[Beignet] [PATCH] Reduce the compilation time of inline pass in runModulePass().

[yan.wang at linux.intel.com]

From: Yan Wang <yan.wang at linux.intel.com>

It could reduce much compilation time when run Luxmark scenes.
Avoid calling inline pass many times in runModulePass when module is
changed by the other pass.
Create a single funtion to run inline pass.
In this single funtion, lower pass and strict math related pass are
also added for keep enough precision.
---
 backend/src/llvm/llvm_to_gen.cpp | 53 +++++++++++++++++++++++++++++++++++++---
 1 file changed, 50 insertions(+), 3 deletions(-)

diff --git a/backend/src/llvm/llvm_to_gen.cpp b/backend/src/llvm/llvm_to_gen.cpp
index e108810..0d51ee3 100644
--- a/backend/src/llvm/llvm_to_gen.cpp
+++ b/backend/src/llvm/llvm_to_gen.cpp
@@ -106,6 +106,55 @@ namespace gbe
     FPM.doFinalization();
   }
 
+  void runInlinePass(Module &mod, TARGETLIBRARY *libraryInfo, const DataLayout &DL, int optLevel, bool strictMath)
+  {
+#if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 7
+    legacy::PassManager MPM;
+#else
+    PassManager MPM;
+#endif
+
+#if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 7
+#elif LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 6
+    MPM.add(new DataLayoutPass());
+#elif LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR == 5
+    MPM.add(new DataLayoutPass(DL));
+#else
+    MPM.add(new DataLayout(DL));
+#endif
+
+#if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 7
+    MPM.add(new TargetLibraryInfoWrapperPass(*libraryInfo));
+#else
+    MPM.add(new TargetLibraryInfo(*libraryInfo));
+#endif
+    MPM.add(createIntrinsicLoweringPass());
+    MPM.add(createBarrierNodupPass(false));   // remove noduplicate fnAttr before inlining.
+    MPM.add(createFunctionInliningPass(20000));
+    //MPM.add(createAlwaysInlinerPass());
+    MPM.add(createBarrierNodupPass(true));    // restore noduplicate fnAttr after inlining.
+#if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 5
+    // FIXME Workaround: we find that CustomLoopUnroll may increase register pressure greatly,
+    // and it may even make som cl kernel cannot compile because of limited scratch memory for spill.
+    // As we observe this under strict math. So we disable CustomLoopUnroll if strict math is enabled.
+    if (!strictMath) {
+#if !defined(__ANDROID__)
+      MPM.add(createCustomLoopUnrollPass()); //1024, 32, 1024, 512)); //Unroll loops
+#endif
+      MPM.add(createLoopUnrollPass()); //1024, 32, 1024, 512)); //Unroll loops
+      if(optLevel > 0) {
+#if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 8
+        MPM.add(createSROAPass());
+#else
+        MPM.add(createSROAPass(/*RequiresDomTree*/ false));
+#endif
+        MPM.add(createGVNPass());                 // Remove redundancies
+      }
+    }
+#endif
+    MPM.run(mod);
+  }
+
   void runModulePass(Module &mod, TARGETLIBRARY *libraryInfo, const DataLayout &DL, int optLevel, bool strictMath)
   {
 #if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 7
@@ -146,9 +195,6 @@ namespace gbe
     MPM.add(createInstructionCombiningPass());// Clean up after IPCP & DAE
     MPM.add(createCFGSimplificationPass());   // Clean up after IPCP & DAE
     MPM.add(createPruneEHPass());             // Remove dead EH info
-    MPM.add(createBarrierNodupPass(false));   // remove noduplicate fnAttr before inlining.
-    MPM.add(createFunctionInliningPass(20000));
-    MPM.add(createBarrierNodupPass(true));    // restore noduplicate fnAttr after inlining.
 #if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 9
     MPM.add(createPostOrderFunctionAttrsLegacyPass());
 #elif LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 8
@@ -344,6 +390,7 @@ namespace gbe
     OUTPUT_BITCODE(AFTER_LINK, mod);
 
     runFuntionPass(mod, libraryInfo, DL);
+    runInlinePass(mod, libraryInfo, DL, optLevel, strictMath);
     runModulePass(mod, libraryInfo, DL, optLevel, strictMath);
 #if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 7
     legacy::PassManager passes;
-- 
2.7.4