Add compiler flags -Os and -Oz to optimize binary size (#14463)

Co-authored-by: Johannes Müller <[email protected]>
Co-authored-by: Sijawusz Pur Rahnama <[email protected]>

man/crystal.1

@@ -496,6 +496,10 @@ Low optimization
 Middle optimization
 .It Fl O3
 High optimization
+.It Fl \!Os
+Middle optimization with focus on file size
+.It Fl Oz
+Middle optimization aggressively focused on file size
 .
 .Sh ENVIRONMENT\ VARIABLES
 .Bl -tag -width "12345678" -compact

src/compiler/crystal/codegen/target.cr

@@ -204,10 +204,10 @@ class Crystal::Codegen::Target
     end
 
     opt_level = case optimization_mode
-                in .o3? then LLVM::CodeGenOptLevel::Aggressive
-                in .o2? then LLVM::CodeGenOptLevel::Default
-                in .o1? then LLVM::CodeGenOptLevel::Less
-                in .o0? then LLVM::CodeGenOptLevel::None
+                in .o3?             then LLVM::CodeGenOptLevel::Aggressive
+                in .o2?, .os?, .oz? then LLVM::CodeGenOptLevel::Default
+                in .o1?             then LLVM::CodeGenOptLevel::Less
+                in .o0?             then LLVM::CodeGenOptLevel::None
                 end
 
     target = LLVM::Target.from_triple(self.to_s)

src/compiler/crystal/command.cr

@@ -521,9 +521,12 @@ class Crystal::Command
         opts.on("--release", "Compile in release mode (-O3 --single-module)") do
           compiler.release!
         end
-        opts.on("-O LEVEL", "Optimization mode: 0 (default), 1, 2, 3") do |level|
-          optimization_mode = level.to_i?.try { |v| Compiler::OptimizationMode.from_value?(v) }
-          compiler.optimization_mode = optimization_mode || raise Error.new("Invalid optimization mode: #{level}")
+        opts.on("-O LEVEL", "Optimization mode: 0 (default), 1, 2, 3, s, z") do |level|
+          if mode = Compiler::OptimizationMode.from_level?(level)
+            compiler.optimization_mode = mode
+          else
+            raise Error.new("Invalid optimization mode: O#{level}")
+          end
         end
       end
 
@@ -661,8 +664,12 @@ class Crystal::Command
     opts.on("--release", "Compile in release mode (-O3 --single-module)") do
       compiler.release!
     end
-    opts.on("-O LEVEL", "Optimization mode: 0 (default), 1, 2, 3") do |level|
-      compiler.optimization_mode = Compiler::OptimizationMode.from_value?(level.to_i) || raise Error.new("Unknown optimization mode #{level}")
+    opts.on("-O LEVEL", "Optimization mode: 0 (default), 1, 2, 3, s, z") do |level|
+      if mode = Compiler::OptimizationMode.from_level?(level)
+        compiler.optimization_mode = mode
+      else
+        raise Error.new("Invalid optimization mode: O#{level}")
+      end
     end
     opts.on("--single-module", "Generate a single LLVM module") do
       compiler.single_module = true

src/compiler/crystal/compiler.cr

@@ -102,9 +102,27 @@ module Crystal
       # enables with --release flag
       O3 = 3
 
+      # optimize for size, enables most O2 optimizations but aims for smaller
+      # code size
+      Os
+
+      # optimize aggressively for size rather than speed
+      Oz
+
       def suffix
         ".#{to_s.downcase}"
       end
+
+      def self.from_level?(level : String) : self?
+        case level
+        when "0" then O0
+        when "1" then O1
+        when "2" then O2
+        when "3" then O3
+        when "s" then Os
+        when "z" then Oz
+        end
+      end
     end
 
     # Sets the Optimization mode.
@@ -674,8 +692,8 @@ module Crystal
       exit 1
     end
 
-    {% if LibLLVM::IS_LT_130 %}
-      protected def optimize(llvm_mod)
+    {% if LibLLVM::IS_LT_170 %}
+      private def optimize_with_pass_manager(llvm_mod)
         fun_pass_manager = llvm_mod.new_function_pass_manager
         pass_manager_builder.populate fun_pass_manager
         fun_pass_manager.run llvm_mod
@@ -700,6 +718,8 @@ module Crystal
           registry.initialize_all
 
           builder = LLVM::PassManagerBuilder.new
+          builder.size_level = 0
+
           case optimization_mode
           in .o3?
             builder.opt_level = 3
@@ -712,26 +732,37 @@ module Crystal
             builder.use_inliner_with_threshold = 150
           in .o0?
             # default behaviour, no optimizations
+          in .os?
+            builder.opt_level = 2
+            builder.size_level = 1
+            builder.use_inliner_with_threshold = 50
+          in .oz?
+            builder.opt_level = 2
+            builder.size_level = 2
+            builder.use_inliner_with_threshold = 5
           end
 
-          builder.size_level = 0
-
           builder
         end
       end
-    {% else %}
-      protected def optimize(llvm_mod)
+    {% end %}
+
+    protected def optimize(llvm_mod)
+      {% if LibLLVM::IS_LT_130 %}
+        optimize_with_pass_manager(llvm_mod)
+      {% else %}
+        {% if LibLLVM::IS_LT_170 %}
+          # PassBuilder doesn't support Os and Oz before LLVM 17
+          if @optimization_mode.os? || @optimization_mode.oz?
+            return optimize_with_pass_manager(llvm_mod)
+          end
+        {% end %}
+
         LLVM::PassBuilderOptions.new do |options|
-          mode = case @optimization_mode
-                 in .o3? then "default<O3>"
-                 in .o2? then "default<O2>"
-                 in .o1? then "default<O1>"
-                 in .o0? then "default<O0>"
-                 end
-          LLVM.run_passes(llvm_mod, mode, target_machine, options)
+          LLVM.run_passes(llvm_mod, "default<#{@optimization_mode}>", target_machine, options)
         end
-      end
-    {% end %}
+      {% end %}
+    end
 
     private def run_linker(linker_name, command, args)
       print_command(command, args) if verbose?

src/llvm/function_pass_manager.cr

@@ -1,4 +1,4 @@
-{% unless LibLLVM::IS_LT_130 %}
+{% unless LibLLVM::IS_LT_170 %}
   @[Deprecated("The legacy pass manager was removed in LLVM 17. Use `LLVM::PassBuilderOptions` instead")]
 {% end %}
 class LLVM::FunctionPassManager
@@ -34,7 +34,7 @@ class LLVM::FunctionPassManager
     LibLLVM.dispose_pass_manager(@unwrap)
   end
 
-  {% unless LibLLVM::IS_LT_130 %}
+  {% unless LibLLVM::IS_LT_170 %}
     @[Deprecated("The legacy pass manager was removed in LLVM 17. Use `LLVM::PassBuilderOptions` instead")]
   {% end %}
   struct Runner

src/llvm/lib_llvm/initialization.cr

@@ -6,11 +6,11 @@ lib LibLLVM
   fun initialize_core = LLVMInitializeCore(r : PassRegistryRef)
   fun initialize_transform_utils = LLVMInitializeTransformUtils(r : PassRegistryRef)
   fun initialize_scalar_opts = LLVMInitializeScalarOpts(r : PassRegistryRef)
-  fun initialize_obj_c_arc_opts = LLVMInitializeObjCARCOpts(r : PassRegistryRef)
+  {% if LibLLVM::IS_LT_160 %} fun initialize_obj_c_arc_opts = LLVMInitializeObjCARCOpts(r : PassRegistryRef) {% end %}
   fun initialize_vectorization = LLVMInitializeVectorization(r : PassRegistryRef)
   fun initialize_inst_combine = LLVMInitializeInstCombine(r : PassRegistryRef)
   fun initialize_ipo = LLVMInitializeIPO(r : PassRegistryRef)
-  fun initialize_instrumentation = LLVMInitializeInstrumentation(r : PassRegistryRef)
+  {% if LibLLVM::IS_LT_160 %} fun initialize_instrumentation = LLVMInitializeInstrumentation(r : PassRegistryRef) {% end %}
   fun initialize_analysis = LLVMInitializeAnalysis(r : PassRegistryRef)
   fun initialize_ipa = LLVMInitializeIPA(r : PassRegistryRef)
   fun initialize_code_gen = LLVMInitializeCodeGen(r : PassRegistryRef)

src/llvm/lib_llvm/transforms/pass_builder.cr

@@ -10,4 +10,10 @@ lib LibLLVM
 
   fun create_pass_builder_options = LLVMCreatePassBuilderOptions : PassBuilderOptionsRef
   fun dispose_pass_builder_options = LLVMDisposePassBuilderOptions(options : PassBuilderOptionsRef)
+  {% unless LibLLVM::IS_LT_170 %}
+    fun pass_builder_options_set_inliner_threshold = LLVMPassBuilderOptionsSetInlinerThreshold(PassBuilderOptionsRef, Int)
+  {% end %}
+  fun pass_builder_options_set_loop_unrolling = LLVMPassBuilderOptionsSetLoopUnrolling(PassBuilderOptionsRef, Bool)
+  fun pass_builder_options_set_loop_vectorization = LLVMPassBuilderOptionsSetLoopVectorization(PassBuilderOptionsRef, Bool)
+  fun pass_builder_options_set_slp_vectorization = LLVMPassBuilderOptionsSetSLPVectorization(PassBuilderOptionsRef, Bool)
 end

src/llvm/module.cr

@@ -84,7 +84,7 @@ class LLVM::Module
     self
   end
 
-  {% unless LibLLVM::IS_LT_130 %}
+  {% unless LibLLVM::IS_LT_170 %}
     @[Deprecated("The legacy pass manager was removed in LLVM 17. Use `LLVM::PassBuilderOptions` instead")]
   {% end %}
   def new_function_pass_manager

src/llvm/module_pass_manager.cr

@@ -1,4 +1,4 @@
-{% unless LibLLVM::IS_LT_130 %}
+{% unless LibLLVM::IS_LT_170 %}
   @[Deprecated("The legacy pass manager was removed in LLVM 17. Use `LLVM::PassBuilderOptions` instead")]
 {% end %}
 class LLVM::ModulePassManager

src/llvm/pass_builder_options.cr

@@ -25,4 +25,22 @@ class LLVM::PassBuilderOptions
 
     LibLLVM.dispose_pass_builder_options(self)
   end
+
+  {% unless LibLLVM::IS_LT_170 %}
+    def set_inliner_threshold(threshold : Int)
+      LibLLVM.pass_builder_options_set_inliner_threshold(self, threshold)
+    end
+  {% end %}
+
+  def set_loop_unrolling(enabled : Bool)
+    LibLLVM.pass_builder_options_set_loop_unrolling(self, enabled)
+  end
+
+  def set_loop_vectorization(enabled : Bool)
+    LibLLVM.pass_builder_options_set_loop_vectorization(self, enabled)
+  end
+
+  def set_slp_vectorization(enabled : Bool)
+    LibLLVM.pass_builder_options_set_slp_vectorization(self, enabled)
+  end
 end

src/llvm/pass_manager_builder.cr

@@ -1,4 +1,4 @@
-{% unless LibLLVM::IS_LT_130 %}
+{% unless LibLLVM::IS_LT_170 %}
   @[Deprecated("The legacy pass manager was removed in LLVM 17. Use `LLVM::PassBuilderOptions` instead")]
 {% end %}
 class LLVM::PassManagerBuilder

src/llvm/pass_registry.cr

@@ -1,4 +1,4 @@
-{% unless LibLLVM::IS_LT_130 %}
+{% unless LibLLVM::IS_LT_170 %}
   @[Deprecated("The legacy pass manager was removed in LLVM 17. Use `LLVM::PassBuilderOptions` instead")]
 {% end %}
 struct LLVM::PassRegistry
@@ -9,17 +9,32 @@ struct LLVM::PassRegistry
   def initialize(@unwrap : LibLLVM::PassRegistryRef)
   end
 
-  Inits = %w(core transform_utils scalar_opts obj_c_arc_opts vectorization inst_combine ipo instrumentation analysis ipa code_gen target)
+  {% begin %}
+    Inits = %w[
+      initialize_core
+      initialize_transform_utils
+      initialize_scalar_opts
+      {% if LibLLVM::IS_LT_160 %} initialize_obj_c_arc_opts {% end %}
+      initialize_vectorization
+      initialize_inst_combine
+      initialize_ipo
+      {% if LibLLVM::IS_LT_160 %} initialize_instrumentation {% end %}
+      initialize_analysis
+      initialize_ipa
+      initialize_code_gen
+      initialize_target
+    ]
+  {% end %}
 
   {% for name in Inits %}
-    def initialize_{{name.id}}
-      LibLLVM.initialize_{{name.id}} self
+    def {{name.id}}
+      LibLLVM.{{name.id}} self
     end
   {% end %}
 
   def initialize_all
     {% for name in Inits %}
-      initialize_{{name.id}}
+      {{name.id}}
     {% end %}
   end