[libc-commits] [libc] [libc][math] Improve fmul performance by using double-double arithmetic. (PR #107517)

Job Henandez Lara via libc-commits libc-commits at lists.llvm.org
Sat Sep 14 12:38:01 PDT 2024


https://github.com/Jobhdez updated https://github.com/llvm/llvm-project/pull/107517

>From 51364fbfe327624bcb6e09d69a6f6d77d6a3447f Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Thu, 5 Sep 2024 21:26:49 -0700
Subject: [PATCH 01/22] add draft, one more test to go

---
 libc/src/math/generic/CMakeLists.txt |  4 +++-
 libc/src/math/generic/fmul.cpp       | 18 ++++++++++++++++--
 2 files changed, 19 insertions(+), 3 deletions(-)

diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index 350072f4b9649d..5a1ee3b8b83c77 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -2958,7 +2958,9 @@ add_entrypoint_object(
   HDRS
     ../fmul.h
   DEPENDS
-    libc.src.__support.FPUtil.generic.mul
+    libc.hdr.errno_macros
+    libc.hdr.fenv_macros
+    libc.src.__support.FPUtil.double_double
   COMPILE_OPTIONS
     -O3
 )
diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 64c27d6e2f9564..fe8f5b0bb7c45b 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -5,16 +5,30 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-
+#include "hdr/errno_macros.h"
+#include "hdr/fenv_macros.h"
 #include "src/math/fmul.h"
 #include "src/__support/FPUtil/generic/mul.h"
 #include "src/__support/common.h"
 #include "src/__support/macros/config.h"
+#include "src/__support/FPUtil/double_double.h"
 
 namespace LIBC_NAMESPACE_DECL {
-
+  /*
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   return fputil::generic::mul<float>(x, y);
 }
+  */
+LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
+  fputil::DoubleDouble prod = fputil::exact_mult(x, y);
+  if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf_or_nan() || fputil::FPBits<double>(prod.hi).is_zero()))
+    return static_cast<float>(prod.hi);
+  if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf() || fputil::FPBits<double>(prod.hi).is_zero())) {
+    fputil::set_errno_if_required(EDOM);
+    fputil::raise_except_if_required(FE_INVALID);
+    return fputil::FPBits<double>::quiet_nan().get_val();
+  }
+  return static_cast<float>(prod.hi + prod.lo);
+}
 
 } // namespace LIBC_NAMESPACE_DECL

>From 693746c38eeac04a6e908cec51e98b2df54dd3db Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Fri, 6 Sep 2024 16:35:42 -0700
Subject: [PATCH 02/22] format code

---
 libc/src/math/generic/fmul.cpp | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index fe8f5b0bb7c45b..c2536645b593bd 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -5,25 +5,27 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
+#include "src/math/fmul.h"
 #include "hdr/errno_macros.h"
 #include "hdr/fenv_macros.h"
-#include "src/math/fmul.h"
+#include "src/__support/FPUtil/double_double.h"
 #include "src/__support/FPUtil/generic/mul.h"
 #include "src/__support/common.h"
 #include "src/__support/macros/config.h"
-#include "src/__support/FPUtil/double_double.h"
 
 namespace LIBC_NAMESPACE_DECL {
-  /*
+/*
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
-  return fputil::generic::mul<float>(x, y);
+return fputil::generic::mul<float>(x, y);
 }
-  */
+*/
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   fputil::DoubleDouble prod = fputil::exact_mult(x, y);
-  if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf_or_nan() || fputil::FPBits<double>(prod.hi).is_zero()))
+  if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf_or_nan() ||
+                    fputil::FPBits<double>(prod.hi).is_zero()))
     return static_cast<float>(prod.hi);
-  if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf() || fputil::FPBits<double>(prod.hi).is_zero())) {
+  if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf() ||
+                    fputil::FPBits<double>(prod.hi).is_zero())) {
     fputil::set_errno_if_required(EDOM);
     fputil::raise_except_if_required(FE_INVALID);
     return fputil::FPBits<double>::quiet_nan().get_val();

>From f4981c1852b86ee47c099214eeaed91428728029 Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Sat, 7 Sep 2024 20:14:50 -0700
Subject: [PATCH 03/22] add new case

---
 libc/src/math/generic/fmul.cpp         | 17 +++++++++++++++--
 libc/test/src/math/smoke/fmul_test.cpp | 17 +++++++++++++++++
 2 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index c2536645b593bd..28d96448cbb2df 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -21,8 +21,9 @@ return fputil::generic::mul<float>(x, y);
 */
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   fputil::DoubleDouble prod = fputil::exact_mult(x, y);
-  if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf_or_nan() ||
-                    fputil::FPBits<double>(prod.hi).is_zero()))
+  fputil::FPBits<double> hi_bits(prod.hi), lo_bits(prod.lo);
+  
+  if (LIBC_UNLIKELY(hi_bits.is_inf_or_nan() || hi_bits.is_zero()))
     return static_cast<float>(prod.hi);
   if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf() ||
                     fputil::FPBits<double>(prod.hi).is_zero())) {
@@ -30,7 +31,19 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
     fputil::raise_except_if_required(FE_INVALID);
     return fputil::FPBits<double>::quiet_nan().get_val();
   }
+  if (prod.lo == 0.0)
+    return static_cast<float>(prod.hi);
+  
+  if (lo_bits.sign() != hi_bits.sign()) {
+     // Check if sticky bit of hi are all 0
+    constexpr uint64_t STICKY_MASK = 0xFFF'FFFF;  // Lower (52 - 23 - 1 = 28 bits)
+    uint64_t sticky_bits = (hi_bits.uintval() & STICKY_MASK);
+    uint64_t result_bits = (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
+    double result = fputil::FPBits<double>(result_bits).get_val();
+    return static_cast<float>(result);
+}
   return static_cast<float>(prod.hi + prod.lo);
+    
 }
 
 } // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/src/math/smoke/fmul_test.cpp b/libc/test/src/math/smoke/fmul_test.cpp
index 3f6df66456bac5..e4165b23da1389 100644
--- a/libc/test/src/math/smoke/fmul_test.cpp
+++ b/libc/test/src/math/smoke/fmul_test.cpp
@@ -11,3 +11,20 @@
 #include "src/math/fmul.h"
 
 LIST_MUL_TESTS(float, double, LIBC_NAMESPACE::fmul)
+
+TEST_F(LlvmLibcMulTest, SpecialInputs) {
+  constexpr double INPUTS[][2] = {
+    {0x1.0100010002p8, 0x1.fffcp14},
+  };
+
+  constexpr float RESULTS[] = {
+    0x1.00fdfep+23f,
+  };
+
+  constexpr size_t N = sizeof(RESULTS) / sizeof(RESULTS[0]);
+
+  for (size_t i = 0; i < N; ++i) {
+    float result = LIBC_NAMESPACE::fmul(INPUTS[i][0], INPUTS[i][1]);
+    EXPECT_FP_EQ(RESULTS[i], result);
+  }
+}

>From be366e7333f638f2c2e1a9fa9a86740f37da03ea Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Sat, 7 Sep 2024 20:15:16 -0700
Subject: [PATCH 04/22] format code

---
 libc/src/math/generic/fmul.cpp         | 15 ++++++++-------
 libc/test/src/math/smoke/fmul_test.cpp |  4 ++--
 2 files changed, 10 insertions(+), 9 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 28d96448cbb2df..a9de90edd2a7ab 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -22,7 +22,7 @@ return fputil::generic::mul<float>(x, y);
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   fputil::DoubleDouble prod = fputil::exact_mult(x, y);
   fputil::FPBits<double> hi_bits(prod.hi), lo_bits(prod.lo);
-  
+
   if (LIBC_UNLIKELY(hi_bits.is_inf_or_nan() || hi_bits.is_zero()))
     return static_cast<float>(prod.hi);
   if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf() ||
@@ -33,17 +33,18 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   }
   if (prod.lo == 0.0)
     return static_cast<float>(prod.hi);
-  
+
   if (lo_bits.sign() != hi_bits.sign()) {
-     // Check if sticky bit of hi are all 0
-    constexpr uint64_t STICKY_MASK = 0xFFF'FFFF;  // Lower (52 - 23 - 1 = 28 bits)
+    // Check if sticky bit of hi are all 0
+    constexpr uint64_t STICKY_MASK =
+        0xFFF'FFFF; // Lower (52 - 23 - 1 = 28 bits)
     uint64_t sticky_bits = (hi_bits.uintval() & STICKY_MASK);
-    uint64_t result_bits = (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
+    uint64_t result_bits =
+        (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
     double result = fputil::FPBits<double>(result_bits).get_val();
     return static_cast<float>(result);
-}
+  }
   return static_cast<float>(prod.hi + prod.lo);
-    
 }
 
 } // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/src/math/smoke/fmul_test.cpp b/libc/test/src/math/smoke/fmul_test.cpp
index e4165b23da1389..48a5afd247bcfb 100644
--- a/libc/test/src/math/smoke/fmul_test.cpp
+++ b/libc/test/src/math/smoke/fmul_test.cpp
@@ -14,11 +14,11 @@ LIST_MUL_TESTS(float, double, LIBC_NAMESPACE::fmul)
 
 TEST_F(LlvmLibcMulTest, SpecialInputs) {
   constexpr double INPUTS[][2] = {
-    {0x1.0100010002p8, 0x1.fffcp14},
+      {0x1.0100010002p8, 0x1.fffcp14},
   };
 
   constexpr float RESULTS[] = {
-    0x1.00fdfep+23f,
+      0x1.00fdfep+23f,
   };
 
   constexpr size_t N = sizeof(RESULTS) / sizeof(RESULTS[0]);

>From 07331c1086c95ff2e15c14d6e4ff2d2025118fff Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Mon, 9 Sep 2024 19:30:08 -0700
Subject: [PATCH 05/22] run mpfr tests

---
 libc/test/src/math/fmul_test.cpp              | 27 +++++++++++++++++++
 .../math/performance_testing/CMakeLists.txt   |  2 ++
 .../math/performance_testing/fmul_perf.cpp    |  4 ++-
 3 files changed, 32 insertions(+), 1 deletion(-)

diff --git a/libc/test/src/math/fmul_test.cpp b/libc/test/src/math/fmul_test.cpp
index 3f6df66456bac5..c134c3fa299d85 100644
--- a/libc/test/src/math/fmul_test.cpp
+++ b/libc/test/src/math/fmul_test.cpp
@@ -10,4 +10,31 @@
 
 #include "src/math/fmul.h"
 
+#include "test/UnitTest/Test.h"
+#include "utils/MPFRWrapper/MPFRUtils.h"
+
 LIST_MUL_TESTS(float, double, LIBC_NAMESPACE::fmul)
+
+TEST_F(LlvmLibcMulTest, SpecialInputs) {
+  namespace mpfr = LIBC_NAMESPACE::testing::mpfr;
+  double INPUTS[][2] = {
+      {0x1.0100010002p8, 0x1.fffcp14},
+  };
+
+  constexpr float RESULTS[] = {
+      0x1.00fdfep+23f,
+  };
+
+  
+  constexpr size_t N = sizeof(RESULTS) / sizeof(RESULTS[0]);
+
+  for (size_t i = 0; i < N; ++i) {
+    double a = INPUTS[i][0];
+
+    for (int j = 0; j < 180; ++j) {
+      a *= 0.5;
+      mpfr::BinaryInput<double> input{INPUTS[i][0], INPUTS[i][1]};
+      ASSERT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Mul, input, LIBC_NAMESPACE::fmul(INPUTS[i][0], INPUTS[i][1]), 0.5);
+    }
+  }
+}
diff --git a/libc/test/src/math/performance_testing/CMakeLists.txt b/libc/test/src/math/performance_testing/CMakeLists.txt
index ed1b03f3493c7d..60c074a248f72a 100644
--- a/libc/test/src/math/performance_testing/CMakeLists.txt
+++ b/libc/test/src/math/performance_testing/CMakeLists.txt
@@ -484,6 +484,8 @@ add_perf_binary(
   DEPENDS
     .binary_op_single_output_diff
     libc.src.math.fmul
+    libc.src.__support.FPUtil.generic.mul
+    libc.src.__support.FPUtil.fp_bits
   COMPILE_OPTIONS
     -fno-builtin
 )
diff --git a/libc/test/src/math/performance_testing/fmul_perf.cpp b/libc/test/src/math/performance_testing/fmul_perf.cpp
index a215405eb6aa5d..d6156f1519c507 100644
--- a/libc/test/src/math/performance_testing/fmul_perf.cpp
+++ b/libc/test/src/math/performance_testing/fmul_perf.cpp
@@ -7,12 +7,14 @@
 //===----------------------------------------------------------------------===//
 
 #include "BinaryOpSingleOutputPerf.h"
+#include "src/__support/FPUtil/FPBits.h"
+#include "src/__support/FPUtil/generic/mul.h"
 #include "src/math/fmul.h"
 
 static constexpr size_t DOUBLE_ROUNDS = 40;
 
 float fmul_placeholder_binary(double x, double y) {
-  return static_cast<float>(x * y);
+  return LIBC_NAMESPACE::fputil::generic::mul<float>(x, y);
 }
 
 int main() {

>From 8ce2aea18760c7aa69a06776af23784873c2b21e Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Mon, 9 Sep 2024 19:30:34 -0700
Subject: [PATCH 06/22] format code

---
 libc/test/src/math/fmul_test.cpp | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/libc/test/src/math/fmul_test.cpp b/libc/test/src/math/fmul_test.cpp
index c134c3fa299d85..d1caca0c38de20 100644
--- a/libc/test/src/math/fmul_test.cpp
+++ b/libc/test/src/math/fmul_test.cpp
@@ -25,7 +25,6 @@ TEST_F(LlvmLibcMulTest, SpecialInputs) {
       0x1.00fdfep+23f,
   };
 
-  
   constexpr size_t N = sizeof(RESULTS) / sizeof(RESULTS[0]);
 
   for (size_t i = 0; i < N; ++i) {
@@ -34,7 +33,9 @@ TEST_F(LlvmLibcMulTest, SpecialInputs) {
     for (int j = 0; j < 180; ++j) {
       a *= 0.5;
       mpfr::BinaryInput<double> input{INPUTS[i][0], INPUTS[i][1]};
-      ASSERT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Mul, input, LIBC_NAMESPACE::fmul(INPUTS[i][0], INPUTS[i][1]), 0.5);
+      ASSERT_MPFR_MATCH_ALL_ROUNDING(
+          mpfr::Operation::Mul, input,
+          LIBC_NAMESPACE::fmul(INPUTS[i][0], INPUTS[i][1]), 0.5);
     }
   }
 }

>From 2d928d30e606db9caf767978434c036346da35e5 Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Tue, 10 Sep 2024 09:08:23 -0700
Subject: [PATCH 07/22] handle all cases

Co-authored-by: Tue Ly <lntue at google.com>
---
 libc/src/math/generic/fmul.cpp         | 4 +++-
 libc/test/src/math/smoke/fmul_test.cpp | 3 +++
 2 files changed, 6 insertions(+), 1 deletion(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index a9de90edd2a7ab..6d3b4c3016ff2b 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -44,7 +44,9 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
     double result = fputil::FPBits<double>(result_bits).get_val();
     return static_cast<float>(result);
   }
-  return static_cast<float>(prod.hi + prod.lo);
+
+  double result = fputil::FPBits<double>(hi_bits.uintval() | 1).get_val();
+  return static_cast<float>(result);
 }
 
 } // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/src/math/smoke/fmul_test.cpp b/libc/test/src/math/smoke/fmul_test.cpp
index 48a5afd247bcfb..dd0d89143f623d 100644
--- a/libc/test/src/math/smoke/fmul_test.cpp
+++ b/libc/test/src/math/smoke/fmul_test.cpp
@@ -15,10 +15,12 @@ LIST_MUL_TESTS(float, double, LIBC_NAMESPACE::fmul)
 TEST_F(LlvmLibcMulTest, SpecialInputs) {
   constexpr double INPUTS[][2] = {
       {0x1.0100010002p8, 0x1.fffcp14},
+      {0x1.000000b92144p-7, 0x1.62p7},
   };
 
   constexpr float RESULTS[] = {
       0x1.00fdfep+23f,
+      0x1.620002p0f,
   };
 
   constexpr size_t N = sizeof(RESULTS) / sizeof(RESULTS[0]);
@@ -28,3 +30,4 @@ TEST_F(LlvmLibcMulTest, SpecialInputs) {
     EXPECT_FP_EQ(RESULTS[i], result);
   }
 }
+    

>From be463f29dc0fa764b543ec02e4eb7d548ebf9f2b Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Tue, 10 Sep 2024 18:03:01 -0700
Subject: [PATCH 08/22] handle exceptions

---
 libc/src/math/generic/fmul.cpp | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 6d3b4c3016ff2b..836779d0cad8ae 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -24,13 +24,9 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   fputil::FPBits<double> hi_bits(prod.hi), lo_bits(prod.lo);
 
   if (LIBC_UNLIKELY(hi_bits.is_inf_or_nan() || hi_bits.is_zero()))
-    return static_cast<float>(prod.hi);
-  if (LIBC_UNLIKELY(fputil::FPBits<double>(prod.hi).is_inf() ||
-                    fputil::FPBits<double>(prod.hi).is_zero())) {
     fputil::set_errno_if_required(EDOM);
     fputil::raise_except_if_required(FE_INVALID);
-    return fputil::FPBits<double>::quiet_nan().get_val();
-  }
+    return static_cast<float>(prod.hi);
   if (prod.lo == 0.0)
     return static_cast<float>(prod.hi);
 

>From e6c377890e7a466068db99450e498557a59afd6b Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Tue, 10 Sep 2024 18:03:27 -0700
Subject: [PATCH 09/22] format code

---
 libc/src/math/generic/fmul.cpp         | 22 +++++++++++-----------
 libc/test/src/math/smoke/fmul_test.cpp |  1 -
 2 files changed, 11 insertions(+), 12 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 836779d0cad8ae..a1e1c65e768fc5 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -27,18 +27,18 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
     fputil::set_errno_if_required(EDOM);
     fputil::raise_except_if_required(FE_INVALID);
     return static_cast<float>(prod.hi);
-  if (prod.lo == 0.0)
-    return static_cast<float>(prod.hi);
+    if (prod.lo == 0.0)
+      return static_cast<float>(prod.hi);
 
-  if (lo_bits.sign() != hi_bits.sign()) {
-    // Check if sticky bit of hi are all 0
-    constexpr uint64_t STICKY_MASK =
-        0xFFF'FFFF; // Lower (52 - 23 - 1 = 28 bits)
-    uint64_t sticky_bits = (hi_bits.uintval() & STICKY_MASK);
-    uint64_t result_bits =
-        (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
-    double result = fputil::FPBits<double>(result_bits).get_val();
-    return static_cast<float>(result);
+    if (lo_bits.sign() != hi_bits.sign()) {
+      // Check if sticky bit of hi are all 0
+      constexpr uint64_t STICKY_MASK =
+          0xFFF'FFFF; // Lower (52 - 23 - 1 = 28 bits)
+      uint64_t sticky_bits = (hi_bits.uintval() & STICKY_MASK);
+      uint64_t result_bits =
+          (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
+      double result = fputil::FPBits<double>(result_bits).get_val();
+      return static_cast<float>(result);
   }
 
   double result = fputil::FPBits<double>(hi_bits.uintval() | 1).get_val();
diff --git a/libc/test/src/math/smoke/fmul_test.cpp b/libc/test/src/math/smoke/fmul_test.cpp
index dd0d89143f623d..3fcf514bcd9f05 100644
--- a/libc/test/src/math/smoke/fmul_test.cpp
+++ b/libc/test/src/math/smoke/fmul_test.cpp
@@ -30,4 +30,3 @@ TEST_F(LlvmLibcMulTest, SpecialInputs) {
     EXPECT_FP_EQ(RESULTS[i], result);
   }
 }
-    

>From b69581b55c3cfc8aef5d44b1b9323ecfb52d7967 Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Tue, 10 Sep 2024 18:28:17 -0700
Subject: [PATCH 10/22] fix typo

---
 libc/src/math/generic/fmul.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index a1e1c65e768fc5..8b2315792a74c1 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -23,10 +23,11 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   fputil::DoubleDouble prod = fputil::exact_mult(x, y);
   fputil::FPBits<double> hi_bits(prod.hi), lo_bits(prod.lo);
 
-  if (LIBC_UNLIKELY(hi_bits.is_inf_or_nan() || hi_bits.is_zero()))
+  if (LIBC_UNLIKELY(hi_bits.is_inf_or_nan() || hi_bits.is_zero())) {
     fputil::set_errno_if_required(EDOM);
     fputil::raise_except_if_required(FE_INVALID);
     return static_cast<float>(prod.hi);
+  }
     if (prod.lo == 0.0)
       return static_cast<float>(prod.hi);
 

>From 9507d99799925a08d9d1509a0d04555119f54005 Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Wed, 11 Sep 2024 15:29:02 -0700
Subject: [PATCH 11/22] handle exceptions

---
 libc/src/math/generic/fmul.cpp | 89 ++++++++++++++++++++++++++++------
 1 file changed, 73 insertions(+), 16 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 8b2315792a74c1..c0974c6413d552 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -23,23 +23,80 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   fputil::DoubleDouble prod = fputil::exact_mult(x, y);
   fputil::FPBits<double> hi_bits(prod.hi), lo_bits(prod.lo);
 
-  if (LIBC_UNLIKELY(hi_bits.is_inf_or_nan() || hi_bits.is_zero())) {
-    fputil::set_errno_if_required(EDOM);
-    fputil::raise_except_if_required(FE_INVALID);
-    return static_cast<float>(prod.hi);
+  float prod_hif = static_cast<float>(prod.hi);
+  fputil::FPBits<float> hif_bits(prod_hif);
+  using OutFPBits = fputil::FPBits<float>;
+  using OutStorageType = typename OutFPBits::StorageType;
+  using InFPBits = FPBits<double>;
+  using InStorageType = typename InFPBits::StorageType;
+
+  InFPBits x_bits(x);
+  InFPBits y_bits(y);
+
+  
+  Sign result_sign = x_bits.sign() == y_bits.sign() ? Sign::POS : Sign::NEG;
+
+  if (LIBC_UNLIKELY(x_bits.is_inf_or_nan() || y_bits.is_inf_or_nan() ||
+                    x_bits.is_zero() || y_bits.is_zero())) {
+    if (x_bits.is_nan() || y_bits.is_nan()) {
+      if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan())
+        raise_except_if_required(FE_INVALID);
+
+      if (x_bits.is_quiet_nan()) {
+        InStorageType x_payload = x_bits.get_mantissa();
+        x_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN;
+        return OutFPBits::quiet_nan(x_bits.sign(),
+                                    static_cast<OutStorageType>(x_payload))
+            .get_val();
+      }
+
+      if (y_bits.is_quiet_nan()) {
+        InStorageType y_payload = y_bits.get_mantissa();
+        y_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN;
+        return OutFPBits::quiet_nan(y_bits.sign(),
+                                    static_cast<OutStorageType>(y_payload))
+            .get_val();
+      }
+
+      return OutFPBits::quiet_nan().get_val();
+    }
+
+    if (x_bits.is_inf()) {
+      if (y_bits.is_zero()) {
+        set_errno_if_required(EDOM);
+        raise_except_if_required(FE_INVALID);
+        return OutFPBits::quiet_nan().get_val();
+      }
+
+      return OutFPBits::inf(result_sign).get_val();
+    }
+
+    if (y_bits.is_inf()) {
+      if (x_bits.is_zero()) {
+        set_errno_if_required(EDOM);
+        raise_except_if_required(FE_INVALID);
+        return OutFPBits::quiet_nan().get_val();
+      }
+
+      return OutFPBits::inf(result_sign).get_val();
+    }
+
+    // Now either x or y is zero, and the other one is finite.
+    return OutFPBits::zero(result_sign).get_val();
   }
-    if (prod.lo == 0.0)
-      return static_cast<float>(prod.hi);
-
-    if (lo_bits.sign() != hi_bits.sign()) {
-      // Check if sticky bit of hi are all 0
-      constexpr uint64_t STICKY_MASK =
-          0xFFF'FFFF; // Lower (52 - 23 - 1 = 28 bits)
-      uint64_t sticky_bits = (hi_bits.uintval() & STICKY_MASK);
-      uint64_t result_bits =
-          (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
-      double result = fputil::FPBits<double>(result_bits).get_val();
-      return static_cast<float>(result);
+  
+  if (prod.lo == 0.0)
+    return static_cast<float>(prod.hi);
+
+  if (lo_bits.sign() != hi_bits.sign()) {
+    // Check if sticky bit of hi are all 0
+    constexpr uint64_t STICKY_MASK =
+      0xFFF'FFFF; // Lower (52 - 23 - 1 = 28 bits)
+    uint64_t sticky_bits = (hi_bits.uintval() & STICKY_MASK);
+    uint64_t result_bits =
+      (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
+    double result = fputil::FPBits<double>(result_bits).get_val();
+    return static_cast<float>(result);
   }
 
   double result = fputil::FPBits<double>(hi_bits.uintval() | 1).get_val();

>From 59f5656a34cc1c1804cd136d0dfa40a20711ab4e Mon Sep 17 00:00:00 2001
From: Job Hernandez Lara <hj93 at protonmail.com>
Date: Wed, 11 Sep 2024 15:29:42 -0700
Subject: [PATCH 12/22] fix format

---
 libc/src/math/generic/fmul.cpp | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index c0974c6413d552..81bf6e7b9aae08 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -33,7 +33,6 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   InFPBits x_bits(x);
   InFPBits y_bits(y);
 
-  
   Sign result_sign = x_bits.sign() == y_bits.sign() ? Sign::POS : Sign::NEG;
 
   if (LIBC_UNLIKELY(x_bits.is_inf_or_nan() || y_bits.is_inf_or_nan() ||
@@ -84,17 +83,17 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
     // Now either x or y is zero, and the other one is finite.
     return OutFPBits::zero(result_sign).get_val();
   }
-  
+
   if (prod.lo == 0.0)
     return static_cast<float>(prod.hi);
 
   if (lo_bits.sign() != hi_bits.sign()) {
     // Check if sticky bit of hi are all 0
     constexpr uint64_t STICKY_MASK =
-      0xFFF'FFFF; // Lower (52 - 23 - 1 = 28 bits)
+        0xFFF'FFFF; // Lower (52 - 23 - 1 = 28 bits)
     uint64_t sticky_bits = (hi_bits.uintval() & STICKY_MASK);
     uint64_t result_bits =
-      (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
+        (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
     double result = fputil::FPBits<double>(result_bits).get_val();
     return static_cast<float>(result);
   }

>From 5f9ec22fc0abb4667b97af3b26139f36a18aa98f Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Fri, 13 Sep 2024 19:12:38 -0700
Subject: [PATCH 13/22] update

---
 libc/src/math/generic/fmul.cpp | 78 ++++++++++++++++++++++------------
 1 file changed, 50 insertions(+), 28 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 81bf6e7b9aae08..5f83a4dcdd0ff1 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -20,14 +20,15 @@ return fputil::generic::mul<float>(x, y);
 }
 */
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
-  fputil::DoubleDouble prod = fputil::exact_mult(x, y);
-  fputil::FPBits<double> hi_bits(prod.hi), lo_bits(prod.lo);
+  //fputil::DoubleDouble prod = fputil::exact_mult(x, y);
+  //fputil::FPBits<double> hi_bits(prod.hi), lo_bits(prod.lo);
 
+  fputil::DoubleDouble prod = fputil::exact_mult(x, y);
   float prod_hif = static_cast<float>(prod.hi);
   fputil::FPBits<float> hif_bits(prod_hif);
   using OutFPBits = fputil::FPBits<float>;
   using OutStorageType = typename OutFPBits::StorageType;
-  using InFPBits = FPBits<double>;
+  using InFPBits = fputil::FPBits<double>;
   using InStorageType = typename InFPBits::StorageType;
 
   InFPBits x_bits(x);
@@ -35,11 +36,38 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
 
   Sign result_sign = x_bits.sign() == y_bits.sign() ? Sign::POS : Sign::NEG;
 
+    using DoubleBits = fputil::FPBits<double>;
+  using FloatBits = fputil::FPBits<float>;
+  double result = prod.hi;
+  DoubleBits hi_bits(prod.hi), lo_bits(prod.lo);
+  // Check for cases where we need to propagate the sticky bits:
+  constexpr uint64_t STICKY_MASK = 0xFFF'FFF; // Lower (52 - 23 - 1 = 28 bits)
+  uint64_t sticky_bits = (hif_bits.uintval() & STICKY_MASK);
+  if (LIBC_UNLIKELY(sticky_bits == 0)) {
+    // Might need to propagate sticky bits:
+    if (!(lo_bits.is_inf_or_nan() || lo_bits.is_zero())) {
+      // Now prod.lo is nonzero and finite, we need to propagate sticky bits.
+      if (lo_bits.sign() != hi_bits.sign())
+        result = DoubleBits(hi_bits.uintval() - 1).get_val();
+      else
+        result = DoubleBits(hi_bits.uintval() | 1).get_val();
+    }
+  }
+
+  float result_f = static_cast<float>(result);
+  FloatBits rf_bits(result_f);
+  uint32_t rf_exp = rf_bits.get_biased_exponent();
+  
+  if (LIBC_LIKELY(rf_exp > 0 && rf_exp < 2*FloatBits::EXP_BIAS + 1))
+    return result_f;
+
+  // Now result_f is either inf/nan/zero/denormal.
+  // Perform all exceptional checks.
   if (LIBC_UNLIKELY(x_bits.is_inf_or_nan() || y_bits.is_inf_or_nan() ||
                     x_bits.is_zero() || y_bits.is_zero())) {
-    if (x_bits.is_nan() || y_bits.is_nan()) {
+  if (x_bits.is_nan() || y_bits.is_nan()) {
       if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan())
-        raise_except_if_required(FE_INVALID);
+	fputil::raise_except_if_required(FE_INVALID);
 
       if (x_bits.is_quiet_nan()) {
         InStorageType x_payload = x_bits.get_mantissa();
@@ -62,8 +90,8 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
 
     if (x_bits.is_inf()) {
       if (y_bits.is_zero()) {
-        set_errno_if_required(EDOM);
-        raise_except_if_required(FE_INVALID);
+	fputil::set_errno_if_required(EDOM);
+	fputil::raise_except_if_required(FE_INVALID);
         return OutFPBits::quiet_nan().get_val();
       }
 
@@ -72,34 +100,28 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
 
     if (y_bits.is_inf()) {
       if (x_bits.is_zero()) {
-        set_errno_if_required(EDOM);
-        raise_except_if_required(FE_INVALID);
+	fputil::set_errno_if_required(EDOM);
+	fputil::raise_except_if_required(FE_INVALID);
         return OutFPBits::quiet_nan().get_val();
       }
 
       return OutFPBits::inf(result_sign).get_val();
     }
+  }
 
     // Now either x or y is zero, and the other one is finite.
-    return OutFPBits::zero(result_sign).get_val();
-  }
+    if (hif_bits.is_inf()) {
+      fputil::set_errno_if_required(ERANGE);
+      return OutFPBits::zero(result_sign).get_val();
+    }
 
-  if (prod.lo == 0.0)
-    return static_cast<float>(prod.hi);
-
-  if (lo_bits.sign() != hi_bits.sign()) {
-    // Check if sticky bit of hi are all 0
-    constexpr uint64_t STICKY_MASK =
-        0xFFF'FFFF; // Lower (52 - 23 - 1 = 28 bits)
-    uint64_t sticky_bits = (hi_bits.uintval() & STICKY_MASK);
-    uint64_t result_bits =
-        (sticky_bits == 0) ? (hi_bits.uintval() - 1) : hi_bits.uintval();
-    double result = fputil::FPBits<double>(result_bits).get_val();
-    return static_cast<float>(result);
-  }
+    if (hif_bits.get_biased_exponent() == 0.0) {
+      fputil::set_errno_if_required(ERANGE);
+      fputil::raise_except_if_required(FE_UNDERFLOW);
+      return OutFPBits::zero(result_sign).get_val();
+    }
 
-  double result = fputil::FPBits<double>(hi_bits.uintval() | 1).get_val();
-  return static_cast<float>(result);
-}
+    return OutFPBits::zero(result_sign).get_val();
 
-} // namespace LIBC_NAMESPACE_DECL
+}
+}// namespace LIBC_NAMESPACE_DECL

>From 2027e828a61279c31df2042e34f0f337c0ec2e19 Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Fri, 13 Sep 2024 21:08:25 -0700
Subject: [PATCH 14/22] fix tests

---
 libc/src/math/generic/fmul.cpp     | 42 +++++++++++++-----------------
 libc/test/src/math/smoke/MulTest.h | 14 ++++++----
 2 files changed, 27 insertions(+), 29 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 5f83a4dcdd0ff1..191124614413ec 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -12,17 +12,15 @@
 #include "src/__support/FPUtil/generic/mul.h"
 #include "src/__support/common.h"
 #include "src/__support/macros/config.h"
+#include <iostream>
 
 namespace LIBC_NAMESPACE_DECL {
-/*
-LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
-return fputil::generic::mul<float>(x, y);
-}
-*/
+
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
-  //fputil::DoubleDouble prod = fputil::exact_mult(x, y);
-  //fputil::FPBits<double> hi_bits(prod.hi), lo_bits(prod.lo);
 
+  #ifndef LIBC_TARGET_CPU_HAS_FMA
+  return fputil::generic::mul<float>(x, y);
+  #else
   fputil::DoubleDouble prod = fputil::exact_mult(x, y);
   float prod_hif = static_cast<float>(prod.hi);
   fputil::FPBits<float> hif_bits(prod_hif);
@@ -36,13 +34,13 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
 
   Sign result_sign = x_bits.sign() == y_bits.sign() ? Sign::POS : Sign::NEG;
 
-    using DoubleBits = fputil::FPBits<double>;
+  using DoubleBits = fputil::FPBits<double>;
   using FloatBits = fputil::FPBits<float>;
   double result = prod.hi;
   DoubleBits hi_bits(prod.hi), lo_bits(prod.lo);
   // Check for cases where we need to propagate the sticky bits:
   constexpr uint64_t STICKY_MASK = 0xFFF'FFF; // Lower (52 - 23 - 1 = 28 bits)
-  uint64_t sticky_bits = (hif_bits.uintval() & STICKY_MASK);
+  uint64_t sticky_bits = (hi_bits.uintval() & STICKY_MASK);
   if (LIBC_UNLIKELY(sticky_bits == 0)) {
     // Might need to propagate sticky bits:
     if (!(lo_bits.is_inf_or_nan() || lo_bits.is_zero())) {
@@ -57,14 +55,11 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   float result_f = static_cast<float>(result);
   FloatBits rf_bits(result_f);
   uint32_t rf_exp = rf_bits.get_biased_exponent();
-  
-  if (LIBC_LIKELY(rf_exp > 0 && rf_exp < 2*FloatBits::EXP_BIAS + 1))
+  if (LIBC_LIKELY(rf_exp > 0 && rf_exp < 2*FloatBits::EXP_BIAS + 1)) {
     return result_f;
+  }
 
   // Now result_f is either inf/nan/zero/denormal.
-  // Perform all exceptional checks.
-  if (LIBC_UNLIKELY(x_bits.is_inf_or_nan() || y_bits.is_inf_or_nan() ||
-                    x_bits.is_zero() || y_bits.is_zero())) {
   if (x_bits.is_nan() || y_bits.is_nan()) {
       if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan())
 	fputil::raise_except_if_required(FE_INVALID);
@@ -92,6 +87,7 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
       if (y_bits.is_zero()) {
 	fputil::set_errno_if_required(EDOM);
 	fputil::raise_except_if_required(FE_INVALID);
+	
         return OutFPBits::quiet_nan().get_val();
       }
 
@@ -107,21 +103,19 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
 
       return OutFPBits::inf(result_sign).get_val();
     }
-  }
 
     // Now either x or y is zero, and the other one is finite.
     if (hif_bits.is_inf()) {
       fputil::set_errno_if_required(ERANGE);
-      return OutFPBits::zero(result_sign).get_val();
-    }
-
-    if (hif_bits.get_biased_exponent() == 0.0) {
-      fputil::set_errno_if_required(ERANGE);
-      fputil::raise_except_if_required(FE_UNDERFLOW);
-      return OutFPBits::zero(result_sign).get_val();
+      return OutFPBits::inf(result_sign).get_val();
     }
 
-    return OutFPBits::zero(result_sign).get_val();
+    if (x_bits.is_zero() || y_bits.is_zero())
+      return FloatBits::zero(result_sign).get_val();
 
+    fputil::set_errno_if_required(ERANGE);
+    fputil::raise_except_if_required(FE_UNDERFLOW);
+    return result_f;
+}
 }
-}// namespace LIBC_NAMESPACE_DECL
+// namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/src/math/smoke/MulTest.h b/libc/test/src/math/smoke/MulTest.h
index 0c847e39687b72..eac0ef1b60b992 100644
--- a/libc/test/src/math/smoke/MulTest.h
+++ b/libc/test/src/math/smoke/MulTest.h
@@ -66,10 +66,11 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
 
   void test_range_errors(MulFunc func) {
     using namespace LIBC_NAMESPACE::fputil::testing;
-
+    
     if (ForceRoundingMode r(RoundingMode::Nearest); r.success) {
       EXPECT_FP_EQ_WITH_EXCEPTION(inf, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
+      
       EXPECT_MATH_ERRNO(ERANGE);
       EXPECT_FP_EQ_WITH_EXCEPTION(neg_inf, func(neg_max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
@@ -82,8 +83,9 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
                                   func(in.neg_min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
+      
     }
-
+    
     if (ForceRoundingMode r(RoundingMode::TowardZero); r.success) {
       EXPECT_FP_EQ_WITH_EXCEPTION(max_normal, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
@@ -98,9 +100,10 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
                                   func(in.neg_min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
-    }
+      }
 
     if (ForceRoundingMode r(RoundingMode::Downward); r.success) {
+      
       EXPECT_FP_EQ_WITH_EXCEPTION(max_normal, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
       EXPECT_FP_EQ_WITH_EXCEPTION(neg_inf, func(neg_max_normal, max_normal),
@@ -110,12 +113,13 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
       EXPECT_FP_EQ_WITH_EXCEPTION(zero, func(in.min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
+      
       EXPECT_FP_EQ_WITH_EXCEPTION(neg_min_denormal,
                                   func(in.neg_min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
     }
-
+    
     if (ForceRoundingMode r(RoundingMode::Upward); r.success) {
       EXPECT_FP_EQ_WITH_EXCEPTION(inf, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
@@ -132,8 +136,8 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
                                   func(in.neg_min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
+      }
     }
-  }
 
   void test_inexact_results(MulFunc func) {
     InFPBits x_bits = InFPBits::one();

>From e125d74e4cbbc9f4b6eba3e2992cd300a5950d98 Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Fri, 13 Sep 2024 21:08:48 -0700
Subject: [PATCH 15/22] format code

---
 libc/src/math/generic/fmul.cpp     |  4 ++--
 libc/test/src/math/smoke/MulTest.h | 17 ++++++++---------
 2 files changed, 10 insertions(+), 11 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 191124614413ec..9330c8c92ae77b 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -18,9 +18,9 @@ namespace LIBC_NAMESPACE_DECL {
 
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
 
-  #ifndef LIBC_TARGET_CPU_HAS_FMA
+#ifndef LIBC_TARGET_CPU_HAS_FMA
   return fputil::generic::mul<float>(x, y);
-  #else
+#else
   fputil::DoubleDouble prod = fputil::exact_mult(x, y);
   float prod_hif = static_cast<float>(prod.hi);
   fputil::FPBits<float> hif_bits(prod_hif);
diff --git a/libc/test/src/math/smoke/MulTest.h b/libc/test/src/math/smoke/MulTest.h
index eac0ef1b60b992..d586423795cedd 100644
--- a/libc/test/src/math/smoke/MulTest.h
+++ b/libc/test/src/math/smoke/MulTest.h
@@ -66,11 +66,11 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
 
   void test_range_errors(MulFunc func) {
     using namespace LIBC_NAMESPACE::fputil::testing;
-    
+
     if (ForceRoundingMode r(RoundingMode::Nearest); r.success) {
       EXPECT_FP_EQ_WITH_EXCEPTION(inf, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
-      
+
       EXPECT_MATH_ERRNO(ERANGE);
       EXPECT_FP_EQ_WITH_EXCEPTION(neg_inf, func(neg_max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
@@ -83,9 +83,8 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
                                   func(in.neg_min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
-      
     }
-    
+
     if (ForceRoundingMode r(RoundingMode::TowardZero); r.success) {
       EXPECT_FP_EQ_WITH_EXCEPTION(max_normal, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
@@ -100,10 +99,10 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
                                   func(in.neg_min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
-      }
+    }
 
     if (ForceRoundingMode r(RoundingMode::Downward); r.success) {
-      
+
       EXPECT_FP_EQ_WITH_EXCEPTION(max_normal, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
       EXPECT_FP_EQ_WITH_EXCEPTION(neg_inf, func(neg_max_normal, max_normal),
@@ -113,13 +112,13 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
       EXPECT_FP_EQ_WITH_EXCEPTION(zero, func(in.min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
-      
+
       EXPECT_FP_EQ_WITH_EXCEPTION(neg_min_denormal,
                                   func(in.neg_min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
     }
-    
+
     if (ForceRoundingMode r(RoundingMode::Upward); r.success) {
       EXPECT_FP_EQ_WITH_EXCEPTION(inf, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
@@ -136,8 +135,8 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
                                   func(in.neg_min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
-      }
     }
+  }
 
   void test_inexact_results(MulFunc func) {
     InFPBits x_bits = InFPBits::one();

>From 801eb3938894ea8615c9a0d3cc1d737f23da14a4 Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Fri, 13 Sep 2024 22:04:18 -0700
Subject: [PATCH 16/22] address review

---
 libc/src/math/generic/fmul.cpp                      |  3 +++
 libc/test/src/math/fmul_test.cpp                    | 13 ++++---------
 .../performance_testing/BinaryOpSingleOutputPerf.h  |  1 -
 libc/test/src/math/smoke/MulTest.h                  |  3 ---
 4 files changed, 7 insertions(+), 13 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 9330c8c92ae77b..f48ad2bc0c1278 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -116,6 +116,9 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
     fputil::set_errno_if_required(ERANGE);
     fputil::raise_except_if_required(FE_UNDERFLOW);
     return result_f;
+
+    #endif
+    
 }
 }
 // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/src/math/fmul_test.cpp b/libc/test/src/math/fmul_test.cpp
index d1caca0c38de20..01898cd6fd50f1 100644
--- a/libc/test/src/math/fmul_test.cpp
+++ b/libc/test/src/math/fmul_test.cpp
@@ -19,23 +19,18 @@ TEST_F(LlvmLibcMulTest, SpecialInputs) {
   namespace mpfr = LIBC_NAMESPACE::testing::mpfr;
   double INPUTS[][2] = {
       {0x1.0100010002p8, 0x1.fffcp14},
+      {0x1.000000b92144p-7, 0x1.62p7},
   };
 
-  constexpr float RESULTS[] = {
-      0x1.00fdfep+23f,
-  };
-
-  constexpr size_t N = sizeof(RESULTS) / sizeof(RESULTS[0]);
-
-  for (size_t i = 0; i < N; ++i) {
+  for (size_t i = 0; i < 2; ++i) {
     double a = INPUTS[i][0];
 
     for (int j = 0; j < 180; ++j) {
       a *= 0.5;
-      mpfr::BinaryInput<double> input{INPUTS[i][0], INPUTS[i][1]};
+      mpfr::BinaryInput<double> input{a, INPUTS[i][1]};
       ASSERT_MPFR_MATCH_ALL_ROUNDING(
           mpfr::Operation::Mul, input,
-          LIBC_NAMESPACE::fmul(INPUTS[i][0], INPUTS[i][1]), 0.5);
+          LIBC_NAMESPACE::fmul(a, INPUTS[i][1]), 0.5);
     }
   }
 }
diff --git a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
index 98a1813bd7b54a..7a0ab5e74b0277 100644
--- a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
+++ b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
@@ -7,7 +7,6 @@
 //===----------------------------------------------------------------------===//
 
 #include "src/__support/CPP/algorithm.h"
-#include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/macros/config.h"
 #include "test/src/math/performance_testing/Timer.h"
 
diff --git a/libc/test/src/math/smoke/MulTest.h b/libc/test/src/math/smoke/MulTest.h
index d586423795cedd..0c847e39687b72 100644
--- a/libc/test/src/math/smoke/MulTest.h
+++ b/libc/test/src/math/smoke/MulTest.h
@@ -70,7 +70,6 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
     if (ForceRoundingMode r(RoundingMode::Nearest); r.success) {
       EXPECT_FP_EQ_WITH_EXCEPTION(inf, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
-
       EXPECT_MATH_ERRNO(ERANGE);
       EXPECT_FP_EQ_WITH_EXCEPTION(neg_inf, func(neg_max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
@@ -102,7 +101,6 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
     }
 
     if (ForceRoundingMode r(RoundingMode::Downward); r.success) {
-
       EXPECT_FP_EQ_WITH_EXCEPTION(max_normal, func(max_normal, max_normal),
                                   FE_OVERFLOW | FE_INEXACT);
       EXPECT_FP_EQ_WITH_EXCEPTION(neg_inf, func(neg_max_normal, max_normal),
@@ -112,7 +110,6 @@ class MulTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
       EXPECT_FP_EQ_WITH_EXCEPTION(zero, func(in.min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);
       EXPECT_MATH_ERRNO(ERANGE);
-
       EXPECT_FP_EQ_WITH_EXCEPTION(neg_min_denormal,
                                   func(in.neg_min_denormal, in.min_denormal),
                                   FE_UNDERFLOW | FE_INEXACT);

>From 8c3af580e23eb193d6235c0ebf9eaeebf80f9c31 Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Fri, 13 Sep 2024 22:11:39 -0700
Subject: [PATCH 17/22] address review

---
 .../test/src/math/performance_testing/BinaryOpSingleOutputPerf.h | 1 +
 libc/test/src/math/performance_testing/fmul_perf.cpp             | 1 -
 2 files changed, 1 insertion(+), 1 deletion(-)

diff --git a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
index 7a0ab5e74b0277..8ddeaf7c5ed813 100644
--- a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
+++ b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
@@ -8,6 +8,7 @@
 
 #include "src/__support/CPP/algorithm.h"
 #include "src/__support/macros/config.h"
+#include "src/__support/FPUtil/FPBits.h"
 #include "test/src/math/performance_testing/Timer.h"
 
 #include <cstddef>
diff --git a/libc/test/src/math/performance_testing/fmul_perf.cpp b/libc/test/src/math/performance_testing/fmul_perf.cpp
index d6156f1519c507..f15cfafbf29451 100644
--- a/libc/test/src/math/performance_testing/fmul_perf.cpp
+++ b/libc/test/src/math/performance_testing/fmul_perf.cpp
@@ -7,7 +7,6 @@
 //===----------------------------------------------------------------------===//
 
 #include "BinaryOpSingleOutputPerf.h"
-#include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/FPUtil/generic/mul.h"
 #include "src/math/fmul.h"
 

>From 0deb4a9eed45a47ce3c5633ee06896081567e17c Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Fri, 13 Sep 2024 22:12:22 -0700
Subject: [PATCH 18/22] format code

---
 libc/src/math/generic/fmul.cpp                | 95 +++++++++----------
 libc/test/src/math/fmul_test.cpp              |  6 +-
 .../BinaryOpSingleOutputPerf.h                |  2 +-
 3 files changed, 51 insertions(+), 52 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index f48ad2bc0c1278..dbc6439484e52a 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -55,70 +55,69 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   float result_f = static_cast<float>(result);
   FloatBits rf_bits(result_f);
   uint32_t rf_exp = rf_bits.get_biased_exponent();
-  if (LIBC_LIKELY(rf_exp > 0 && rf_exp < 2*FloatBits::EXP_BIAS + 1)) {
+  if (LIBC_LIKELY(rf_exp > 0 && rf_exp < 2 * FloatBits::EXP_BIAS + 1)) {
     return result_f;
   }
 
   // Now result_f is either inf/nan/zero/denormal.
   if (x_bits.is_nan() || y_bits.is_nan()) {
-      if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan())
-	fputil::raise_except_if_required(FE_INVALID);
-
-      if (x_bits.is_quiet_nan()) {
-        InStorageType x_payload = x_bits.get_mantissa();
-        x_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN;
-        return OutFPBits::quiet_nan(x_bits.sign(),
-                                    static_cast<OutStorageType>(x_payload))
-            .get_val();
-      }
-
-      if (y_bits.is_quiet_nan()) {
-        InStorageType y_payload = y_bits.get_mantissa();
-        y_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN;
-        return OutFPBits::quiet_nan(y_bits.sign(),
-                                    static_cast<OutStorageType>(y_payload))
-            .get_val();
-      }
-
-      return OutFPBits::quiet_nan().get_val();
+    if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan())
+      fputil::raise_except_if_required(FE_INVALID);
+
+    if (x_bits.is_quiet_nan()) {
+      InStorageType x_payload = x_bits.get_mantissa();
+      x_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN;
+      return OutFPBits::quiet_nan(x_bits.sign(),
+                                  static_cast<OutStorageType>(x_payload))
+          .get_val();
     }
 
-    if (x_bits.is_inf()) {
-      if (y_bits.is_zero()) {
-	fputil::set_errno_if_required(EDOM);
-	fputil::raise_except_if_required(FE_INVALID);
-	
-        return OutFPBits::quiet_nan().get_val();
-      }
-
-      return OutFPBits::inf(result_sign).get_val();
+    if (y_bits.is_quiet_nan()) {
+      InStorageType y_payload = y_bits.get_mantissa();
+      y_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN;
+      return OutFPBits::quiet_nan(y_bits.sign(),
+                                  static_cast<OutStorageType>(y_payload))
+          .get_val();
     }
 
-    if (y_bits.is_inf()) {
-      if (x_bits.is_zero()) {
-	fputil::set_errno_if_required(EDOM);
-	fputil::raise_except_if_required(FE_INVALID);
-        return OutFPBits::quiet_nan().get_val();
-      }
+    return OutFPBits::quiet_nan().get_val();
+  }
+
+  if (x_bits.is_inf()) {
+    if (y_bits.is_zero()) {
+      fputil::set_errno_if_required(EDOM);
+      fputil::raise_except_if_required(FE_INVALID);
 
-      return OutFPBits::inf(result_sign).get_val();
+      return OutFPBits::quiet_nan().get_val();
     }
 
-    // Now either x or y is zero, and the other one is finite.
-    if (hif_bits.is_inf()) {
-      fputil::set_errno_if_required(ERANGE);
-      return OutFPBits::inf(result_sign).get_val();
+    return OutFPBits::inf(result_sign).get_val();
+  }
+
+  if (y_bits.is_inf()) {
+    if (x_bits.is_zero()) {
+      fputil::set_errno_if_required(EDOM);
+      fputil::raise_except_if_required(FE_INVALID);
+      return OutFPBits::quiet_nan().get_val();
     }
 
-    if (x_bits.is_zero() || y_bits.is_zero())
-      return FloatBits::zero(result_sign).get_val();
+    return OutFPBits::inf(result_sign).get_val();
+  }
 
+  // Now either x or y is zero, and the other one is finite.
+  if (hif_bits.is_inf()) {
     fputil::set_errno_if_required(ERANGE);
-    fputil::raise_except_if_required(FE_UNDERFLOW);
-    return result_f;
+    return OutFPBits::inf(result_sign).get_val();
+  }
 
-    #endif
-    
-}
+  if (x_bits.is_zero() || y_bits.is_zero())
+    return FloatBits::zero(result_sign).get_val();
+
+  fputil::set_errno_if_required(ERANGE);
+  fputil::raise_except_if_required(FE_UNDERFLOW);
+  return result_f;
+
+#endif
 }
+} // namespace LIBC_NAMESPACE_DECL
 // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/src/math/fmul_test.cpp b/libc/test/src/math/fmul_test.cpp
index 01898cd6fd50f1..488e087a325205 100644
--- a/libc/test/src/math/fmul_test.cpp
+++ b/libc/test/src/math/fmul_test.cpp
@@ -28,9 +28,9 @@ TEST_F(LlvmLibcMulTest, SpecialInputs) {
     for (int j = 0; j < 180; ++j) {
       a *= 0.5;
       mpfr::BinaryInput<double> input{a, INPUTS[i][1]};
-      ASSERT_MPFR_MATCH_ALL_ROUNDING(
-          mpfr::Operation::Mul, input,
-          LIBC_NAMESPACE::fmul(a, INPUTS[i][1]), 0.5);
+      ASSERT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Mul, input,
+                                     LIBC_NAMESPACE::fmul(a, INPUTS[i][1]),
+                                     0.5);
     }
   }
 }
diff --git a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
index 8ddeaf7c5ed813..98a1813bd7b54a 100644
--- a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
+++ b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
@@ -7,8 +7,8 @@
 //===----------------------------------------------------------------------===//
 
 #include "src/__support/CPP/algorithm.h"
-#include "src/__support/macros/config.h"
 #include "src/__support/FPUtil/FPBits.h"
+#include "src/__support/macros/config.h"
 #include "test/src/math/performance_testing/Timer.h"
 
 #include <cstddef>

>From 2161862d8dc0a2fda5487ea4854b7592f5c0433c Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Sat, 14 Sep 2024 12:13:49 -0700
Subject: [PATCH 19/22] address review

---
 libc/src/math/generic/fmul.cpp | 53 ++++++++++++++++------------------
 1 file changed, 25 insertions(+), 28 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index dbc6439484e52a..ea5f7d928c0a6d 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -18,24 +18,22 @@ namespace LIBC_NAMESPACE_DECL {
 
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
 
+  /* Without FMA instructions, fputil::exact_mult is not
+  correctly rounded for all rounding modes, so we fall
+  back to the generic `fmul` implementation
+  */ 
 #ifndef LIBC_TARGET_CPU_HAS_FMA
   return fputil::generic::mul<float>(x, y);
 #else
   fputil::DoubleDouble prod = fputil::exact_mult(x, y);
-  float prod_hif = static_cast<float>(prod.hi);
-  fputil::FPBits<float> hif_bits(prod_hif);
-  using OutFPBits = fputil::FPBits<float>;
-  using OutStorageType = typename OutFPBits::StorageType;
-  using InFPBits = fputil::FPBits<double>;
-  using InStorageType = typename InFPBits::StorageType;
-
-  InFPBits x_bits(x);
-  InFPBits y_bits(y);
-
-  Sign result_sign = x_bits.sign() == y_bits.sign() ? Sign::POS : Sign::NEG;
-
   using DoubleBits = fputil::FPBits<double>;
+  using DoubleStorageType = typename DoubleBits::StorageType;
   using FloatBits = fputil::FPBits<float>;
+  using FloatStorageType = typename FloatBits::StorageType;
+  DoubleBits x_bits(x);
+  DoubleBits y_bits(y);
+
+  Sign result_sign = x_bits.sign() == y_bits.sign() ? Sign::POS : Sign::NEG;
   double result = prod.hi;
   DoubleBits hi_bits(prod.hi), lo_bits(prod.lo);
   // Check for cases where we need to propagate the sticky bits:
@@ -65,22 +63,22 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
       fputil::raise_except_if_required(FE_INVALID);
 
     if (x_bits.is_quiet_nan()) {
-      InStorageType x_payload = x_bits.get_mantissa();
-      x_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN;
-      return OutFPBits::quiet_nan(x_bits.sign(),
-                                  static_cast<OutStorageType>(x_payload))
+      DoubleStorageType x_payload = x_bits.get_mantissa();
+      x_payload >>= DoubleBits::FRACTION_LEN - FloatBits::FRACTION_LEN;
+      return FloatBits::quiet_nan(x_bits.sign(),
+                                  static_cast<FloatStorageType>(x_payload))
           .get_val();
     }
 
     if (y_bits.is_quiet_nan()) {
-      InStorageType y_payload = y_bits.get_mantissa();
-      y_payload >>= InFPBits::FRACTION_LEN - OutFPBits::FRACTION_LEN;
-      return OutFPBits::quiet_nan(y_bits.sign(),
-                                  static_cast<OutStorageType>(y_payload))
+      DoubleStorageType y_payload = y_bits.get_mantissa();
+      y_payload >>= DoubleBits::FRACTION_LEN - FloatBits::FRACTION_LEN;
+      return FloatBits::quiet_nan(y_bits.sign(),
+                                  static_cast<FloatStorageType>(y_payload))
           .get_val();
     }
 
-    return OutFPBits::quiet_nan().get_val();
+    return FloatBits::quiet_nan().get_val();
   }
 
   if (x_bits.is_inf()) {
@@ -88,26 +86,26 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
       fputil::set_errno_if_required(EDOM);
       fputil::raise_except_if_required(FE_INVALID);
 
-      return OutFPBits::quiet_nan().get_val();
+      return FloatBits::quiet_nan().get_val();
     }
 
-    return OutFPBits::inf(result_sign).get_val();
+    return FloatBits::inf(result_sign).get_val();
   }
 
   if (y_bits.is_inf()) {
     if (x_bits.is_zero()) {
       fputil::set_errno_if_required(EDOM);
       fputil::raise_except_if_required(FE_INVALID);
-      return OutFPBits::quiet_nan().get_val();
+      return FloatBits::quiet_nan().get_val();
     }
 
-    return OutFPBits::inf(result_sign).get_val();
+    return FloatBits::inf(result_sign).get_val();
   }
 
   // Now either x or y is zero, and the other one is finite.
-  if (hif_bits.is_inf()) {
+  if (rf_bits.is_inf()) {
     fputil::set_errno_if_required(ERANGE);
-    return OutFPBits::inf(result_sign).get_val();
+    return FloatBits::inf(result_sign).get_val();
   }
 
   if (x_bits.is_zero() || y_bits.is_zero())
@@ -120,4 +118,3 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
 #endif
 }
 } // namespace LIBC_NAMESPACE_DECL
-// namespace LIBC_NAMESPACE_DECL

>From fa5a8af5e0e022301ab4cbd0f790081ee66034a2 Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Sat, 14 Sep 2024 12:14:10 -0700
Subject: [PATCH 20/22] format code

---
 libc/src/math/generic/fmul.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index ea5f7d928c0a6d..a8d82ee0bd07eb 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -21,7 +21,7 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   /* Without FMA instructions, fputil::exact_mult is not
   correctly rounded for all rounding modes, so we fall
   back to the generic `fmul` implementation
-  */ 
+  */
 #ifndef LIBC_TARGET_CPU_HAS_FMA
   return fputil::generic::mul<float>(x, y);
 #else

>From 78d388c74f1b29888451936f8af23ac76dd4ae14 Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Sat, 14 Sep 2024 12:37:12 -0700
Subject: [PATCH 21/22] address review

---
 libc/src/math/generic/fmul.cpp | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index a8d82ee0bd07eb..2454359e717481 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -12,16 +12,15 @@
 #include "src/__support/FPUtil/generic/mul.h"
 #include "src/__support/common.h"
 #include "src/__support/macros/config.h"
-#include <iostream>
 
 namespace LIBC_NAMESPACE_DECL {
 
 LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
 
-  /* Without FMA instructions, fputil::exact_mult is not
-  correctly rounded for all rounding modes, so we fall
-  back to the generic `fmul` implementation
-  */
+  // Without FMA instructions, fputil::exact_mult is not
+  // correctly rounded for all rounding modes, so we fall
+  // back to the generic `fmul` implementation
+   
 #ifndef LIBC_TARGET_CPU_HAS_FMA
   return fputil::generic::mul<float>(x, y);
 #else

>From f19954caffe70f6406918f5fe948eb13312feba9 Mon Sep 17 00:00:00 2001
From: Job Hernandez <hj93 at protonmail.com>
Date: Sat, 14 Sep 2024 12:37:34 -0700
Subject: [PATCH 22/22] format code

---
 libc/src/math/generic/fmul.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/libc/src/math/generic/fmul.cpp b/libc/src/math/generic/fmul.cpp
index 2454359e717481..e759e48cd6989a 100644
--- a/libc/src/math/generic/fmul.cpp
+++ b/libc/src/math/generic/fmul.cpp
@@ -20,7 +20,7 @@ LLVM_LIBC_FUNCTION(float, fmul, (double x, double y)) {
   // Without FMA instructions, fputil::exact_mult is not
   // correctly rounded for all rounding modes, so we fall
   // back to the generic `fmul` implementation
-   
+
 #ifndef LIBC_TARGET_CPU_HAS_FMA
   return fputil::generic::mul<float>(x, y);
 #else



More information about the libc-commits mailing list