LLVM freeze instruction between mask and div 2/?

Author: alexbaden

third_party/intel/lib/CMakeLists.txt

@@ -1,6 +1,7 @@
 add_subdirectory(Analysis)
 add_subdirectory(Dialect)
 add_subdirectory(GPUToTritonGEN)
+add_subdirectory(LLVMIR)
 add_subdirectory(Target)
 add_subdirectory(TritonAnnotateModule)
 add_subdirectory(TritonGENToLLVM)

third_party/intel/triton_xpu.cc

@@ -17,6 +17,7 @@
 #include "intel/include/TritonAnnotateModule/Passes.h"
 #include "intel/include/TritonIntelGPUToLLVM/Passes.h"
 #include "intel/include/TritonToTritonGPUWarp/Passes.h"
+#include "intel/lib/LLVMIR/LLVMPasses.h"
 
 #include "triton/Target/SPIRV/SPIRVTranslation.h"
 #include "triton/Tools/Sys/GetEnv.hpp"
@@ -207,6 +208,7 @@ void init_triton_intel(py::module &&m) {
           fpm.addPass(BreakStructPhiNodesPass());
           fpm.addPass(InstCombinePass());
         });
+#if 1
     pb.registerPeepholeEPCallback(
         [&](llvm::FunctionPassManager &fpm, llvm::OptimizationLevel level) {
           // The Triton masked load pattern can generate instances where the
@@ -216,8 +218,12 @@ void init_triton_intel(py::module &&m) {
           // an incorrect result for the kernel. Adding `DivRemPairsPass`
           // introduces freeze instructions which prevent UB from leaking into
           // div/rem instructions.
-          fpm.addPass(DivRemPairsPass());
+          // fpm.addPass(DivRemPairsPass());
+          fpm.addPass(FreezeMaskedDivRemPass());
         });
+#else
+    mpm.addPass(createModuleToFunctionPassAdaptor(FreezeMaskedDivRemPass()));
+#endif
     mpm.addPass(pb.buildPerModuleDefaultPipeline(opt));
     mpm.run(*mod, mam);
   });