ac: run LLVM optimization passes only on the final function after inlining

src/amd/common/ac_llvm_helper.cpp

@@ -36,6 +36,7 @@
 #include <llvm/Target/TargetMachine.h>
 #include <llvm/IR/IRBuilder.h>
 #include <llvm/Analysis/TargetLibraryInfo.h>
+#include <llvm/Transforms/IPO.h>
 
 #include <llvm/IR/LegacyPassManager.h>
 #if HAVE_LLVM < 0x0700
@@ -165,3 +166,8 @@ bool ac_compile_module_to_binary(struct ac_compiler_passes *p, LLVMModuleRef mod
 		fprintf(stderr, "amd: cannot read an ELF shader binary\n");
 	return success;
 }
+
+void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr)
+{
+	llvm::unwrap(passmgr)->add(llvm::createBarrierNoopPass());
+}

src/amd/common/ac_llvm_util.c

@@ -181,6 +181,13 @@ static LLVMPassManagerRef ac_create_passmgr(LLVMTargetLibraryInfoRef target_libr
 	if (check_ir)
 		LLVMAddVerifierPass(passmgr);
 	LLVMAddAlwaysInlinerPass(passmgr);
+	/* Normally, the pass manager runs all passes on one function before
+	 * moving onto another. Adding a barrier no-op pass forces the pass
+	 * manager to run the inliner on all functions first, which makes sure
+	 * that the following passes are only run on the remaining non-inline
+	 * function, so it removes useless work done on dead inline functions.
+	 */
+	ac_llvm_add_barrier_noop_pass(passmgr);
 	/* This pass should eliminate all the load and store instructions. */
 	LLVMAddPromoteMemoryToRegisterPass(passmgr);
 	LLVMAddScalarReplAggregatesPass(passmgr);

src/amd/common/ac_llvm_util.h

@@ -133,6 +133,7 @@ struct ac_compiler_passes *ac_create_llvm_passes(LLVMTargetMachineRef tm);
 void ac_destroy_llvm_passes(struct ac_compiler_passes *p);
 bool ac_compile_module_to_binary(struct ac_compiler_passes *p, LLVMModuleRef module,
 				 struct ac_shader_binary *binary);
+void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr);
 
 #ifdef __cplusplus
 }