[clang][docs] Revise documentation for __builtin_reduce_(max|min).

[c8ef]

The function operation described in the document did not match its actual semantic meaning, this patch resolved the problem.

[llvmbot]

@llvm/pr-subscribers-clang

Author: None (c8ef)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/114637.diff

1 Files Affected:

• (modified) clang/docs/LanguageExtensions.rst (+2-6)

diff --git a/clang/docs/LanguageExtensions.rst b/clang/docs/LanguageExtensions.rst
index 10232ff41da15a..b7676bab623ef8 100644
--- a/clang/docs/LanguageExtensions.rst
+++ b/clang/docs/LanguageExtensions.rst
@@ -745,12 +745,8 @@ Let ``VT`` be a vector type and ``ET`` the element type of ``VT``.
 ======================================= ====================================================================== ==================================
          Name                            Operation                                                              Supported element types
 ======================================= ====================================================================== ==================================
- ET __builtin_reduce_max(VT a)           return x or y, whichever is larger; If exactly one argument is         integer and floating point types
-                                         a NaN, return the other argument. If both arguments are NaNs,
-                                         fmax() return a NaN.
- ET __builtin_reduce_min(VT a)           return x or y, whichever is smaller; If exactly one argument           integer and floating point types
-                                         is a NaN, return the other argument. If both arguments are
-                                         NaNs, fmax() return a NaN.
+ ET __builtin_reduce_max(VT a)           return the largest element of the vector.                              integer and floating point types
+ ET __builtin_reduce_min(VT a)           return the smallest element of the vector.                             integer and floating point types
  ET __builtin_reduce_add(VT a)           \+                                                                     integer types
  ET __builtin_reduce_mul(VT a)           \*                                                                     integer types
  ET __builtin_reduce_and(VT a)           &                                                                      integer types

[RKSimon]

Please can you add back the NaN semantics - AFAICT they should always return the largest/smallest non-NAN value if any vector element is non-NAN, otherwise it could return any of the NAN values in the (all NAN) vector. There might be a better way of phrasing that......

[c8ef]

LLVM LangRef states: "This instruction has the same comparison semantics as the ‘llvm.maxnum.*’ intrinsic." Can I add a link to https://llvm.org/docs/LangRef.html#vector-reduction-intrinsics?

[c8ef]

Add pointers to the floating point min/max function.

[arsenm]

Do we have any kind of conformance testing for this?

[c8ef]

I'm not sure if we had any conformance testing before, but since this is just a document change and the patch is keeping the clang builtin document aligned with the llvm intrinsic document, maybe there's no need to add extra tests here?

[arsenm]

Looking at the implementation, this is going to use reduce_fmax intrinsic, which is not the same as reduce_maximum

I think this builtin trying to support floating point is an anti-feature, and if it has to, it should have matched minimum/maximum. But that's not what the implementation is doing

[c8ef]

I apologize for my earlier mistake. The function __builtin_reduce_max behaves differently from __builtin_reduce_maximum when working with floating point numbers. The former will only return nan if all elements are nan, while the latter will return nan if any element in the vector is nan. The most recent patch corrects the wording error and restores the nan semantics. Please take another look.
@arsenm @RKSimon