From 7e40ff9e35699ea14a6addd2ce20cd23be519430 Mon Sep 17 00:00:00 2001
From: Eric Fiselier <eric@efcs.ca>
Date: Mon, 11 Jul 2016 14:58:50 -0600
Subject: [PATCH] Provide a better implementation of DoNotOptimize(...).

This implementation is less likely to ICE compilers, and is more correct.
It also acts as a memory barrier which will help prevent writes to global memory
from being optimized away.
---
 README.md                         | 45 ++++++++++++++++++++++++++++++-
 include/benchmark/benchmark_api.h | 19 +++++++------
 2 files changed, 53 insertions(+), 11 deletions(-)

diff --git a/README.md b/README.md
index e1ac17a0..a0bcc61e 100644
--- a/README.md
+++ b/README.md
@@ -279,7 +279,8 @@ BENCHMARK(BM_ManualTiming)->Range(1, 1<<17)->UseManualTime();
 
 ### Preventing optimisation
 To prevent a value or expression from being optimized away by the compiler
-the `benchmark::DoNotOptimize(...)` function can be used.
+the `benchmark::DoNotOptimize(...)` and `benchmark::ClobberMemory()`
+functions can be used.
 
 ```c++
 static void BM_test(benchmark::State& state) {
@@ -292,6 +293,48 @@ static void BM_test(benchmark::State& state) {
 }
 ```
 
+`DoNotOptimize(<expr>)` forces the  *result* of `<expr>` to be stored in either
+memory or a register. For GNU based compilers it acts as read/write barrier
+for global memory. More specifically it forces the compiler to flush pending
+writes to memory and reload any other values as necessary.
+
+Note that `DoNotOptimize(<expr>)` does not prevent optimizations on `<expr>`
+in any way. `<expr>` may even be removed entirely when the result is already
+known. For example:
+
+```c++
+  /* Example 1: `<expr>` is removed entirely. */
+  int foo(int x) { return x + 42; }
+  while (...) DoNotOptimize(foo(0)); // Optimized to DoNotOptimize(42);
+
+  /*  Example 2: Result of '<expr>' is only reused */
+  int bar(int) __attribute__((const));
+  while (...) DoNotOptimize(bar(0)); // Optimized to:
+  // int __result__ = bar(0);
+  // while (...) DoNotOptimize(__result__);
+```
+
+The second tool for preventing optimizations is `ClobberMemory()`. In essence
+`ClobberMemory()` forces the compiler to perform all pending writes to global
+memory. Memory managed by block scope objects must be "escaped" using
+`DoNotOptimize(...)` before it can be clobbered. In the below example
+`ClobberMemory()` prevents the call to `v.push_back(42)` from being optimized
+away.
+
+```c++
+static void BM_vector_push_back(benchmark::State& state) {
+  while (state.KeepRunning()) {
+    std::vector<int> v;
+    v.reserve(1);
+    benchmark::DoNotOptimize(v.data()); // Allow v.data() to be clobbered.
+    v.push_back(42);
+    benchmark::ClobberMemory(); // Force 42 to be written to memory.
+  }
+}
+```
+
+Note that `ClobberMemory()` is only available for GNU based compilers.
+
 ### Set time unit manually
 If a benchmark runs a few milliseconds it may be hard to visually compare the
 measured times, since the output data is given in nanoseconds per default. In
diff --git a/include/benchmark/benchmark_api.h b/include/benchmark/benchmark_api.h
index f38dc974..664ca2a9 100644
--- a/include/benchmark/benchmark_api.h
+++ b/include/benchmark/benchmark_api.h
@@ -207,25 +207,24 @@ Benchmark* RegisterBenchmarkInternal(Benchmark*);
 
 // The DoNotOptimize(...) function can be used to prevent a value or
 // expression from being optimized away by the compiler. This function is
-// intented to add little to no overhead.
-// See: http://stackoverflow.com/questions/28287064
+// intended to add little to no overhead.
+// See: https://youtu.be/nXaxk27zwlk?t=2441
 #if defined(__GNUC__)
-// TODO(ericwf): Clang has a bug where it tries to always use a register
-// even if value must be stored in memory. This causes codegen to fail.
-// To work around this we remove the "r" modifier so the operand is always
-// loaded into memory.
-// GCC also has a bug where it complains about inconsistent operand constraints
-// when "+rm" is used for a type larger than can fit in a register or two.
-// For now force the operand to memory for both GCC and Clang.
 template <class Tp>
 inline BENCHMARK_ALWAYS_INLINE void DoNotOptimize(Tp const& value) {
-    asm volatile("" : "+m" (const_cast<Tp&>(value)));
+    asm volatile("" : : "g"(value) : "memory");
+}
+// Force the compiler to flush pending writes to global memory. Acts as an
+// effective read/write barrier
+inline BENCHMARK_ALWAYS_INLINE void ClobberMemory() {
+    asm volatile("" : : : "memory");
 }
 #else
 template <class Tp>
 inline BENCHMARK_ALWAYS_INLINE void DoNotOptimize(Tp const& value) {
     internal::UseCharPointer(&reinterpret_cast<char const volatile&>(value));
 }
+// FIXME Add ClobberMemory() for non-gnu compilers
 #endif
 
 // TimeUnit is passed to a benchmark in order to specify the order of magnitude