Fix SUBALIGN linker script feature

include/eld/Diagnostics/DiagLDScript.inc

@@ -122,6 +122,8 @@ DIAG(error_cannot_specify_lma_and_memory_region, DiagnosticEngine::Error,
      "cannot specify AT and LMA memory region for output section : %0 defined in script %1")
 DIAG(warn_non_power_of_2_value_to_align_builtin, DiagnosticEngine::Warning,
      "%0: non-power-of-2 value 0x%1 passed to ALIGN builtin function")
+DIAG(warn_subalign_less_than_section_alignment, DiagnosticEngine::Warning,
+     "SUBALIGN(0x%0) is less than the section alignment (0x%1) for section '%2'")
 DIAG(error_non_power_of_2_value_to_align_output_section, DiagnosticEngine::Error,
      "%0: non-power-of-2 value 0x%1 passed to ALIGN in '%2' output section description, value must "
      "be 0 or a power of 2")

include/eld/Object/ObjectLinker.h

@@ -312,6 +312,8 @@ class ObjectLinker {
 
   bool initializeOutputSectionsAndRunPlugin();
 
+  void applySubAlign();
+
   // Get Plugin list for Relocation registration callback.
   LinkerScript::PluginVectorT getLinkerPluginWithLinkerConfigs();
 

lib/LayoutMap/TextLayoutPrinter.cpp

@@ -108,6 +108,10 @@ void TextLayoutPrinter::printOnlyLayoutSection(GNULDBackend const &Backend,
   outputStream().write_hex(Section->size());
   outputStream() << "\t# Alignment: 0x";
   outputStream().write_hex(Section->getAddrAlign());
+  if (OS->prolog().hasSubAlign()) {
+    outputStream() << ", SubAlign: 0x";
+    outputStream().write_hex(OS->prolog().subAlign().resultOrZero());
+  }
   if (Section->isAlloc())
     printSegments(Backend, OS);
   outputStream() << "\n";
@@ -277,6 +281,12 @@ void TextLayoutPrinter::printSection(GNULDBackend const &Backend,
   outputStream() << ", Alignment: 0x";
   outputStream().write_hex(Section->getAddrAlign());
 
+  // Print SubAlign if present
+  if (OS->prolog().hasSubAlign()) {
+    outputStream() << ", SubAlign: 0x";
+    outputStream().write_hex(OS->prolog().subAlign().resultOrZero());
+  }
+
   // Print flags
   outputStream() << ", Flags: "
                  << ELFSection::getELFPermissionsStr(Section->getFlags());

lib/Object/ObjectLinker.cpp

@@ -61,6 +61,7 @@
 #include "eld/SymbolResolver/ResolveInfo.h"
 #include "eld/Target/ELFFileFormat.h"
 #include "eld/Target/GNULDBackend.h"
+#include "eld/Target/LDFileFormat.h"
 #include "eld/Target/Relocator.h"
 #include "llvm/ADT/ScopeExit.h"
 #include "llvm/ADT/StringSwitch.h"
@@ -81,6 +82,7 @@
 #include <chrono>
 #include <mutex>
 #include <sstream>
+#include <unordered_set>
 
 using namespace llvm;
 using namespace eld;
@@ -970,7 +972,7 @@ bool ObjectLinker::sortSections(RuleContainer *I, bool SortRule) {
 bool ObjectLinker::createOutputSection(ObjectBuilder &Builder,
                                        OutputSectionEntry *Output,
                                        bool PostLayout) {
-  uint64_t OutAlign = 0x0, InAlign = 0x0;
+  uint64_t OutAlign = 0x0;
   bool IsPartialLink = (LinkerConfig::Object == ThisConfig.codeGenType());
 
   ELFSection *OutSect = Output->getSection();
@@ -980,15 +982,6 @@ bool ObjectLinker::createOutputSection(ObjectBuilder &Builder,
   OutputSectionEntry::iterator In, InBegin, InEnd;
   InBegin = Output->begin();
   InEnd = Output->end();
-  bool HasSubAlign = false;
-
-  // force input alignment from ldscript if any
-  if (Output->prolog().hasSubAlign()) {
-    Output->prolog().subAlign().eval();
-    Output->prolog().subAlign().commit();
-    InAlign = Output->prolog().subAlign().result();
-    HasSubAlign = true;
-  }
 
   // force output alignment from ldscript if any
   if (Output->prolog().hasAlign()) {
@@ -1030,12 +1023,6 @@ bool ObjectLinker::createOutputSection(ObjectBuilder &Builder,
       }
       InSect->setAddrAlign(Alignment);
     }
-    if (HasSubAlign && (InSect->getAddrAlign() < InAlign)) {
-      if (InSect->getFragmentList().size()) {
-        InSect->getFragmentList().front()->setAlignment(InAlign);
-        InSect->setAddrAlign(InAlign);
-      }
-    }
     if (InSect->getFragmentList().size() && !FirstNonEmptyRule)
       FirstNonEmptyRule = *In;
 
@@ -1286,6 +1273,9 @@ bool ObjectLinker::mergeSections() {
   {
     eld::RegisterTimer T("Create Output Section", "Merge Sections",
                          ThisConfig.options().printTimingStats());
+    // Prepass: apply SUBALIGN to first fragments per input section per output
+    // section serially to avoid races in parallel output creation.
+    applySubAlign();
     std::vector<OutputSectionEntry *> OutSections;
     for (Out = OutBegin; Out != OutEnd; ++Out) {
       OutSections.push_back(*Out);
@@ -1329,6 +1319,41 @@ bool ObjectLinker::initializeOutputSectionsAndRunPlugin() {
   return runOutputSectionIteratorPlugin();
 }
 
+void ObjectLinker::applySubAlign() {
+  std::unordered_set<ELFSection *> seen;
+
+  for (auto *O : ThisModule->getScript().sectionMap()) {
+    auto &prolog = O->prolog();
+    if (!prolog.hasSubAlign())
+      continue;
+    uint64_t subAlign;
+    prolog.subAlign().eval();
+    prolog.subAlign().commit();
+    subAlign = prolog.subAlign().result();
+
+    for (RuleContainer *R : *O) {
+      ELFSection *inSect = R->getSection();
+      for (Fragment *F : inSect->getFragmentList()) {
+        ELFSection *owningSect = F->getOwningSection();
+        if (owningSect->getKind() == LDFileFormat::Kind::OutputSectData) {
+          continue;
+        }
+        if (owningSect && seen.insert(owningSect).second) {
+          // Warn if SUBALIGN is reducing the section alignment
+          if (ThisConfig.showLinkerScriptWarnings() && F->alignment() > subAlign) {
+            ThisConfig.raise(Diag::warn_subalign_less_than_section_alignment)
+                << utility::toHex(subAlign) << utility::toHex(F->alignment())
+                << owningSect->getLocation(0, ThisConfig.options());
+          }
+          F->setAlignment(subAlign);
+          inSect->setAddrAlign(std::max(
+              static_cast<uint64_t>(inSect->getAddrAlign()), subAlign));
+        }
+      }
+    }
+  }
+}
+
 void ObjectLinker::assignOffset(OutputSectionEntry *Out) {
   int64_t O = 0;
   const ELFSection *OutSection = Out->getSection();

test/Common/standalone/SubAlign/Inputs/1.c

@@ -1,3 +1,3 @@
 int foo() { return 0; }
-int bar() { return 0; }
+int bar() { return foo(); }
 int car() { return 0; }

test/Common/standalone/SubAlign/Inputs/2.c

@@ -0,0 +1,8 @@
+__attribute__((aligned((0x4))))
+int foo() { return 0; }
+
+__attribute__((aligned((0x4))))
+int bar() { return foo(); }
+
+__attribute__((aligned((0x4))))
+int car() { return 0; }

test/Common/standalone/SubAlign/Inputs/script.2.t

@@ -0,0 +1,14 @@
+PHDRS {
+  A PT_LOAD;
+}
+
+SECTIONS {
+  .a : SUBALIGN(64) {
+   . = . + 1;
+    *(.text*)
+  }:A
+  .b : SUBALIGN(64) {
+   . = . + 1;
+  }:A
+  /DISCARD/ : { *(.ARM.exidx*) }
+}

test/Common/standalone/SubAlign/Inputs/script.data.t

@@ -0,0 +1,12 @@
+SECTIONS {
+  .explicit_data : SUBALIGN(64) {
+    BYTE(0x11)
+    SHORT(0x2222)
+    LONG(0x44444444)
+    QUAD(0x8888888888888888)
+    *(.text*)
+    BYTE(0xAA)
+    QUAD(0xBBBBBBBBBBBBBBBB)
+  }
+  /DISCARD/ : { *(.ARM.exidx*) }
+}

test/Common/standalone/SubAlign/Inputs/script.discard.t

@@ -0,0 +1,14 @@
+PHDRS {
+  A PT_LOAD;
+}
+
+SECTIONS {
+  .keep : SUBALIGN(64) {
+    *(.text.foo)
+    *(.text.bar)
+  }:A
+  /DISCARD/ : SUBALIGN(32) {
+    *(.text.car)
+    *(.ARM.exidx*)
+  }
+}

test/Common/standalone/SubAlign/Inputs/script.empty.t

@@ -0,0 +1,13 @@
+SECTIONS {
+  .nonempty : SUBALIGN(64) {
+    *(.text.foo)
+  }
+  .empty1 : SUBALIGN(32) {
+    *(.nonexistent1)
+  }
+  .empty2 : SUBALIGN(128) {
+    /* Empty section with just location counter */
+    . = . + 0;
+  }
+  /DISCARD/ : { *(.ARM.exidx*) }
+}

test/Common/standalone/SubAlign/Inputs/script.nonalloc.t

@@ -0,0 +1,12 @@
+SECTIONS {
+  .text : SUBALIGN(64) {
+    *(.text*)
+  }
+  .debug_info : SUBALIGN(32) {
+    *(.debug_info)
+  }
+  .comment : SUBALIGN(16) {
+    *(.comment)
+  }
+  /DISCARD/ : { *(.ARM.exidx*) }
+}

test/Common/standalone/SubAlign/Inputs/script.partial.t

@@ -0,0 +1,6 @@
+SECTIONS {
+  .partial : SUBALIGN(64) {
+    *(.text*)
+  }
+  /DISCARD/ : { *(.ARM.exidx*) }
+}

test/Common/standalone/SubAlign/Inputs/script.reduce.t

@@ -0,0 +1,6 @@
+SECTIONS {
+  .reduce : SUBALIGN(2) {
+    *(.text*)
+  }
+  /DISCARD/ : { *(.ARM.exidx*) }
+}

test/Common/standalone/SubAlign/Inputs/script.relocs.t

@@ -0,0 +1,7 @@
+SECTIONS {
+  .reloc_text : SUBALIGN(64) {
+    *(.text*)
+    *(.rel.*)
+  }
+  /DISCARD/ : { *(.ARM.exidx*) }
+}

test/Common/standalone/SubAlign/Inputs/script.shared.t

@@ -0,0 +1,6 @@
+SECTIONS {
+  .shared_text (0x1000) : SUBALIGN(64) {
+    *(.text*)
+  }
+  /DISCARD/ : { *(.ARM.exidx*) }
+}

test/Common/standalone/SubAlign/SubAlign.test

@@ -3,12 +3,100 @@
 # Test for linker script SUBALIGN directive.
 #END_COMMENT
 #START_TEST
+
+# Basic SUBALIGN test
 RUN: %clang %clangopts -c %p/Inputs/1.c -o %t1.1.o -ffunction-sections
-RUN: %link -MapStyle txt %linkopts %t1.1.o -T %p/Inputs/script.t -o %t1.1.o -Map %t2.map
-RUN: %filecheck %s < %t2.map
-#CHECK: .a {{.*}} Alignment: 0x40
-#CHECK: .text.foo	0x40
-#CHECK: .text.bar	0x80
-#CHECK: .text.car	0xc0
-#CHECK: .b {{.*}} Alignment: 0x1
+RUN: %link -MapStyle txt %linkopts %t1.1.o -T %p/Inputs/script.t -o %t1.1.out -Map %t1.1.map.txt
+RUN: %filecheck %s < %t1.1.map.txt
+RUN: %link -MapStyle txt %linkopts %t1.1.o -T %p/Inputs/script.2.t -o %t1.2.out -Map %t1.2.map.txt
+RUN: %filecheck %s < %t1.1.map.txt
+RUN: %filecheck %s < %t1.2.map.txt
+CHECK: .a {{.*}} Alignment: 0x40
+CHECK: .text.foo	0x40
+CHECK: .text.bar	0x80
+CHECK: .text.car	0xc0
+CHECK: .b {{.*}} Alignment: 0x1
+
+# Test SubAlign display in map file
+RUN: %filecheck %s --check-prefix=SUBALIGN-DISPLAY < %t1.1.map.txt
+SUBALIGN-DISPLAY: .a {{.*}} Alignment: 0x40, SubAlign: 0x40
+SUBALIGN-DISPLAY: .b {{.*}} Alignment: 0x1, SubAlign: 0x40
+
+# Test SUBALIGN with DISCARD sections
+RUN: %link -MapStyle txt %linkopts %t1.1.o -T %p/Inputs/script.discard.t -o %t.discard.out -Map %t.discard.map.txt
+RUN: %filecheck %s --check-prefix=DISCARD < %t.discard.map.txt
+DISCARD: .keep {{.*}} Alignment: 0x40, SubAlign: 0x40
+DISCARD: .text.foo	0x0
+DISCARD: .text.bar	0x40
+DISCARD: /DISCARD/ {{.*}} Alignment: 0x1, SubAlign: 0x20
+DISCARD: # .text.car
+
+# Test SUBALIGN with emit-relocs
+RUN: %link -MapStyle txt %linkopts %t1.1.o -T %p/Inputs/script.relocs.t -o %t.relocs.out -Map %t.relocs.map.txt --emit-relocs
+RUN: %filecheck %s --check-prefix=RELOCS < %t.relocs.map.txt
+RUN: %readelf -r %t.relocs.out | %filecheck %s --check-prefix=RELOCS-READELF
+RELOCS: .reloc_text {{.*}} Alignment: 0x40, SubAlign: 0x40
+RELOCS: .text.foo	0x0
+RELOCS: .text.bar	0x40
+RELOCS: .text.car	0x80
+#RELOCS-READELF: Relocation section '.rel{{a?}}.reloc_text' {{.*}}
+
+# Test SUBALIGN with empty sections
+RUN: %link -MapStyle txt %linkopts %t1.1.o -T %p/Inputs/script.empty.t -o %t.empty.out -Map %t.empty.map.txt
+RUN: %filecheck %s --check-prefix=EMPTY < %t.empty.map.txt
+EMPTY: .nonempty {{.*}} Alignment: 0x40, SubAlign: 0x40
+EMPTY: .text.foo	0x0
+EMPTY: .empty1 {{.*}} 0x0 {{.*}} Alignment: 0x1, SubAlign: 0x20
+EMPTY: .empty2 {{.*}} 0x0 {{.*}} Alignment: 0x1, SubAlign: 0x80
+
+# Test SUBALIGN with explicit data (BYTE/QUAD/...)
+RUN: %link -MapStyle txt %linkopts %t1.1.o -T %p/Inputs/script.data.t -o %t.data.out -Map %t.data.map.txt
+RUN: %filecheck %s --check-prefix=DATA < %t.data.map.txt
+DATA: .explicit_data {{.*}} Alignment: 0x40, SubAlign: 0x40
+DATA: BYTE (0x11) 0x0 0x1
+DATA: SHORT (0x2222) 0x1 0x3
+DATA: LONG (0x44444444) 0x3 0x7
+DATA: QUAD (0x8888888888888888) 0x7 0xf
+DATA: .text.foo 0x40
+DATA: .text.bar 0x80
+
+# Test SUBALIGN with partial linking
+RUN: %link %linkopts %t1.1.o -T %p/Inputs/script.partial.t -o %t.partial.tmp.o -r -Map %t.partial.map.txt
+RUN: %filecheck %s --check-prefix=PARTIAL < %t.partial.map.txt
+PARTIAL: .partial {{.*}} Alignment: 0x40, SubAlign: 0x40
+PARTIAL: .text.foo	0x0
+PARTIAL: .text.bar	0x40
+PARTIAL: .text.car	0x80
+
+# Test SUBALIGN with shared libraries
+RUN: %clang %clangopts -c %p/Inputs/1.c -o %t.shared.o -ffunction-sections -fPIC
+RUN: %link -MapStyle txt %linkopts %t.shared.o -T %p/Inputs/script.shared.t -o %t.shared.so -Map %t.shared.map.txt -shared
+RUN: %filecheck %s --check-prefix=SHARED < %t.shared.map.txt
+#END_TEST
+SHARED: .shared_text {{.*}} Alignment: 0x40, SubAlign: 0x40
+SHARED: .text.foo	0x1000
+SHARED: .text.bar	0x1040
+SHARED: .text.car	0x1080
+
+# Test SUBALIGN with non-allocatable sections
+RUN: %clang %clangopts -c %p/Inputs/1.c -o %t.nonalloc.o -ffunction-sections -g
+RUN: %link -MapStyle txt %linkopts %t.nonalloc.o -T %p/Inputs/script.nonalloc.t -o %t.nonalloc.out -Map %t.nonalloc.map.txt
+RUN: %filecheck %s --check-prefix=NONALLOC < %t.nonalloc.map.txt
+NONALLOC: .text {{.*}} Alignment: 0x40, SubAlign: 0x40
+NONALLOC: .text.foo	0x0
+NONALLOC: .text.bar	0x40
+NONALLOC: .text.car	0x80
+NONALLOC: .debug_info {{.*}} Alignment: 0x20, SubAlign: 0x20
+NONALLOC: .comment {{.*}} Alignment: 0x10, SubAlign: 0x10
+
+# Test SUBALIGN reducing section alignment (should warn)
+RUN: %clang %clangopts -c %p/Inputs/2.c -o %t1.reduce.o -ffunction-sections
+RUN: %link -MapStyle txt %linkopts %t1.reduce.o -T %p/Inputs/script.reduce.t -o %t1.1.out -Map %t.reduce.map.txt -Wlinker-script 2>&1 | %filecheck %s --check-prefix=REDUCE-WARN
+RUN: %link -MapStyle txt %linkopts %t1.reduce.o -T %p/Inputs/script.reduce.t -o %t1.1.out -Map %t.reduce.map.txt -Wno-linker-script 2>&1 | %filecheck %s --allow-empty --check-prefix=REDUCE-NOWARN
+RUN: %filecheck %s --check-prefix=REDUCE < %t.reduce.map.txt
+REDUCE-WARN: Warning: SUBALIGN(0x2) is less than the section alignment (0x{{.*}}) for section '{{.*reduce.o}}:(.text.foo)'
+REDUCE-WARN: Warning: SUBALIGN(0x2) is less than the section alignment (0x{{.*}}) for section '{{.*reduce.o}}:(.text.bar)'
+REDUCE-WARN: Warning: SUBALIGN(0x2) is less than the section alignment (0x{{.*}}) for section '{{.*reduce.o}}:(.text.car)'
+REDUCE-NOWARN-NOT: Warning
+REDUCE: .reduce {{.*}} Alignment: {{.*}}, SubAlign: 0x2
 #END_TEST