Pass code-model through Module IR to LTO which will use is

[cmtice]

Summary

Currently the code-model does not get saved in the module IR, so if a code model is specified when compiling with LTO, it gets lost and is not propagated properly to LTO. This patch (along with one to the Clang front end: https://reviews.llvm.org/D52323) fixes that.

Depends on https://reviews.llvm.org/D52323

Fixes PR33306.

[dexonsmith]

lib/IR/Module.cpp

522

Is this the correct linking behaviour for code model? I.e., is it illegal to link two object files together that have different code models?

[tmsriram]

lib/IR/Module.cpp

522

I believe it is undefined behavior to link two object files with different code models as the compiler has to generate additional code if a larger code model would be used. As an example:

extern int *a;
int main() {

return a[10];

}

$ clang a.cc -S -mcmodel=large && cat a.s
...
movabsq $a, %rax
movq (%rax), %rax
movl 40(%rax), %eax
popq %rbp
..

$ clang a.cc -S -mcmodel=small && cat a.s
...
movq a, %rax
movl 40(%rax), %eax
popq %rbp

The data would not be accessible with a 32-bit relocation for x86_64 and the compiler must explicitly handle this.

[tejohnson]

lib/IR/Module.cpp

522

Presumably the largest could be picked for code generation, but if it is illegal to link 2 native objects with different code models then the question is why we should support linking to LTO IR objects with different code models. Also, not sure how something like -code-model=kernel interact with the -code-model=small/medium/large.

Unless otherwise shown that we need to mix and match different code models, I think Error is the right behavior. Suggest adding a comment summarizing why it is being treated as an Error.

test/LTO/X86/codemodel-3.ll

3

The Inputs/codemodel-3.ll file is missing from patch.

[dexonsmith]

lib/IR/Module.cpp

522

Agreed; if it's not correct for normal object files we should error here as in the patch.

[cmtice]

Add comment explaining why mixing two code models is an error. Add missing Inputs/codemodel-3.ll file.

[tejohnson]

lib/LTO/LTOBackend.cpp

143

With this change the CodeModel passed in to the LTO link (if any) is ignored. I think this should be like the code that sets the RelocModel just above here.

(As an aside, please upload your patches with context. You can either do so automatically if you use the 'arc' tool, or I think there is a way to do when you create the diff directly - see https://llvm.org/docs/Phabricator.html)

[cmtice]

Use Conf.CodeModel in LTO, if it exists.

[tejohnson]

LGTM with a minor change suggested.

lib/LTO/LTOBackend.cpp

145

No need to do this conditionally - just unconditionally set CodeModel to M.getCodeModel() in the else since it should get None if not set on the Module.

include/llvm/IR/Module.h

 #define LLVM_IR_MODULE_H
 
 #include "llvm-c/Types.h"
+#include "llvm/ADT/Optional.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/StringMap.h"
 #include "llvm/ADT/StringRef.h"

   void setPIELevel(PIELevel::Level PL);
 /// @}
 
+  /// @}
+  /// @name Utility function for querying and setting code model
+  /// @{
+
+  /// Returns the code model (tiny, small, kernel, medium or large model)
+  Optional<CodeModel::Model> getCodeModel() const;
+
+  /// Set the code model (tiny, small, kernel, medium or large)
+  void setCodeModel(CodeModel::Model CL);
+  /// @}
+
   /// @name Utility functions for querying and setting PGO summary
   /// @{
 

lib/IR/Module.cpp

 
 #include "llvm/IR/Module.h"
 #include "SymbolTableListTraitsImpl.h"
+#include "llvm/ADT/Optional.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallString.h"
 #include "llvm/ADT/SmallVector.h"

   addModuleFlag(ModFlagBehavior::Max, "PIE Level", PL);
 }
 
+Optional<CodeModel::Model> Module::getCodeModel() const {
+  auto *Val = cast_or_null<ConstantAsMetadata>(getModuleFlag("Code Model"));
+
+  if (!Val)
+    return None;
+
+  return static_cast<CodeModel::Model>(
+      cast<ConstantInt>(Val->getValue())->getZExtValue());
+}
+
+void Module::setCodeModel(CodeModel::Model CL) {
+  // Linking object files with different code models is undefined behavior
+  // because the compiler would have to generate additional code (to span
+  // longer jumps) if a larger code model is used with a smaller one.
+  // Therefore we will treat attempts to mix code models as an error.
+  addModuleFlag(ModFlagBehavior::Error, "Code Model", CL);
+}
+
 void Module::setProfileSummary(Metadata *M) {
   addModuleFlag(ModFlagBehavior::Error, "ProfileSummary", M);
 }

lib/LTO/LTOBackend.cpp

 
   return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
       TheTriple, Conf.CPU, Features.getString(), Conf.Options, RelocModel,
-      Conf.CodeModel, Conf.CGOptLevel));
+      M.getCodeModel(), Conf.CGOptLevel));
 }
 
 static void runNewPMPasses(Config &Conf, Module &Mod, TargetMachine *TM,

test/LTO/X86/Inputs/codemodel-3.ll

+target triple = "x86_64-unknown-linux-gnu"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+!llvm.module.flags = !{!0, !1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 1, !"Code Model", i32 1}
+
+%struct.rtx_def = type { i16, i16 }
+
+define void @bar(%struct.rtx_def* %a, i8 %b, i32 %c) {
+  call void  @llvm.memset.p0struct.rtx_def.i32(%struct.rtx_def* align 4 %a, i8 %b, i32 %c, i1 true)
+  ret void
+}
+
+declare void @llvm.memset.p0struct.rtx_def.i32(%struct.rtx_def*, i8, i32, i1)

test/LTO/X86/codemodel-1.ll

+; RUN: llvm-as %s -o %t.o
+; RUN: llvm-lto2 run -r %t.o,_start,px %t.o -o %t.s
+; RUN: llvm-objdump -d %t.s.0 | FileCheck %s --check-prefix=CHECK-SMALL
+
+target triple = "x86_64-unknown-linux-gnu"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+!llvm.module.flags = !{!0, !1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 1, !"Code Model", i32 1}
+
+@data = internal constant [0 x i32] []
+
+define i32* @_start() nounwind readonly {
+entry:
+; CHECK-SMALL-LABEL:  _start:
+; CHECK-SMALL: leaq    (%rip), %rax
+    ret i32* getelementptr ([0 x i32], [0 x i32]* @data, i64 0, i64 0)
+}

test/LTO/X86/codemodel-2.ll

+; RUN: llvm-as %s -o %t.o
+; RUN: llvm-lto2 run -r %t.o,_start,px %t.o -o %t.s
+; RUN: llvm-objdump -d %t.s.0 | FileCheck %s --check-prefix=CHECK-LARGE
+
+target triple = "x86_64-unknown-linux-gnu"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+!llvm.module.flags = !{!0, !1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 1, !"Code Model", i32 4}
+
+@data = internal constant [0 x i32] []
+
+define i32* @_start() nounwind readonly {
+entry:
+; CHECK-LARGE-LABEL: _start:
+; CHECK-LARGE: movabsq $0, %rax
+    ret i32* getelementptr ([0 x i32], [0 x i32]* @data, i64 0, i64 0)
+}

test/LTO/X86/codemodel-3.ll

+; RUN: llvm-as %s -o %t0.o
+; RUN: llvm-as < %p/Inputs/codemodel-3.ll > %t1.o
+; RUN: not llvm-lto2 run -r %t0.o,_start,px -r %t1.o,bar,px %t0.o %t1.o -o %t2.s 2>&1 | FileCheck %s 
+
+target triple = "x86_64-unknown-linux-gnu"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+!llvm.module.flags = !{!0, !1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 1, !"Code Model", i32 4}
+
+@data = internal constant [0 x i32] []
+
+define i32* @_start() nounwind readonly {
+entry:
+    ret i32* getelementptr ([0 x i32], [0 x i32]* @data, i64 0, i64 0)
+}
+
+; CHECK: 'Code Model': IDs have conflicting values