[libc-commits] [libc] [libc][math][c23] Add exp2m1f C23 math function (PR #86996)

[via libc-commits]

================
@@ -0,0 +1,144 @@
+//===-- Implementation of exp2m1f function --------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "src/math/exp2m1f.h"
+#include "src/__support/FPUtil/FEnvImpl.h"
+#include "src/__support/FPUtil/FPBits.h"
+#include "src/__support/FPUtil/PolyEval.h"
+#include "src/__support/FPUtil/multiply_add.h"
+#include "src/__support/FPUtil/rounding_mode.h"
+#include "src/__support/common.h"
+#include "src/__support/macros/optimization.h"
+#include "src/__support/macros/properties/cpu_features.h"
+#include "src/errno/libc_errno.h"
+
+#include "explogxf.h"
+
+namespace LIBC_NAMESPACE {
+
+LLVM_LIBC_FUNCTION(float, exp2m1f, (float x)) {
+  using FPBits = fputil::FPBits<float>;
+  FPBits xbits(x);
+
+  uint32_t x_u = xbits.uintval();
+  uint32_t x_abs = x_u & 0x7fff'ffffU;
+
+  // When |x| >= 128, or x is nan, or |x| <= 2^-5
+  if (LIBC_UNLIKELY(x_abs >= 0x4300'0000U || x_abs <= 0x3d00'0000U)) {
+    // |x| <= 2^-5
+    if (x_abs <= 0x3d00'0000) {
+      // TODO: Handle exceptional values.
+
+      // Minimax polynomial generated by Sollya with:
+      // > display = hexadecimal;
+      // > fpminimax((2^x - 1)/x, 5, [|D...|], [-2^-5, 2^-5]);
+      constexpr double COEFFS[] = {
+          0x1.62e42fefa39f3p-1, 0x1.ebfbdff82c57bp-3,  0x1.c6b08d6f2d7aap-5,
+          0x1.3b2ab6fc92f5dp-7, 0x1.5d897cfe27125p-10, 0x1.43090e61e6af1p-13};
+      double xd = x;
+      double xsq = xd * xd;
+      double c0 = fputil::multiply_add(xd, COEFFS[1], COEFFS[0]);
+      double c1 = fputil::multiply_add(xd, COEFFS[3], COEFFS[2]);
+      double c2 = fputil::multiply_add(xd, COEFFS[5], COEFFS[4]);
+      double p = fputil::polyeval(xsq, c0, c1, c2);
+      return static_cast<float>(p * xd);
+    }
+
+    // x >= 128, or x is nan
+    if (xbits.is_pos()) {
+      if (xbits.is_finite()) {
+        int rounding = fputil::quick_get_round();
+        if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO)
+          return FPBits::max_normal().get_val();
+
+        fputil::set_errno_if_required(ERANGE);
+        fputil::raise_except_if_required(FE_OVERFLOW);
+      }
+
+      // x >= 128 and 2^x - 1 rounds to +inf, or x is +inf or nan
+      return x + FPBits::inf().get_val();
+    }
+
+    // x <= -25
+    if (x <= -25.0f) {
----------------
lntue wrote:

You can benchmark later, but there are several reasons:
- the first one is that inside CPUs, floating point instructions and integer instructions use different ports, and switching values between them does have some hidden cost, like the above we use integer comparisons to omit inf / nan.
- second is that with multiple tight conditions, keeping them as integers will help compilers to optimize better, since the full semantic of floating point is more complicated, such as it's easy for us to see that NaNs cannot get here, but not so easy for the compilers.

And of course everything is just guessing.  Once you finish the implementations and accuracy tests, you can try both with performance tests to see if it has any affects.

https://github.com/llvm/llvm-project/pull/86996