[test-suite] Fix clang build failure due to uniform_real_distribution in SLPVectorization

[jsji]

Summary

The newly added file will cause build failure with clang.

test-suite/MicroBenchmarks/SLPVectorization/Versioning.cpp:10:37:
required from ‘void init_data(T*, unsigned int) [with T = int]’
test-suite/MicroBenchmarks/SLPVectorization/Versioning.cpp:28:12:
required from ‘void benchmark_xor_runtime_checks_pass(benchmark::State&)
[with unsigned int N = 4; T = int]’
test-suite/MicroBenchmarks/SLPVectorization/Versioning.cpp:70:1:
required from here
/usr/include/c++/7/bits/random.h:1704:7: error: static assertion failed:
result_type must be a floating point type

static_assert(std::is_floating_point<_RealType>::value,

[jsji]

This is breaking our internal buildbot.

[Meinersbur]

Did you consider just casting the range to the desired type?

template <typename T> static void init_data(T *A, unsigned N) {
  std::uniform_real_distribution<T> dist((T)-100, (T)100);
  std::mt19937 rng(12345);
  for (unsigned i = 0; i < N; i++)
    A[i] = dist(rng);
}

[jsji]

In D107204#2917943, @Meinersbur wrote:

Did you consider just casting the range to the desired type?

template <typename T> static void init_data(T *A, unsigned N) {
  std::uniform_real_distribution<T> dist((T)-100, (T)100);
  std::mt19937 rng(12345);
  for (unsigned i = 0; i < N; i++)
    A[i] = dist(rng);
}

the problem is uniform_real_distribution require floating result_type ,
it will assert when T is int, so casting won't fix the problem.

[jsji]

Casting the result is OK for me, if that is better for you.

template <typename T> static void init_data(T *A, unsigned N) {
  std::uniform_real_distribution<> dist(-100, 100);
  std::mt19937 rng(12345);
  for (unsigned i = 0; i < N; i++)
    A[i] = (T)dist(rng);
}

[jsji]

Casting the result of distribution only.

[hubert.reinterpretcast]

LGTM; thanks!

MicroBenchmarks/SLPVectorization/Versioning.cpp

13

I think this is fine for the purposes of this benchmark (this doesn't need to be fully generic, e.g., for long double). Could alternatively go with explicitly using uniform_real_distribution<double> for double and uniform_int_distribution<int> for int.

[jsji]

MicroBenchmarks/SLPVectorization/Versioning.cpp

13

:), that was my 1st version, but anyway, committing this version first.

MicroBenchmarks/SLPVectorization/Versioning.cpp

 #include <iostream>
 #include <math.h>
 #include <memory>
 #include <random>
 
 #include "benchmark/benchmark.h"
 
 // Initialize arrays A, B and T with random numbers.
 template <typename T> static void init_data(T *A, unsigned N) {
-  std::uniform_real_distribution<T> dist(-100, 100);
+  std::uniform_real_distribution<> dist(-100, 100);
   std::mt19937 rng(12345);
   for (unsigned i = 0; i < N; i++)
-    A[i] = dist(rng);
+    A[i] = (T)dist(rng);
 }
 
 template <unsigned N, typename T>
 __attribute__((noinline)) void do_xor(T *A, T *B) {
 #pragma clang loop unroll(full)
   for (int i = 0; i < N; i++) {
     A[i] ^= B[i];
   }