[libc][math] Fix signed zeros for powf when underflow happens.

Author: lntue

libc/src/math/generic/powf.cpp

@@ -855,9 +855,11 @@ LLVM_LIBC_FUNCTION(float, powf, (float x, float y)) {
           : 0.0;
   exp2_hi_mid_dd.hi = exp2_hi_mid;
 
-  return static_cast<float>(
-             powf_double_double(idx_x, dx, y6, lo6_hi, exp2_hi_mid_dd)) +
-         0.0f;
+  double r_dd = powf_double_double(idx_x, dx, y6, lo6_hi, exp2_hi_mid_dd);
+  float r_f = static_cast<float>(r_dd);
+
+  // Only fix signed zeros for exact zeros results.
+  return r_dd != 0 ? r_f : r_f + 0.0f;
 }
 
 } // namespace LIBC_NAMESPACE_DECL

libc/test/src/math/smoke/powf_test.cpp

@@ -190,4 +190,6 @@ TEST_F(LlvmLibcPowfTest, SpecialNumbers) {
                                   FE_UNDERFLOW);
     }
   }
+
+  EXPECT_FP_EQ(-0.0f, LIBC_NAMESPACE::powf(-0.015625f, 25.0f));
 }