Optimize `absl::Duration` division and modulo: Allow the compiler to inline `time_internal::IDivDuration`, by splitting the slow path to a separate function.

With that change, the compiler can inline the fast path. This is specially important in the context of `Duration::operator%=`, because it allows proving that the return value is unused, therefore avoiding expensive multiplies and divides (e.g. `*q = num_hi / den_hi;`).

```
name                         old cpu/op   new cpu/op   delta
BM_Duration_Modulo           23.1ns ± 0%  22.5ns ± 0%   -2.42%  (p=0.000 n=20+16)
BM_Duration_Modulo_FastPath  7.05ns ± 0%  4.85ns ± 0%  -31.17%  (p=0.000 n=20+20)

name                         old time/op          new time/op          delta
BM_Duration_Modulo           23.1ns ± 0%          22.6ns ± 0%   -2.43%  (p=0.000 n=20+16)
BM_Duration_Modulo_FastPath  7.06ns ± 0%          4.86ns ± 0%  -31.18%  (p=0.000 n=20+20)

name                         old INSTRUCTIONS/op  new INSTRUCTIONS/op  delta
BM_Duration_Modulo              188 ± 0%             178 ± 0%   -5.32%  (p=0.000 n=20+20)
BM_Duration_Modulo_FastPath    84.0 ± 0%            62.0 ± 0%  -26.19%  (p=0.000 n=20+20)

name                         old CYCLES/op        new CYCLES/op        delta
BM_Duration_Modulo             73.8 ± 0%            72.1 ± 0%   -2.27%  (p=0.000 n=19+20)
BM_Duration_Modulo_FastPath    22.5 ± 0%            15.5 ± 0%  -31.13%  (p=0.000 n=19+20)
```

Note: We don't need to expose `absl::time_internal::IDivDuration` at all given that we have a public `absl::IDivDuration`.
PiperOrigin-RevId: 610710635
Change-Id: Ief7c3d5b1c000b397d931e9249edcaef96e7151e

3 files changed, 73 insertions, 54 deletions

diff --git a/absl/time/duration.cc b/absl/time/duration.cc
index 13b268c0..4c25dc40 100644
--- a/absl/time/duration.cc
+++ b/absl/time/duration.cc
@@ -342,19 +342,10 @@ inline bool IDivFastPath(const Duration num, const Duration den, int64_t* q,
 
 }  // namespace
 
-namespace time_internal {
+namespace {
 
-// The 'satq' argument indicates whether the quotient should saturate at the
-// bounds of int64_t.  If it does saturate, the difference will spill over to
-// the remainder.  If it does not saturate, the remainder remain accurate,
-// but the returned quotient will over/underflow int64_t and should not be used.
-int64_t IDivDuration(bool satq, const Duration num, const Duration den,
+int64_t IDivSlowPath(bool satq, const Duration num, const Duration den,
                      Duration* rem) {
-  int64_t q = 0;
-  if (IDivFastPath(num, den, &q, rem)) {
-    return q;
-  }
-
   const bool num_neg = num < ZeroDuration();
   const bool den_neg = den < ZeroDuration();
   const bool quotient_neg = num_neg != den_neg;
@@ -391,7 +382,27 @@ int64_t IDivDuration(bool satq, const Duration num, const Duration den,
   return -static_cast<int64_t>(Uint128Low64(quotient128 - 1) & kint64max) - 1;
 }
 
-}  // namespace time_internal
+// The 'satq' argument indicates whether the quotient should saturate at the
+// bounds of int64_t.  If it does saturate, the difference will spill over to
+// the remainder.  If it does not saturate, the remainder remain accurate,
+// but the returned quotient will over/underflow int64_t and should not be used.
+ABSL_ATTRIBUTE_ALWAYS_INLINE inline int64_t IDivDurationImpl(bool satq,
+                                                             const Duration num,
+                                                             const Duration den,
+                                                             Duration* rem) {
+  int64_t q = 0;
+  if (IDivFastPath(num, den, &q, rem)) {
+    return q;
+  }
+  return IDivSlowPath(satq, num, den, rem);
+}
+
+}  // namespace
+
+int64_t IDivDuration(Duration num, Duration den, Duration* rem) {
+  return IDivDurationImpl(true, num, den,
+                          rem);  // trunc towards zero
+}
 
 //
 // Additive operators.
@@ -475,7 +486,7 @@ Duration& Duration::operator/=(double r) {
 }
 
 Duration& Duration::operator%=(Duration rhs) {
-  time_internal::IDivDuration(false, *this, rhs, this);
+  IDivDurationImpl(false, *this, rhs, this);
   return *this;
 }
 
@@ -501,9 +512,7 @@ double FDivDuration(Duration num, Duration den) {
 // Trunc/Floor/Ceil.
 //
 
-Duration Trunc(Duration d, Duration unit) {
-  return d - (d % unit);
-}
+Duration Trunc(Duration d, Duration unit) { return d - (d % unit); }
 
 Duration Floor(const Duration d, const Duration unit) {
   const absl::Duration td = Trunc(d, unit);
@@ -591,15 +600,9 @@ double ToDoubleMicroseconds(Duration d) {
 double ToDoubleMilliseconds(Duration d) {
   return FDivDuration(d, Milliseconds(1));
 }
-double ToDoubleSeconds(Duration d) {
-  return FDivDuration(d, Seconds(1));
-}
-double ToDoubleMinutes(Duration d) {
-  return FDivDuration(d, Minutes(1));
-}
-double ToDoubleHours(Duration d) {
-  return FDivDuration(d, Hours(1));
-}
+double ToDoubleSeconds(Duration d) { return FDivDuration(d, Seconds(1)); }
+double ToDoubleMinutes(Duration d) { return FDivDuration(d, Minutes(1)); }
+double ToDoubleHours(Duration d) { return FDivDuration(d, Hours(1)); }
 
 timespec ToTimespec(Duration d) {
   timespec ts;
diff --git a/absl/time/duration_benchmark.cc b/absl/time/duration_benchmark.cc
index 56820f37..fdb26bb3 100644
--- a/absl/time/duration_benchmark.cc
+++ b/absl/time/duration_benchmark.cc
@@ -290,6 +290,26 @@ void BM_Duration_IDivDuration_Hours(benchmark::State& state) {
 }
 BENCHMARK(BM_Duration_IDivDuration_Hours);
 
+void BM_Duration_Modulo(benchmark::State& state) {
+  int i = 0;
+  while (state.KeepRunning()) {
+    auto mod = absl::Seconds(i) % absl::Nanoseconds(12345);
+    benchmark::DoNotOptimize(mod);
+    ++i;
+  }
+}
+BENCHMARK(BM_Duration_Modulo);
+
+void BM_Duration_Modulo_FastPath(benchmark::State& state) {
+  int i = 0;
+  while (state.KeepRunning()) {
+    auto mod = absl::Seconds(i) % absl::Milliseconds(1);
+    benchmark::DoNotOptimize(mod);
+    ++i;
+  }
+}
+BENCHMARK(BM_Duration_Modulo_FastPath);
+
 void BM_Duration_ToInt64Nanoseconds(benchmark::State& state) {
   absl::Duration d = absl::Seconds(100000);
   while (state.KeepRunning()) {
diff --git a/absl/time/time.h b/absl/time/time.h
index 37580805..d367ace2 100644
--- a/absl/time/time.h
+++ b/absl/time/time.h
@@ -98,7 +98,6 @@ class Time;      // Defined below
 class TimeZone;  // Defined below
 
 namespace time_internal {
-int64_t IDivDuration(bool satq, Duration num, Duration den, Duration* rem);
 ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixDuration(Duration d);
 ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration ToUnixDuration(Time t);
 ABSL_ATTRIBUTE_CONST_FUNCTION constexpr int64_t GetRepHi(Duration d);
@@ -338,30 +337,6 @@ ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration operator-(Duration lhs,
   return lhs -= rhs;
 }
 
-// Multiplicative Operators
-// Integer operands must be representable as int64_t.
-template <typename T>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration operator*(Duration lhs, T rhs) {
-  return lhs *= rhs;
-}
-template <typename T>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration operator*(T lhs, Duration rhs) {
-  return rhs *= lhs;
-}
-template <typename T>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration operator/(Duration lhs, T rhs) {
-  return lhs /= rhs;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline int64_t operator/(Duration lhs,
-                                                       Duration rhs) {
-  return time_internal::IDivDuration(true, lhs, rhs,
-                                     &lhs);  // trunc towards zero
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration operator%(Duration lhs,
-                                                        Duration rhs) {
-  return lhs %= rhs;
-}
-
 // IDivDuration()
 //
 // Divides a numerator `Duration` by a denominator `Duration`, returning the
@@ -390,10 +365,7 @@ ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration operator%(Duration lhs,
 //   // Here, q would overflow int64_t, so rem accounts for the difference.
 //   int64_t q = absl::IDivDuration(a, b, &rem);
 //   // q == std::numeric_limits<int64_t>::max(), rem == a - b * q
-inline int64_t IDivDuration(Duration num, Duration den, Duration* rem) {
-  return time_internal::IDivDuration(true, num, den,
-                                     rem);  // trunc towards zero
-}
+int64_t IDivDuration(Duration num, Duration den, Duration* rem);
 
 // FDivDuration()
 //
@@ -409,6 +381,30 @@ inline int64_t IDivDuration(Duration num, Duration den, Duration* rem) {
 //   // d == 1.5
 ABSL_ATTRIBUTE_CONST_FUNCTION double FDivDuration(Duration num, Duration den);
 
+// Multiplicative Operators
+// Integer operands must be representable as int64_t.
+template <typename T>
+ABSL_ATTRIBUTE_CONST_FUNCTION Duration operator*(Duration lhs, T rhs) {
+  return lhs *= rhs;
+}
+template <typename T>
+ABSL_ATTRIBUTE_CONST_FUNCTION Duration operator*(T lhs, Duration rhs) {
+  return rhs *= lhs;
+}
+template <typename T>
+ABSL_ATTRIBUTE_CONST_FUNCTION Duration operator/(Duration lhs, T rhs) {
+  return lhs /= rhs;
+}
+ABSL_ATTRIBUTE_CONST_FUNCTION inline int64_t operator/(Duration lhs,
+                                                       Duration rhs) {
+  return IDivDuration(lhs, rhs,
+                      &lhs);  // trunc towards zero
+}
+ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration operator%(Duration lhs,
+                                                        Duration rhs) {
+  return lhs %= rhs;
+}
+
 // ZeroDuration()
 //
 // Returns a zero-length duration. This function behaves just like the default