From 85cccfb81f4c01863817a3e8da54cd4fb6904e1d Mon Sep 17 00:00:00 2001
From: Keren Zhou <kerenzhou@openai.com>
Date: Sun, 20 Nov 2022 21:40:36 -0800
Subject: [PATCH] [BUILD] Fix compilation problems in the release build (#897)

---
 .../TritonGPUToLLVM/TritonGPUToLLVM.cpp       | 15 +++++++++-----
 python/setup.py                               | 20 +++++++++----------
 python/src/triton.cc                          |  2 +-
 3 files changed, 21 insertions(+), 16 deletions(-)

diff --git a/lib/Conversion/TritonGPUToLLVM/TritonGPUToLLVM.cpp b/lib/Conversion/TritonGPUToLLVM/TritonGPUToLLVM.cpp
index 0d0b62280..8338388be 100644
--- a/lib/Conversion/TritonGPUToLLVM/TritonGPUToLLVM.cpp
+++ b/lib/Conversion/TritonGPUToLLVM/TritonGPUToLLVM.cpp
@@ -1438,9 +1438,10 @@ struct BroadcastOpConversion
     SmallVector<int64_t> resultLogicalShape(2 * rank);
     SmallVector<unsigned> broadcastDims;
     for (unsigned d = 0; d < rank; ++d) {
-      unsigned resultShapePerCTA = triton::gpu::getSizePerThread(resultLayout)[d] *
-                                   triton::gpu::getThreadsPerWarp(resultLayout)[d] *
-                                   triton::gpu::getWarpsPerCTA(resultLayout)[d];
+      unsigned resultShapePerCTA =
+          triton::gpu::getSizePerThread(resultLayout)[d] *
+          triton::gpu::getThreadsPerWarp(resultLayout)[d] *
+          triton::gpu::getWarpsPerCTA(resultLayout)[d];
       int64_t numCtas = ceil<unsigned>(resultShape[d], resultShapePerCTA);
       if (srcShape[d] != resultShape[d]) {
         assert(srcShape[d] == 1);
@@ -1450,10 +1451,12 @@ struct BroadcastOpConversion
             std::max<unsigned>(1, triton::gpu::getSizePerThread(srcLayout)[d]);
       } else {
         srcLogicalShape[d] = numCtas;
-        srcLogicalShape[d + rank] = triton::gpu::getSizePerThread(resultLayout)[d];
+        srcLogicalShape[d + rank] =
+            triton::gpu::getSizePerThread(resultLayout)[d];
       }
       resultLogicalShape[d] = numCtas;
-      resultLogicalShape[d + rank] = triton::gpu::getSizePerThread(resultLayout)[d];
+      resultLogicalShape[d + rank] =
+          triton::gpu::getSizePerThread(resultLayout)[d];
 
       srcLogicalOrder[d] = order[d] + rank;
       srcLogicalOrder[d + rank] = order[d];
@@ -1968,6 +1971,7 @@ struct PrintfOpConversion
       return "%u";
     }
     assert(false && "not supported type");
+    return "";
   }
 
   // declare vprintf(i8*, i8*) as external function
@@ -5482,6 +5486,7 @@ Value convertSplatLikeOpWithMmaLayout(const MmaEncodingAttr &layout,
   }
 
   assert(false && "Unsupported mma layout found");
+  return {};
 }
 
 class TritonGPUToLLVMTypeConverter : public LLVMTypeConverter {
diff --git a/python/setup.py b/python/setup.py
index a39aa1e5b..73b03d55f 100644
--- a/python/setup.py
+++ b/python/setup.py
@@ -6,7 +6,6 @@ import shutil
 import subprocess
 import sys
 import tarfile
-import tempfile
 import urllib.request
 from distutils.version import LooseVersion
 from typing import NamedTuple
@@ -26,7 +25,9 @@ def get_build_type():
     elif check_env_flag("REL_WITH_DEB_INFO"):
         return "RelWithDebInfo"
     else:
-        return "Release"
+        return "Debug"
+        # TODO(Keren): Restore this before we merge into master
+        #return "Release"
 
 
 # --- third party packages -----
@@ -124,19 +125,14 @@ class CMakeBuild(build_ext):
             self.build_extension(ext)
 
     def build_extension(self, ext):
-        self.debug = True
         lit_dir = shutil.which('lit')
         triton_cache_path = os.path.join(os.environ["HOME"], ".triton")
         # lit is used by the test suite
         thirdparty_cmake_args = get_thirdparty_packages(triton_cache_path)
         extdir = os.path.abspath(os.path.dirname(self.get_ext_fullpath(ext.path)))
         # create build directories
-        build_suffix = 'debug' if self.debug else 'release'
-        llvm_build_dir = os.path.join(tempfile.gettempdir(), "llvm-" + build_suffix)
         if not os.path.exists(self.build_temp):
             os.makedirs(self.build_temp)
-        if not os.path.exists(llvm_build_dir):
-            os.makedirs(llvm_build_dir)
         # python directories
         python_include_dir = distutils.sysconfig.get_python_inc()
         cmake_args = [
@@ -145,13 +141,13 @@ class CMakeBuild(build_ext):
             "-DTRITON_BUILD_TUTORIALS=OFF",
             "-DTRITON_BUILD_PYTHON_MODULE=ON",
             # '-DPYTHON_EXECUTABLE=' + sys.executable,
-            # '-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON',
+            '-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON',
             "-DPYTHON_INCLUDE_DIRS=" + python_include_dir,
             "-DLLVM_EXTERNAL_LIT=" + lit_dir
         ] + thirdparty_cmake_args
 
         # configuration
-        cfg = "Debug" if self.debug else "Release"
+        cfg = get_build_type()
         build_args = ["--config", cfg]
 
         if platform.system() == "Windows":
@@ -183,7 +179,11 @@ setup(
         "torch",
         "lit",
     ],
-    package_data={"triton/ops": ["*.c"], "triton/ops/blocksparse": ["*.c"]},
+    package_data={
+        "triton/ops": ["*.c"],
+        "triton/ops/blocksparse": ["*.c"],
+        "triton/language": ["*.bc"]
+    },
     include_package_data=True,
     ext_modules=[CMakeExtension("triton", "triton/_C/")],
     cmdclass={"build_ext": CMakeBuild},
diff --git a/python/src/triton.cc b/python/src/triton.cc
index f07d4cfee..f7c29dd4a 100644
--- a/python/src/triton.cc
+++ b/python/src/triton.cc
@@ -551,7 +551,7 @@ void init_triton_ir(py::module &&m) {
                return llvm::dyn_cast<mlir::FuncOp>(funcOperation);
              auto loc = self.getUnknownLoc();
              if (auto funcTy = funcType.dyn_cast<mlir::FunctionType>()) {
-               mlir::ArrayRef<mlir::NamedAttribute> attrs = {
+               llvm::SmallVector<mlir::NamedAttribute> attrs = {
                    mlir::NamedAttribute(self.getStringAttr("sym_visibility"),
                                         self.getStringAttr(visibility))};
                return self.create<mlir::FuncOp>(loc, funcName, funcTy, attrs);