Support Os or Oz inside the LTO

[jinlin]

Summary

The changes are to support Os and Oz inside the LTO.

[fhahn]

Thanks for the patch.

I think this probably should be split up in 2 different changes:

1. Changes to the pipeline (llvm/lib/Transforms/IPO/PassManagerBuilder.cpp). Ideally those come with some data motivating the change and why there won't be regressions.
2. Changes to LTOCodeGenerator & co. They also need a test.

[MaskRay]

I don' think mapping 's' to 4 and 'z' to 5 is suitable. It is quite different. If you just want clang -fuse-ld=lld -Oz -flto{,=thin} to work, it already does after D79919

[jinlin]

In D95541#2718205, @MaskRay wrote:

I don' think mapping 's' to 4 and 'z' to 5 is suitable. It is quite different. If you just want clang -fuse-ld=lld -Oz -flto{,=thin} to work, it already does after D79919

First of all, the changes in D719919 does not pass the SizeLevel information to any optimizations including inlining.

Secondly, the changes in D79919 is not general. It does not work in MacOS.

llvm/include/llvm/LTO/Config.h

   CodeGenOpt::Level CGOptLevel = CodeGenOpt::Default;
   CodeGenFileType CGFileType = CGFT_ObjectFile;
   unsigned OptLevel = 2;
+  unsigned SizeLevel = 0;
   bool DisableVerify = false;
 
   /// Use the new pass manager

llvm/lib/LTO/LTOBackend.cpp

   PMB.LibraryInfo = new TargetLibraryInfoImpl(Triple(TM->getTargetTriple()));
   if (Conf.Freestanding)
     PMB.LibraryInfo->disableAllFunctions();
-  PMB.Inliner = createFunctionInliningPass();
+  PMB.Inliner =
+      createFunctionInliningPass(Conf.OptLevel, Conf.SizeLevel, false);
   PMB.ExportSummary = ExportSummary;
   PMB.ImportSummary = ImportSummary;
   // Unconditionally verify input since it is not verified before this

llvm/lib/LTO/LTOCodeGenerator.cpp

 }
 
 void LTOCodeGenerator::setOptLevel(unsigned Level) {
+  if (Level > 3) {
+    Config.SizeLevel = Level - 3;
+    Level = 2;
+  }
   Config.OptLevel = Level;
   Config.PTO.LoopVectorization = Config.OptLevel > 1;
   Config.PTO.SLPVectorization = Config.OptLevel > 1;

llvm/test/tools/lto/lto_optsize.ll

+; RUN: llvm-as %s -o %t.o
+
+; RUN: %ld64 -lto_library %llvmshlibdir/libLTO.dylib -dylib -arch x86_64 -macosx_version_min 10.10.0 -o %t.dylib %t.o -save-temps  -undefined dynamic_lookup -exported_symbol inner -exported_symbol outer -exported_symbol outer2 -exported_symbol extern -exported_symbol a -mllvm -Os
+; RUN: llvm-dis %t.dylib.lto.opt.bc -o - | FileCheck --check-prefix=OS %s
+
+; RUN: %ld64 -lto_library %llvmshlibdir/libLTO.dylib -dylib -arch x86_64 -macosx_version_min 10.10.0 -o %t.dylib %t.o -save-temps  -undefined dynamic_lookup -exported_symbol inner -exported_symbol outer -exported_symbol outer2 -exported_symbol extern -exported_symbol a -mllvm -Oz
+; RUN: llvm-dis %t.dylib.lto.opt.bc -o - | FileCheck --check-prefix=OZ %s
+
+; The inline threshold for a function with the optsize attribute is currently
+; the same as the global inline threshold for -Os. Check that the optsize
+; function attribute doesn't alter the function-specific inline threshold if the
+; global inline threshold is lower (as for -Oz).
+
+target triple = "x86_64-apple-macosx10.8.0"
+
+@a = global i32 4
+
+; This function should be larger than the inline threshold for -Oz (25), but
+; smaller than the inline threshold for optsize (75).
+define i32 @inner() {
+  call void @extern()
+  %a1 = load volatile i32, i32* @a
+  %x1 = add i32 %a1,  %a1
+  %a2 = load volatile i32, i32* @a
+  %x2 = add i32 %x1, %a2
+  %a3 = load volatile i32, i32* @a
+  %x3 = add i32 %x2, %a3
+  %a4 = load volatile i32, i32* @a
+  %x4 = add i32 %x3, %a4
+  %a5 = load volatile i32, i32* @a
+  %x5 = add i32 %x3, %a5
+  ret i32 %x5
+}
+
+; @inner() should be inlined for -Os but not for -Oz.
+; OZ: call
+; OS-NOT: call
+define i32 @outer() optsize {
+   %r = call i32 @inner()
+   ret i32 %r
+}
+
+; @inner() should not be inlined for -Os and -Oz.
+; OZ: call
+; OS: call
+define i32 @outer2() minsize {
+   %r = call i32 @inner()
+   ret i32 %r
+}
+
+declare void @extern()

llvm/tools/lto/lto.cpp

   LTOCodeGenerator *CG = unwrap(cg);
   CG->setAttrs(codegen::getMAttrs());
 
-  if (OptLevel < '0' || OptLevel > '3')
-    report_fatal_error("Optimization level must be between 0 and 3");
+  if (OptLevel == 's')
+    OptLevel = '4';
+  else if (OptLevel == 'z')
+    OptLevel = '5';
+  if (OptLevel < '0' || OptLevel > '5')
+    report_fatal_error(
+        "Optimization level must be between 0 and 3 or Os or Oz");
   CG->setOptLevel(OptLevel - '0');
   CG->setFreestanding(EnableFreestanding);
   CG->setDisableVerify(DisableVerify);

   CodeGen->setFreestanding(EnableFreestanding);
 
   if (OptLevel.getNumOccurrences()) {
-    if (OptLevel < '0' || OptLevel > '3')
-      report_fatal_error("Optimization level must be between 0 and 3");
+    if (OptLevel == 's')
+      OptLevel = '4';
+    else if (OptLevel == 'z')
+      OptLevel = '5';
+    if (OptLevel < '0' || OptLevel > '5')
+      report_fatal_error(
+          "Optimization level must be between 0 and 3 or Os or Oz");
     CodeGen->setOptLevel(OptLevel - '0');
     switch (OptLevel) {
     case '0':