3146p-126}, /* Newlib         1.20805 */
>   {-0x1.6b05c4p-127,0x1.6b3146p-126}, /* OpenLibm       1.20805 */
>   {0x1.26b188p-127,-0x1.a4f2fp-128},  /* Musl           0.926707 */
>   {0x1.e2eff6p+97,-0x1.044cb2p+108},  /* Darwin 20.4.0  0.500001 */

Thanks Paul for the test cases!  I've added these and the other 1.2k cases in your attachment to the test.

I have also tried replacing the final get_round with the corresponding floating point operations, but for some reason, -O3 always partially optimized it away even with -frounding-math flag together, and that make the results incorrect.
If I change the flag to -O2, then it is correct, but I see virtually no performance improvement.  So I decided to keep the get_round() at the end so that it is easier to read and less sensitive to optimization flags.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D117590/new/

https://reviews.llvm.org/D117590