Range-based cilk_for

[arvid220u]

Overview

Implements a range-based cilk_for syntax supporting containers with random access iterators.

For example, this PR adds support for syntax on the form:

vector<int> v;
cilk_for (auto x : v) {
    f(x);
}

Previously, you had to do:

vector<int> v;
cilk_for (int i = 0; i < v.size(); i++) {
    f(v[i]);
}

Implementation

We introduce a new AST type, CilkForRangeStmt, taking inspiration from both CilkForStmt and CXXForRangeStmt.

The CilkForRangeStmt internally stores a CXXForRangeStmt, and lets the CXXForRangeStmt handle the parsing of the for loop. It then outputs IR similar to a CilkForStmt. The loop control differs from what one might expect: instead of having the loop variable be an iterator going from container.begin() to container.end() as in a normal for range loop, we emit a loop index of type difference_type going from 0 to container.end() - container.begin(). This makes it possible for the Tapir loop pass to transform the loop into a divide and conquer.

Testing

1. Diagnostics tested in clang/test/Cilk/rangelooptest.cpp.
2. Correct IR output tested in clang/test/Cilk/cilkforrange-ir.cpp.
3. Tested on the example files from arvid220u/llvmurop/examples, involving simple examples including random access iterators and non-random access iterators.
4. Each example above has been compiled using -fsanitize=cilk (Cilksan), reporting 0 determinancy races.
5. Benchmarked on the microbenchmarks from arvid220u/llvmurop/microbenchmarks, being 10% slower on O0, and having no observable difference on any higher optimization levels.