Reland [X86] With large code model, put functions into .ltext with large section flag (#73037)

So that when mixing small and large text, large text stays out of the
way of the rest of the binary.

This is useful for mixing precompiled small code model object files and
built-from-source large code model binaries so that the the text
sections don't get merged.

The reland fixes an issue where a function in the large code model would reference small data without GOTOFF.

This was incorrectly reverted in 76f78ec.

Author: aeubanks

llvm/include/llvm/Target/TargetMachine.h

@@ -239,7 +239,7 @@ class TargetMachine {
   void setCodeModel(CodeModel::Model CM) { CMModel = CM; }
 
   void setLargeDataThreshold(uint64_t LDT) { LargeDataThreshold = LDT; }
-  bool isLargeData(const GlobalVariable *GV) const;
+  bool isLargeGlobalObject(const GlobalObject *GO) const;
 
   bool isPositionIndependent() const;
 

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

@@ -616,7 +616,7 @@ static unsigned getEntrySizeForKind(SectionKind Kind) {
 /// DataSections.
 static StringRef getSectionPrefixForGlobal(SectionKind Kind, bool IsLarge) {
   if (Kind.isText())
-    return ".text";
+    return IsLarge ? ".ltext" : ".text";
   if (Kind.isReadOnly())
     return IsLarge ? ".lrodata" : ".rodata";
   if (Kind.isBSS())
@@ -650,10 +650,7 @@ getELFSectionNameForGlobal(const GlobalObject *GO, SectionKind Kind,
     Name = ".rodata.cst";
     Name += utostr(EntrySize);
   } else {
-    bool IsLarge = false;
-    if (auto *GV = dyn_cast<GlobalVariable>(GO))
-      IsLarge = TM.isLargeData(GV);
-    Name = getSectionPrefixForGlobal(Kind, IsLarge);
+    Name = getSectionPrefixForGlobal(Kind, TM.isLargeGlobalObject(GO));
   }
 
   bool HasPrefix = false;
@@ -773,12 +770,8 @@ getGlobalObjectInfo(const GlobalObject *GO, const TargetMachine &TM) {
     Group = C->getName();
     IsComdat = C->getSelectionKind() == Comdat::Any;
   }
-  if (auto *GV = dyn_cast<GlobalVariable>(GO)) {
-    if (TM.isLargeData(GV)) {
-      assert(TM.getTargetTriple().getArch() == Triple::x86_64);
-      Flags |= ELF::SHF_X86_64_LARGE;
-    }
-  }
+  if (TM.isLargeGlobalObject(GO))
+    Flags |= ELF::SHF_X86_64_LARGE;
   return {Group, IsComdat, Flags};
 }
 

llvm/lib/Target/TargetMachine.cpp

@@ -39,13 +39,21 @@ TargetMachine::TargetMachine(const Target &T, StringRef DataLayoutString,
 
 TargetMachine::~TargetMachine() = default;
 
-bool TargetMachine::isLargeData(const GlobalVariable *GV) const {
-  if (getTargetTriple().getArch() != Triple::x86_64 || GV->isThreadLocal())
+bool TargetMachine::isLargeGlobalObject(const GlobalObject *GO) const {
+  if (getTargetTriple().getArch() != Triple::x86_64)
     return false;
 
   if (getCodeModel() != CodeModel::Medium && getCodeModel() != CodeModel::Large)
     return false;
 
+  if (isa<Function>(GO))
+    return getCodeModel() == CodeModel::Large;
+
+  auto *GV = cast<GlobalVariable>(GO);
+
+  if (GV->isThreadLocal())
+    return false;
+
   // Allowing large metadata sections in the presence of an explicit section is
   // useful, even if GCC does not allow them. However, we should not mark
   // certain well-known prefixes as large, because it would make the whole

llvm/lib/Target/X86/X86Subtarget.cpp

@@ -83,32 +83,20 @@ X86Subtarget::classifyLocalReference(const GlobalValue *GV) const {
   if (is64Bit()) {
     // 64-bit ELF PIC local references may use GOTOFF relocations.
     if (isTargetELF()) {
-      switch (TM.getCodeModel()) {
-      // 64-bit small code model is simple: All rip-relative.
-      case CodeModel::Tiny:
-        llvm_unreachable("Tiny codesize model not supported on X86");
-      case CodeModel::Small:
-      case CodeModel::Kernel:
-        return X86II::MO_NO_FLAG;
-
-      // The large PIC code model uses GOTOFF.
-      case CodeModel::Large:
+      CodeModel::Model CM = TM.getCodeModel();
+      assert(CM != CodeModel::Tiny &&
+             "Tiny codesize model not supported on X86");
+      // In the large code model, even referencing a global under the large data
+      // threshold which is considered "small", we need to use GOTOFF.
+      if (CM == CodeModel::Large)
         return X86II::MO_GOTOFF;
-
-      // Medium is a hybrid: RIP-rel for code and non-large data, GOTOFF for
-      // remaining DSO local data.
-      case CodeModel::Medium:
-        // Constant pool and jump table handling pass a nullptr to this
-        // function so we need to use isa_and_nonnull.
-        if (isa_and_nonnull<Function>(GV))
-          return X86II::MO_NO_FLAG; // All code is RIP-relative
-        if (auto *GVar = dyn_cast_or_null<GlobalVariable>(GV)) {
-          if (TM.isLargeData(GVar))
-            return X86II::MO_GOTOFF;
-        }
-        return X86II::MO_NO_FLAG;    // Local symbols use GOTOFF.
-      }
-      llvm_unreachable("invalid code model");
+      // Large objects use GOTOFF, otherwise use RIP-rel access.
+      if (auto *GO = dyn_cast_or_null<GlobalObject>(GV))
+        return TM.isLargeGlobalObject(GO) ? X86II::MO_GOTOFF
+                                          : X86II::MO_NO_FLAG;
+      // For non-GlobalObjects, the small and medium code models treat them as
+      // accessible with a RIP-rel access.
+      return X86II::MO_NO_FLAG;
     }
 
     // Otherwise, this is either a RIP-relative reference or a 64-bit movabsq,

llvm/test/CodeGen/X86/code-model-elf-text-sections.ll

@@ -0,0 +1,25 @@
+; RUN: llc < %s -relocation-model=pic -filetype=obj -code-model=small -o %t
+; RUN: llvm-readelf -S %t | FileCheck %s --check-prefix=SMALL
+; RUN: llc < %s -relocation-model=pic -filetype=obj -code-model=medium -o %t
+; RUN: llvm-readelf -S %t | FileCheck %s --check-prefix=SMALL
+; RUN: llc < %s -relocation-model=pic -filetype=obj -code-model=large -o %t
+; RUN: llvm-readelf -S %t | FileCheck %s --check-prefix=LARGE
+
+; RUN: llc < %s -relocation-model=pic -filetype=obj -code-model=small -function-sections -o %t
+; RUN: llvm-readelf -S %t | FileCheck %s --check-prefix=SMALL-DS
+; RUN: llc < %s -relocation-model=pic -filetype=obj -code-model=medium -function-sections -o %t
+; RUN: llvm-readelf -S %t | FileCheck %s --check-prefix=SMALL-DS
+; RUN: llc < %s -relocation-model=pic -filetype=obj -code-model=large -function-sections -o %t
+; RUN: llvm-readelf -S %t | FileCheck %s --check-prefix=LARGE-DS
+
+; SMALL: .text {{.*}} AX {{.*}}
+; SMALL-DS: .text.func {{.*}} AX {{.*}}
+; LARGE: .ltext {{.*}} AXl {{.*}}
+; LARGE-DS: .ltext.func {{.*}} AXl {{.*}}
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64--linux"
+
+define void @func() {
+  ret void
+}

llvm/test/CodeGen/X86/code-model-elf.ll

@@ -9,6 +9,7 @@
 ; RUN: llc -verify-machineinstrs < %s -relocation-model=pic    -code-model=medium -large-data-threshold=1000 | FileCheck %s --check-prefix=CHECK --check-prefix=MEDIUM-SMALL-DATA-PIC
 ; RUN: llc -verify-machineinstrs < %s -relocation-model=pic    -code-model=medium | FileCheck %s --check-prefix=CHECK --check-prefix=MEDIUM-PIC
 ; RUN: llc -verify-machineinstrs < %s -relocation-model=pic    -code-model=large  | FileCheck %s --check-prefix=CHECK --check-prefix=LARGE-PIC
+; RUN: llc -verify-machineinstrs < %s -relocation-model=pic    -code-model=large  -large-data-threshold=1000 | FileCheck %s --check-prefix=CHECK --check-prefix=LARGE-PIC
 
 ; Generated from this C source:
 ;

llvm/test/CodeGen/X86/pcsections.ll

@@ -19,12 +19,12 @@ define void @empty_no_aux() !pcsections !0 {
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    retq
 ; CHECK-NEXT:  .Lfunc_end0:
-; CHECK:       .section	section_no_aux,"awo",@progbits,.text
+; CHECK:       .section	section_no_aux,"awo",@progbits,.{{l?}}text
 ; CHECK-NEXT:  .Lpcsection_base0:
 ; DEFCM-NEXT:  .long	.Lfunc_begin0-.Lpcsection_base0
 ; LARGE-NEXT:  .quad	.Lfunc_begin0-.Lpcsection_base0
 ; CHECK-NEXT:  .long	.Lfunc_end0-.Lfunc_begin0
-; CHECK-NEXT:  .text
+; CHECK-NEXT:  .{{l?}}text
 entry:
   ret void
 }
@@ -35,15 +35,15 @@ define void @empty_aux() !pcsections !1 {
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    retq
 ; CHECK-NEXT:  .Lfunc_end1:
-; CHECK:       .section	section_aux,"awo",@progbits,.text
+; CHECK:       .section	section_aux,"awo",@progbits,.{{l?}}text
 ; CHECK-NEXT:  .Lpcsection_base1:
 ; DEFCM-NEXT:  .long	.Lfunc_begin1-.Lpcsection_base1
 ; LARGE-NEXT:  .quad	.Lfunc_begin1-.Lpcsection_base1
 ; CHECK-NEXT:  .long	.Lfunc_end1-.Lfunc_begin1
 ; CHECK-NEXT:  .long	10
 ; CHECK-NEXT:  .long	20
 ; CHECK-NEXT:  .long	30
-; CHECK-NEXT:  .text
+; CHECK-NEXT:  .{{l?}}text
 entry:
   ret void
 }
@@ -56,44 +56,44 @@ define i64 @multiple() !pcsections !0 {
 ; CHECK-NEXT:    movq
 ; CHECK-NEXT:    retq
 ; CHECK-NEXT:  .Lfunc_end2:
-; CHECK:       .section	section_no_aux,"awo",@progbits,.text
+; CHECK:       .section	section_no_aux,"awo",@progbits,.{{l?}}text
 ; CHECK-NEXT:  .Lpcsection_base2:
 ; DEFCM-NEXT:  .long	.Lfunc_begin2-.Lpcsection_base2
 ; LARGE-NEXT:  .quad	.Lfunc_begin2-.Lpcsection_base2
 ; CHECK-NEXT:  .long	.Lfunc_end2-.Lfunc_begin2
-; CHECK-NEXT:  .section	section_aux_42,"awo",@progbits,.text
+; CHECK-NEXT:  .section	section_aux_42,"awo",@progbits,.{{l?}}text
 ; CHECK-NEXT:  .Lpcsection_base3:
 ; DEFCM-NEXT:  .long	.Lpcsection0-.Lpcsection_base3
 ; LARGE-NEXT:  .quad	.Lpcsection0-.Lpcsection_base3
 ; CHECK-NEXT:  .long	42
-; CHECK-NEXT:  .section	section_aux_21264,"awo",@progbits,.text
+; CHECK-NEXT:  .section	section_aux_21264,"awo",@progbits,.{{l?}}text
 ; CHECK-NEXT:  .Lpcsection_base4:
 ; DEFCM-NEXT:  .long	.Lpcsection0-.Lpcsection_base4
 ; LARGE-NEXT:  .quad	.Lpcsection0-.Lpcsection_base4
 ; CHECK-NEXT:  .long	21264
-; CHECK-NEXT:  .text
+; CHECK-NEXT:  .{{l?}}text
 entry:
   %0 = load i64, ptr @bar, align 8, !pcsections !2
   ret i64 %0
 }
 
 define void @multiple_uleb128() !pcsections !6 {
 ; CHECK-LABEL: multiple_uleb128:
-; CHECK:       .section	section_aux,"awo",@progbits,.text
+; CHECK:       .section	section_aux,"awo",@progbits,.{{l?}}text
 ; CHECK-NEXT:  .Lpcsection_base5:
 ; DEFCM-NEXT:  .long	.Lfunc_begin3-.Lpcsection_base5
 ; LARGE-NEXT:  .quad	.Lfunc_begin3-.Lpcsection_base5
 ; CHECK-NEXT:  .uleb128	.Lfunc_end3-.Lfunc_begin3
 ; CHECK-NEXT:  .byte	42
 ; CHECK-NEXT:  .ascii	"\345\216&"
 ; CHECK-NEXT:  .byte	255
-; CHECK-NEXT:  .section	section_aux_21264,"awo",@progbits,.text
+; CHECK-NEXT:  .section	section_aux_21264,"awo",@progbits,.{{l?}}text
 ; CHECK-NEXT:  .Lpcsection_base6:
 ; DEFCM-NEXT:  .long	.Lfunc_begin3-.Lpcsection_base6
 ; LARGE-NEXT:  .quad	.Lfunc_begin3-.Lpcsection_base6
 ; CHECK-NEXT:  .long	.Lfunc_end3-.Lfunc_begin3
 ; CHECK-NEXT:  .long	21264
-; CHECK-NEXT:  .text
+; CHECK-NEXT:  .{{l?}}text
 entry:
   ret void
 }

llvm/test/ExecutionEngine/OrcLazy/debug-objects-elf-minimal.ll

@@ -44,7 +44,7 @@
 ; RUN:     --generate=__dump_jit_debug_objects %s | llvm-objdump --section-headers - | \
 ; RUN:     FileCheck --check-prefix=CHECK_LOAD_ADDR %s
 ;
-; CHECK_LOAD_ADDR-NOT: {{[0-9]*}} .text {{.*}} 0000000000000000 TEXT
+; CHECK_LOAD_ADDR-NOT: {{[0-9]*}} .ltext {{.*}} 0000000000000000 TEXT
 
 target triple = "x86_64-unknown-unknown-elf"