[PATCH] D35703: [GPGPU] Add support for NVIDIA libdevice

[Siddharth via Phabricator via llvm-commits]

bollu added inline comments.


================
Comment at: lib/CodeGen/PPCGCodeGeneration.cpp:1314
+/// A list of functions that are available in NVIDIA's libdevice.
+std::vector<std::string> LibDeviceFunctions = {"exp", "expf", "expl", "cos",
+                                               "cosf"};
----------------
Consider changing to `SmallSet`? That feels correct in terms of semantics (a `set` of functions.)


================
Comment at: lib/CodeGen/PPCGCodeGeneration.cpp:1418
+
+  bool AllowLibDevice = Arch == GPUArch::NVPTX64;
+
----------------
`const bool`? :)


================
Comment at: lib/CodeGen/PPCGCodeGeneration.cpp:2131
+    }
+  }
+
----------------
consider moving computing `RequiresLibDevice` to a pure function? 


================
Comment at: lib/CodeGen/PPCGCodeGeneration.cpp:2138
+
+    if (!LibDeviceModule) {
+      BuildSuccessful = false;
----------------
I believe we can `assert` at this point, since this is not a "mis-compile" in the strictest sense of the word?. 


================
Comment at: lib/CodeGen/PPCGCodeGeneration.cpp:2148
+    // Set an nvptx64 target triple to avoid linker warnings. The original
+    // trible of the libdevice files are nvptx-unknown-unknown.
+    LibDeviceModule->setTargetTriple(Triple::normalize("nvptx64-nvidia-cuda"));
----------------
nit: `trible` -> `triple`.


https://reviews.llvm.org/D35703