Add semantics validation for cBench benchmarks.

Add semantics validation for cBench benchmarks. This is achieved by
adding a new validation callback mechanism that, when invoked,
compiles the given cBench benchmark to a binary and executes it using
prepared datasets. The output of the program, along with any generated
output files, is differential tested against a copy of the program
compiled without optimizations. A change in program behavior that is
detected by this mechanism is reported.

Calling `compiler_gym.validate_state()` on a benchmark that supports
semantics validation will automatically run it.

The core of the implementation is in
compiler_gym/envs/llvm/dataset.py. It defines a set of library
functions so that these validation callbacks can be defined ad-hoc for
cBench in quite a succinct form, e.g.:

    validator(
        benchmark="benchmark://cBench-v0/ghostscript",
        cmd="$BIN -sDEVICE=ppm -dNOPAUSE -dQUIET -sOutputFile=output.ppm -- 1.ps",
        data=["office_data/1.ps"],
        outs=["output.ppm"],
        linkopts=["-lm", "-lz"],
        pre_execution_callback=setup_ghostscript_library_files,
    )

As part of #45 we may want to make a public API similar to this and
move it into the dataset definitions.

Multiple validation callbacks can be defined for a single
benchmark. Where a benchmark matches multiple validators, they are
executed in parallel.

Compiling binaries from cBench benchmarks requires that the bitcodes
be compiled against the system-specific standard library, so this
patch also splits the cBench dataset into macOS and Linux versions.

WORKSPACE

@@ -1,6 +1,6 @@
 workspace(name = "CompilerGym")
 
-load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")
+load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive", "http_file")
 
 # === Google test ===
 
@@ -201,11 +201,93 @@ filegroup(
 )
 
 http_archive(
+    name = "ctuning-ai",
+    build_file_content = """
+filegroup(
+    name = "all",
+    srcs = glob(["**"]),
+    visibility = ["//visibility:public"],
+)
+
+filegroup(
+    name = "readme",
+    srcs = ["README.md"],
+    visibility = ["//visibility:public"],
+)
+""",
+    sha256 = "2e12233a3d898a96a2ce279123ee256bee357f649e3c579f974628b6c35ea53d",
+    strip_prefix = "ai-afaba3e2900523a0813693fd614a114b4a6f6729",
+    urls = ["https://github.com/ctuning/ai/archive/afaba3e2900523a0813693fd614a114b4a6f6729.tar.gz"],
+)
+
+# Datasets.
+
+http_file(
+    name = "cBench_consumer_tiff_data",
+    sha256 = "779abb7b7fee8733313e462e6066c16375e9209a9f7ff692fd06c7598946939a",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_consumer_tiff_data.tar.gz"],
+)
+
+http_file(
+    name = "cBench_office_data",
+    sha256 = "cfa09cd37cb93aba57415033905dc6308653c7b833feba5a25067bfb62999f32",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_office_data.tar.gz"],
+)
+
+http_file(
+    name = "cBench_telecom_data",
+    sha256 = "e5cb6663beefe32fd12f90c8f533f8e1bce2f05ee4e3836efb5556d5e1089df0",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_telecom_data.tar.gz"],
+)
+
+http_file(
+    name = "cBench_consumer_jpeg_data",
+    sha256 = "bec5ffc15cd2f952d9a786f3cd31d90955c318a5e4f69c5ba472f79d5a3e8f0b",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_consumer_jpeg_data.tar.gz"],
+)
+
+http_file(
+    name = "cBench_telecom_gsm_data",
+    sha256 = "52545d3a0ce15021131c62d96d3a3d7e6670e2d6c34226ac9a3d5191a1ee214a",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_telecom_gsm_data.tar.gz"],
+)
+
+http_file(
+    name = "cBench_consumer_data",
+    sha256 = "a4d40344af3022bfd7b4c6fcf6d59d598825b07d9e37769dbf1b3effa39aa445",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_consumer_data.tar.gz"],
+)
+
+http_file(
+    name = "cBench_bzip2_data",
+    sha256 = "46e5760eeef77e6b0c273af92de971bc45f33a59e0efc183073d9aa6b716c302",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_bzip2_data.tar.gz"],
+)
+
+http_file(
+    name = "cBench_network_patricia_data",
+    sha256 = "72dae0e670d93ef929e50aca7a138463e0915502281ccafe793e378cb2a85dfb",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_network_patricia_data.tar.gz"],
+)
+
+http_file(
     name = "cBench_network_dijkstra_data",
-    build_file_content = all_content,
+    sha256 = "41c13f59cdfbc772081cd941f499b030370bc570fc2ba60a5c4b7194bc36ca5f",
     urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_network_dijkstra_data.tar.gz"],
 )
 
+http_file(
+    name = "cBench_automotive_susan_data",
+    sha256 = "df56e1e44ccc560072381cdb001d770003ac74f92593dd5dbdfdd4ff9332a8e6",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_automotive_susan_data.tar.gz"],
+)
+
+http_file(
+    name = "cBench_automotive_qsort_data",
+    sha256 = "510b4225021408ac190f6f793e7d7171d3553c9916cfa8b2fb4ace005105e768",
+    urls = ["https://downloads.sourceforge.net/project/cbenchmark/cDatasets/V1.1/cDatasets_V1.1_automotive_qsort_data.tar.gz"],
+)
+
 # === C++ cpuinfo ===
 
 http_archive(

compiler_gym/envs/llvm/BUILD

@@ -33,8 +33,12 @@ py_library(
 py_library(
     name = "datasets",
     srcs = ["datasets.py"],
+    data = [
+        "//compiler_gym/third_party/llvm:clang",
+    ],
     deps = [
         "//compiler_gym/datasets:dataset",
+        "//compiler_gym/util",
     ],
 )