Merge pull request #31101 from JuliaLang/tb/skip_lower_intrinsics

Make it possible to add optimization passes without lowering intrinsics.
JuliaLang · Feb 25, 2019 · d2cd3f4 · d2cd3f4
2 parents 31234a5 + e165c90
commit d2cd3f4
Show file tree

Hide file tree

Showing 2 changed files with 30 additions and 24 deletions.
diff --git a/src/jitlayers.cpp b/src/jitlayers.cpp
@@ -92,7 +92,8 @@ void addTargetPasses(legacy::PassManagerBase *PM, TargetMachine *TM)
 
 // this defines the set of optimization passes defined for Julia at various optimization levels.
 // it assumes that the TLI and TTI wrapper passes have already been added.
-void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level, bool dump_native)
+void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
+ bool lower_intrinsics, bool dump_native)
 {
 #ifdef JL_DEBUG_BUILD
  PM->add(createGCInvariantVerifierPass(true));
@@ -114,12 +115,14 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level, bool dump
  }
  PM->add(createMemCpyOptPass()); // Remove memcpy / form memset
  PM->add(createAlwaysInlinerLegacyPass()); // Respect always_inline
- PM->add(createBarrierNoopPass());
- PM->add(createLowerExcHandlersPass());
- PM->add(createGCInvariantVerifierPass(false));
- PM->add(createLateLowerGCFramePass());
- PM->add(createLowerPTLSPass(dump_native));
- PM->add(createLowerSimdLoopPass()); // Annotate loop marked with "simdloop" as LLVM parallel loop
+ if (lower_intrinsics) {
+ PM->add(createBarrierNoopPass());
+ PM->add(createLowerExcHandlersPass());
+ PM->add(createGCInvariantVerifierPass(false));
+ PM->add(createLateLowerGCFramePass());
+ PM->add(createLowerPTLSPass(dump_native));
+ PM->add(createLowerSimdLoopPass()); // Annotate loop marked with "simdloop" as LLVM parallel loop
+ }
  if (dump_native)
  PM->add(createMultiVersioningPass());
  return;
@@ -219,24 +222,27 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level, bool dump
  PM->add(createInstructionCombiningPass()); // Clean up after SLP loop vectorizer
  PM->add(createLoopVectorizePass()); // Vectorize loops
  PM->add(createInstructionCombiningPass()); // Clean up after loop vectorizer
- // LowerPTLS removes an indirect call. As a result, it is likely to trigger
- // LLVM's devirtualization heuristics, which would result in the entire
- // pass pipeline being re-exectuted. Prevent this by inserting a barrier.
- PM->add(createBarrierNoopPass());
- PM->add(createLowerExcHandlersPass());
- PM->add(createGCInvariantVerifierPass(false));
- PM->add(createLateLowerGCFramePass());
- // Remove dead use of ptls
- PM->add(createDeadCodeEliminationPass());
- PM->add(createLowerPTLSPass(dump_native));
- // Clean up write barrier and ptls lowering
- PM->add(createCFGSimplificationPass());
- PM->add(createCombineMulAddPass());
+
+ if (lower_intrinsics) {
+ // LowerPTLS removes an indirect call. As a result, it is likely to trigger
+ // LLVM's devirtualization heuristics, which would result in the entire
+ // pass pipeline being re-exectuted. Prevent this by inserting a barrier.
+ PM->add(createBarrierNoopPass());
+ PM->add(createLowerExcHandlersPass());
+ PM->add(createGCInvariantVerifierPass(false));
+ PM->add(createLateLowerGCFramePass());
+ // Remove dead use of ptls
+ PM->add(createDeadCodeEliminationPass());
+ PM->add(createLowerPTLSPass(dump_native));
+ // Clean up write barrier and ptls lowering
+ PM->add(createCFGSimplificationPass());
+ PM->add(createCombineMulAddPass());
+ }
 }
 
 extern "C" JL_DLLEXPORT
-void jl_add_optimization_passes(LLVMPassManagerRef PM, int opt_level) {
- addOptimizationPasses(unwrap(PM), opt_level);
+void jl_add_optimization_passes(LLVMPassManagerRef PM, int opt_level, int lower_intrinsics) {
+ addOptimizationPasses(unwrap(PM), opt_level, lower_intrinsics);
 }
 
 #if defined(_OS_LINUX_) || defined(_OS_WINDOWS_) || defined(_OS_FREEBSD_)
@@ -946,7 +952,7 @@ void jl_dump_native(const char *bc_fname, const char *unopt_bc_fname, const char
  if (unopt_bc_fname)
  PM.add(createBitcodeWriterPass(unopt_bc_OS));
  if (bc_fname || obj_fname)
- addOptimizationPasses(&PM, jl_options.opt_level, true);
+ addOptimizationPasses(&PM, jl_options.opt_level, true, true);
  if (bc_fname)
  PM.add(createBitcodeWriterPass(bc_OS));
  if (obj_fname)

diff --git a/src/jitlayers.h b/src/jitlayers.h
@@ -34,7 +34,7 @@ extern Function *juliapersonality_func;
 typedef struct {Value *gv; int32_t index;} jl_value_llvm; // uses 1-based indexing
 
 void addTargetPasses(legacy::PassManagerBase *PM, TargetMachine *TM);
-void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level, bool dump_native=false);
+void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level, bool lower_intrinsics=true, bool dump_native=false);
 void* jl_emit_and_add_to_shadow(GlobalVariable *gv, void *gvarinit = NULL);
 void* jl_get_globalvar(GlobalVariable *gv);
 GlobalVariable *jl_get_global_for(const char *cname, void *addr, Module *M);