[libc-commits] [libc] [libc] Add config option for fast math optimizations (PR #98029)

[Joseph Huber via libc-commits]

https://github.com/jhuber6 updated https://github.com/llvm/llvm-project/pull/98029

>From 814ee280db0a60e647c3a598f59834ffcd8aa1e4 Mon Sep 17 00:00:00 2001
From: Joseph Huber <huberjn at outlook.com>
Date: Mon, 8 Jul 2024 09:37:10 -0500
Subject: [PATCH] [libc] Add config option for fast math optimizations

Summary:
This patch adds `LIBC_COPT_MATH` that allows users to configure the
different math optimizations.
---
 libc/cmake/modules/LLVMLibCCompileOptionRules.cmake | 3 +++
 libc/config/config.json                             | 6 ++++++
 libc/config/gpu/config.json                         | 5 +++++
 libc/docs/configure.rst                             | 2 ++
 libc/src/math/generic/tan.cpp                       | 5 +++--
 5 files changed, 19 insertions(+), 2 deletions(-)

diff --git a/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake b/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
index 6d38bb491044e..58813f50d101c 100644
--- a/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
+++ b/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
@@ -104,6 +104,9 @@ function(_get_common_compile_options output_var flags)
       list(APPEND compile_options "-Wthread-safety")
       list(APPEND compile_options "-Wglobal-constructors")
     endif()
+    if(LIBC_CONF_MATH_OPTIMIZATIONS)
+      list(APPEND compile_options "-DLIBC_MATH=${LIBC_CONF_MATH_OPTIMIZATIONS}")
+    endif()
   elseif(MSVC)
     list(APPEND compile_options "/EHs-c-")
     list(APPEND compile_options "/GR-")
diff --git a/libc/config/config.json b/libc/config/config.json
index 11433c15d762e..cf80f9993e595 100644
--- a/libc/config/config.json
+++ b/libc/config/config.json
@@ -60,5 +60,11 @@
       "value": 1073741824,
       "doc": "Default size for the constinit freelist buffer used for the freelist malloc implementation (default 1o 1GB)."
     }
+  },
+  "math": {
+    "LIBC_CONF_MATH_OPTIMIZATIONS": {
+      "value": 0,
+      "doc": "Configures optimizations for math functions. Values accepted are LIBC_MATH_SKIP_ACCURATE_PASS, LIBC_MATH_SMALL_TABLES, LIBC_MATH_NO_ERRNO, LIBC_MATH_NO_EXCEPT, and LIBC_MATH_FAST."
+    }
   }
 }
diff --git a/libc/config/gpu/config.json b/libc/config/gpu/config.json
index 53f232e31cc8a..73bf810efbac6 100644
--- a/libc/config/gpu/config.json
+++ b/libc/config/gpu/config.json
@@ -12,5 +12,10 @@
     "LIBC_CONF_PRINTF_FLOAT_TO_STR_USE_MEGA_LONG_DOUBLE_TABLE": {
       "value": false
     }
+  },
+  "math": {
+    "LIBC_CONF_MATH_OPTIMIZATIONS": {
+      "value": "(LIBC_MATH_SKIP_ACCURATE_PASS | LIBC_MATH_SMALL_TABLES | LIBC_MATH_NO_ERRNO | LIBC_MATH_NO_EXCEPT)"
+    }
   }
 }
diff --git a/libc/docs/configure.rst b/libc/docs/configure.rst
index 016e2e5aa5876..d309d5b772038 100644
--- a/libc/docs/configure.rst
+++ b/libc/docs/configure.rst
@@ -30,6 +30,8 @@ to learn about the defaults for your platform and target.
     - ``LIBC_CONF_KEEP_FRAME_POINTER``: Keep frame pointer in functions for better debugging experience.
 * **"malloc" options**
     - ``LIBC_CONF_FREELIST_MALLOC_BUFFER_SIZE``: Default size for the constinit freelist buffer used for the freelist malloc implementation (default 1o 1GB).
+* **"math" options**
+    - ``LIBC_CONF_MATH_OPTIMIZATIONS``: Configures optimizations for math functions. Values accepted are LIBC_MATH_SKIP_ACCURATE_PASS, LIBC_MATH_SMALL_TABLES, LIBC_MATH_NO_ERRNO, LIBC_MATH_NO_EXCEPT, and LIBC_MATH_FAST.
 * **"printf" options**
     - ``LIBC_CONF_PRINTF_DISABLE_FIXED_POINT``: Disable printing fixed point values in printf and friends.
     - ``LIBC_CONF_PRINTF_DISABLE_FLOAT``: Disable printing floating point values in printf and friends.
diff --git a/libc/src/math/generic/tan.cpp b/libc/src/math/generic/tan.cpp
index 7f1385c4717dd..d153cc53acc80 100644
--- a/libc/src/math/generic/tan.cpp
+++ b/libc/src/math/generic/tan.cpp
@@ -127,7 +127,8 @@ LIBC_INLINE DoubleDouble tan_eval(const DoubleDouble &u) {
 // Calculation a / b = a * (1/b) for Float128.
 // Using the initial approximation of q ~ (1/b), then apply 2 Newton-Raphson
 // iterations, before multiplying by a.
-[[maybe_unused]] Float128 newton_raphson_div(const Float128 &a, Float128 b, double q) {
+[[maybe_unused]] Float128 newton_raphson_div(const Float128 &a, Float128 b,
+                                             double q) {
   Float128 q0(q);
   constexpr Float128 TWO(2.0);
   b.sign = (b.sign == Sign::POS) ? Sign::NEG : Sign::POS;
@@ -158,7 +159,7 @@ LLVM_LIBC_FUNCTION(double, tan, (double x)) {
       if (LIBC_UNLIKELY(x == 0.0))
         return x;
 
-        // For |x| < 2^-27, |tan(x) - x| < ulp(x)/2.
+      // For |x| < 2^-27, |tan(x) - x| < ulp(x)/2.
 #ifdef LIBC_TARGET_CPU_HAS_FMA
       return fputil::multiply_add(x, 0x1.0p-54, x);
 #else