[llvm] [SystemZ][z/OS] Add z/OS archive writing support (PR #200087)
Amy Kwan via llvm-commits
llvm-commits at lists.llvm.org
Mon Jul 13 13:49:56 PDT 2026
https://github.com/amy-kwan updated https://github.com/llvm/llvm-project/pull/200087
>From ddffa75edc99a5f73aeaf5db60c4da13dde6023e Mon Sep 17 00:00:00 2001
From: Amy Kwan <amy.kwan1 at ibm.com>
Date: Wed, 27 May 2026 18:56:30 -0500
Subject: [PATCH 1/9] [SystemZ][z/OS] Add z/OS archive writing support
This patch implement z/OS archive writing in ArchiveWriter and adds the option
`llvm-ar --format=zos`.
Moreover, This patch teaches `llvm-ar` to emit z/OS-compatible archives by:
- Detecting GOFF object files as z/OS members
- Writing z/OS member headers and archive magic
- Converting archive headers and symbol-name string tables to EBCDIC
- Emitting z/OS symbol table entries
- Using EBCDIC newline padding bytes
- Add a z/OS-specific workaround for empty symbol tables by emitting a dummy
blank symbol to satisfy binder requirements.
---
llvm/lib/Object/ArchiveWriter.cpp | 75 ++++++++++++++++++++++++--
llvm/test/tools/llvm-ar/zos-write.test | 36 +++++++++++++
llvm/tools/llvm-ar/llvm-ar.cpp | 18 ++++++-
3 files changed, 125 insertions(+), 4 deletions(-)
create mode 100644 llvm/test/tools/llvm-ar/zos-write.test
diff --git a/llvm/lib/Object/ArchiveWriter.cpp b/llvm/lib/Object/ArchiveWriter.cpp
index 3b2d8c30bb301..8479340af26a3 100644
--- a/llvm/lib/Object/ArchiveWriter.cpp
+++ b/llvm/lib/Object/ArchiveWriter.cpp
@@ -20,6 +20,7 @@
#include "llvm/Object/COFF.h"
#include "llvm/Object/COFFImportFile.h"
#include "llvm/Object/Error.h"
+#include "llvm/Object/GOFFObjectFile.h"
#include "llvm/Object/IRObjectFile.h"
#include "llvm/Object/MachO.h"
#include "llvm/Object/ObjectFile.h"
@@ -70,6 +71,8 @@ object::Archive::Kind NewArchiveMember::detectKindFromObject() const {
if (isa<object::COFFObjectFile>(**OptionalObject) ||
isa<object::COFFImportFile>(**OptionalObject))
return object::Archive::K_COFF;
+ if (isa<object::GOFFObjectFile>(**OptionalObject))
+ return object::Archive::K_ZOS;
return object::Archive::K_GNU;
}
@@ -186,6 +189,10 @@ static bool isCOFFArchive(object::Archive::Kind Kind) {
return Kind == object::Archive::K_COFF;
}
+static bool isZOSArchive(object::Archive::Kind Kind) {
+ return Kind == object::Archive::K_ZOS;
+}
+
static bool isBSDLike(object::Archive::Kind Kind) {
switch (Kind) {
case object::Archive::K_GNU:
@@ -253,6 +260,24 @@ printBSDMemberHeader(raw_ostream &Out, uint64_t Pos, StringRef Name,
Out.write(uint8_t(0));
}
+static void
+printZOSMemberHeader(raw_ostream &Out, uint64_t Pos, StringRef Name,
+ const sys::TimePoint<std::chrono::seconds> &ModTime,
+ unsigned UID, unsigned GID, unsigned Perms,
+ uint64_t Size) {
+ std::string AHeader;
+ raw_string_ostream AOut(AHeader);
+ if (Name.size() <= 16) {
+ printWithSpacePadding(AOut, Twine(Name), 16);
+ printRestOfMemberHeader(AOut, ModTime, UID, GID, Perms, Size);
+ } else {
+ printBSDMemberHeader(AOut, Pos, Name, ModTime, UID, GID, Perms, Size);
+ }
+ SmallString<256> EHeader;
+ ConverterEBCDIC::convertToEBCDIC(AHeader, EHeader);
+ Out << EHeader.str();
+}
+
static void
printBigArchiveMemberHeader(raw_ostream &Out, StringRef Name,
const sys::TimePoint<std::chrono::seconds> &ModTime,
@@ -306,6 +331,9 @@ printMemberHeader(raw_ostream &Out, uint64_t Pos, raw_ostream &StringTable,
if (isBSDLike(Kind))
return printBSDMemberHeader(Out, Pos, MemberName, ModTime, M.UID, M.GID,
M.Perms, Size);
+ if (isZOSArchive(Kind))
+ return printZOSMemberHeader(Out, Pos, MemberName, ModTime, M.UID, M.GID,
+ M.Perms, Size);
if (!useStringTable(Thin, MemberName))
return printGNUSmallMemberHeader(Out, MemberName, ModTime, M.UID, M.GID,
M.Perms, Size);
@@ -392,8 +420,11 @@ static uint64_t computeSymbolTableSize(object::Archive::Kind Kind,
uint64_t Size = OffsetSize; // Number of entries
if (isBSDLike(Kind))
Size += NumSyms * OffsetSize * 2; // Table
+ else if (isZOSArchive(Kind))
+ Size += NumSyms * OffsetSize * 2; // MemberOffset + symbol flags
else
Size += NumSyms * OffsetSize; // Table
+
if (isBSDLike(Kind))
Size += OffsetSize; // byte count
Size += StringTableSize;
@@ -453,6 +484,10 @@ static void writeSymbolTableHeader(raw_ostream &Out, object::Archive::Kind Kind,
} else if (isAIXBigArchive(Kind)) {
printBigArchiveMemberHeader(Out, "", now(Deterministic), 0, 0, 0, Size,
PrevMemberOffset, NextMemberOffset);
+ } else if (isZOSArchive(Kind)) {
+ const char *Name = "__.SYMDEF";
+ printZOSMemberHeader(Out, Out.tell(), Name, now(Deterministic), 0, 0, 0,
+ Size);
} else {
const char *Name = is64BitKind(Kind) ? "/SYM64" : "";
printGNUSmallMemberHeader(Out, Name, now(Deterministic), 0, 0, 0, Size);
@@ -613,6 +648,10 @@ static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
StringTable.size(), &Pad);
writeSymbolTableHeader(Out, Kind, Deterministic, Size, PrevMemberOffset,
NextMemberOffset);
+ // Padding size is not included in the Size field of the z/OS symbol table
+ // header
+ if (isZOSArchive(Kind))
+ Size -= Pad;
if (isBSDLike(Kind))
printNBits(Out, Kind, NumSyms * 2 * OffsetSize);
@@ -633,6 +672,8 @@ static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
if (isBSDLike(Kind))
printNBits(Out, Kind, StringOffset);
printNBits(Out, Kind, Pos); // member offset
+ if (isZOSArchive(Kind))
+ printNBits(Out, Kind, Pos); // symbol flags
}
Pos += M.Header.size() + M.Data.size() + M.Padding.size();
}
@@ -640,7 +681,16 @@ static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
if (isBSDLike(Kind))
// byte count of the string table
printNBits(Out, Kind, StringTable.size());
- Out << StringTable;
+ if (isZOSArchive(Kind)) {
+ std::string AStringTable;
+ raw_string_ostream AOut(AStringTable);
+ AOut << StringTable;
+ SmallString<256> EStringTable;
+ ConverterEBCDIC::convertToEBCDIC(AStringTable, EStringTable);
+ Out << EStringTable.str();
+ } else {
+ Out << StringTable;
+ }
while (Pad--)
Out.write(uint8_t(0));
@@ -785,6 +835,8 @@ computeMemberData(raw_ostream &StringTable, raw_ostream &SymNames,
LLVMContext &Context, ArrayRef<NewArchiveMember> NewMembers,
std::optional<bool> IsEC, function_ref<void(Error)> Warn) {
static char PaddingData[8] = {'\n', '\n', '\n', '\n', '\n', '\n', '\n', '\n'};
+ static char ZOSPaddingData[8] = {0x15, 0x15, 0x15, 0x15,
+ 0x15, 0x15, 0x15, 0x15}; // EBCDIC newlines.
uint64_t Pos =
isAIXBigArchive(Kind) ? sizeof(object::BigArchive::FixLenHdr) : 0;
@@ -846,6 +898,9 @@ computeMemberData(raw_ostream &StringTable, raw_ostream &SymNames,
Entry.second = Entry.second > 1 ? 1 : 0;
}
+ uint32_t LastZosObjIndex =
+ UINT_MAX; // Only set when writing symbol table in z/OS archive.
+
for (const NewArchiveMember &M : NewMembers) {
MemberData &D = Ret.emplace_back();
D.Data = M.Buf->getBuffer();
@@ -858,7 +913,6 @@ computeMemberData(raw_ostream &StringTable, raw_ostream &SymNames,
if (!SymFileOrErr)
return createFileError(M.MemberName, SymFileOrErr.takeError());
D.SymFile = std::move(*SymFileOrErr);
-
if (SymMap && D.SymFile.get()) {
auto COFFObj = dyn_cast<COFFObjectFile>(D.SymFile.get());
std::optional<MemoryBufferRef> HybridView;
@@ -887,6 +941,9 @@ computeMemberData(raw_ostream &StringTable, raw_ostream &SymNames,
.str();
}
}
+
+ if (isZOSArchive(Kind) && D.SymFile.get())
+ LastZosObjIndex = Ret.size() - 1;
}
}
@@ -941,7 +998,8 @@ computeMemberData(raw_ostream &StringTable, raw_ostream &SymNames,
isDarwin(Kind) ? offsetToAlignment(D.Data.size(), Align(8)) : 0;
unsigned TailPadding =
offsetToAlignment(D.Data.size() + MemberPadding, Align(2));
- D.Padding = StringRef(PaddingData, MemberPadding + TailPadding);
+ D.Padding = StringRef(isZOSArchive(Kind) ? ZOSPaddingData : PaddingData,
+ MemberPadding + TailPadding);
StringRef MemberName = D.HybridName.size() ? D.HybridName : M->MemberName;
@@ -1006,6 +1064,15 @@ computeMemberData(raw_ostream &StringTable, raw_ostream &SymNames,
if (D.SymFile)
HasObject = true;
}
+ // On z/OS, when there are no symbols, add a dummy blank symbol
+ // into the symbol table. This is done since the z/OS binder:
+ // - emits an error if there is no symbol table in the archive
+ // - emits an error if the symbol table has 0 symbols
+ // - should not find any references to a blank symbol
+ if ((LastZosObjIndex == Index) && (SymNames.tell() == 0)) {
+ D.Symbols.push_back(0);
+ SymNames << ' ' << '\0';
+ }
Pos += D.Header.size() + D.Data.size() + D.Padding.size();
}
@@ -1180,6 +1247,8 @@ Error writeArchiveToStream(raw_ostream &Out,
Out << "!<thin>\n";
else if (isAIXBigArchive(Kind))
Out << "<bigaf>\n";
+ else if (isZOSArchive(Kind))
+ Out << ZOSArchiveMagic;
else
Out << "!<arch>\n";
diff --git a/llvm/test/tools/llvm-ar/zos-write.test b/llvm/test/tools/llvm-ar/zos-write.test
new file mode 100644
index 0000000000000..447e4562b83aa
--- /dev/null
+++ b/llvm/test/tools/llvm-ar/zos-write.test
@@ -0,0 +1,36 @@
+# Test writing a z/OS archive.
+
+# Create a z/OS archive from a plain text file, without a symbol table.
+# RUN: rm -rf %t.dir && mkdir -p %t.dir
+# RUN: printf 'abcd' > %t.dir/test.txt
+# RUN: llvm-ar rcSD --format=zos %t.a %t.dir/test.txt
+
+# The archive should be written and read back successfully through llvm-ar.
+# RUN: llvm-ar t %t.a | FileCheck %s --check-prefix=LIST
+# RUN: llvm-ar p %t.a test.txt | FileCheck %s --check-prefix=CONTENT
+
+# Check raw bytes:
+# - The first 8 bytes are the z/OS archive magic in EBCDIC.
+# - Bytes 66..67 are the first member header terminator (`\n in EBCDIC => 79 15)
+# RUN: %python -c "d=open(r'%t.a','rb').read(); print(d[:8].hex()); print(d[66:68].hex())" | FileCheck %s --check-prefix=BYTES
+
+# LIST: test.txt
+# CONTENT: abcd
+# BYTES: 5a4c819983886e15
+# BYTES-NEXT: 7915
+
+# Odd-sized members should be padded with EBCDIC newline (0x15).
+# RUN: printf 'abc' > %t.dir/odd.txt
+# RUN: llvm-ar rcSD --format=zos %t.odd.a %t.dir/odd.txt
+# RUN: %python -c "d=open(r'%t.odd.a','rb').read(); print(f'{d[-1]:02x}')" | FileCheck %s --check-prefix=PAD
+
+# PAD: 15
+
+# Long member names should written and read back correctly.
+# RUN: printf 'xyz' > %t.dir/very_long_member_name.txt
+# RUN: llvm-ar rcSD --format=zos %t.long.a %t.dir/very_long_member_name.txt
+# RUN: llvm-ar t %t.long.a | FileCheck %s --check-prefix=LONG-LIST
+# RUN: llvm-ar p %t.long.a very_long_member_name.txt | FileCheck %s --check-prefix=LONG-CONTENT
+
+# LONG-LIST: very_long_member_name.txt
+# LONG-CONTENT: xyz
diff --git a/llvm/tools/llvm-ar/llvm-ar.cpp b/llvm/tools/llvm-ar/llvm-ar.cpp
index bac99c548a178..38f929f63b39f 100644
--- a/llvm/tools/llvm-ar/llvm-ar.cpp
+++ b/llvm/tools/llvm-ar/llvm-ar.cpp
@@ -83,6 +83,7 @@ static void printArHelp(StringRef ToolName) {
=darwin - darwin
=bsd - bsd
=bigarchive - big archive (AIX OS)
+ =zos - zos archive (z/OS OS)
=coff - coff
--plugin=<string> - ignored for compatibility
-h --help - display this help and exit
@@ -195,7 +196,16 @@ static SmallVector<const char *, 256> PositionalArgs;
static bool MRI;
namespace {
-enum Format { Default, GNU, COFF, BSD, DARWIN, BIGARCHIVE, Unknown };
+enum Format {
+ Default,
+ GNU,
+ COFF,
+ BSD,
+ DARWIN,
+ BIGARCHIVE,
+ ZOSARCHIVE,
+ Unknown
+};
}
static Format FormatType = Default;
@@ -1079,6 +1089,11 @@ static void performWriteOperation(ArchiveOperation Operation,
fail("only the gnu format has a thin mode");
Kind = object::Archive::K_AIXBIG;
break;
+ case ZOSARCHIVE:
+ if (Thin)
+ fail("only the gnu format has a thin mode");
+ Kind = object::Archive::K_ZOS;
+ break;
case Unknown:
llvm_unreachable("");
}
@@ -1397,6 +1412,7 @@ static int ar_main(int argc, char **argv) {
.Case("bsd", BSD)
.Case("bigarchive", BIGARCHIVE)
.Case("coff", COFF)
+ .Case("zos", ZOSARCHIVE)
.Default(Unknown);
if (FormatType == Unknown)
fail(std::string("Invalid format ") + Match);
>From 37ad1267dc5ba2ed0925dfadffc8405fdfc39e4f Mon Sep 17 00:00:00 2001
From: Amy Kwan <amy.kwan1 at ibm.com>
Date: Fri, 12 Jun 2026 07:11:24 -0400
Subject: [PATCH 2/9] Address review comments.
---
llvm/lib/Object/ArchiveWriter.cpp | 14 +++++------
llvm/test/tools/llvm-ar/zos-write.test | 34 ++++++++++++++++++++------
2 files changed, 32 insertions(+), 16 deletions(-)
diff --git a/llvm/lib/Object/ArchiveWriter.cpp b/llvm/lib/Object/ArchiveWriter.cpp
index 8479340af26a3..3bfb1d06c7056 100644
--- a/llvm/lib/Object/ArchiveWriter.cpp
+++ b/llvm/lib/Object/ArchiveWriter.cpp
@@ -418,10 +418,11 @@ static uint64_t computeSymbolTableSize(object::Archive::Kind Kind,
uint32_t *Padding = nullptr) {
assert((OffsetSize == 4 || OffsetSize == 8) && "Unsupported OffsetSize");
uint64_t Size = OffsetSize; // Number of entries
- if (isBSDLike(Kind))
+ // Each symbol table entry consists of a member offset.
+ // For BSD, each entry also includes a string table offset.
+ // For z/OS, each entry also includes a flag field.
+ if (isBSDLike(Kind) || isZOSArchive(Kind))
Size += NumSyms * OffsetSize * 2; // Table
- else if (isZOSArchive(Kind))
- Size += NumSyms * OffsetSize * 2; // MemberOffset + symbol flags
else
Size += NumSyms * OffsetSize; // Table
@@ -649,7 +650,7 @@ static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
writeSymbolTableHeader(Out, Kind, Deterministic, Size, PrevMemberOffset,
NextMemberOffset);
// Padding size is not included in the Size field of the z/OS symbol table
- // header
+ // header.
if (isZOSArchive(Kind))
Size -= Pad;
@@ -682,11 +683,8 @@ static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
// byte count of the string table
printNBits(Out, Kind, StringTable.size());
if (isZOSArchive(Kind)) {
- std::string AStringTable;
- raw_string_ostream AOut(AStringTable);
- AOut << StringTable;
SmallString<256> EStringTable;
- ConverterEBCDIC::convertToEBCDIC(AStringTable, EStringTable);
+ ConverterEBCDIC::convertToEBCDIC(StringTable, EStringTable);
Out << EStringTable.str();
} else {
Out << StringTable;
diff --git a/llvm/test/tools/llvm-ar/zos-write.test b/llvm/test/tools/llvm-ar/zos-write.test
index 447e4562b83aa..d26b6d092b961 100644
--- a/llvm/test/tools/llvm-ar/zos-write.test
+++ b/llvm/test/tools/llvm-ar/zos-write.test
@@ -1,17 +1,17 @@
-# Test writing a z/OS archive.
+## Test writing a z/OS archive.
-# Create a z/OS archive from a plain text file, without a symbol table.
+## Create a z/OS archive from a plain text file, without a symbol table.
# RUN: rm -rf %t.dir && mkdir -p %t.dir
# RUN: printf 'abcd' > %t.dir/test.txt
# RUN: llvm-ar rcSD --format=zos %t.a %t.dir/test.txt
-# The archive should be written and read back successfully through llvm-ar.
+## The archive should be written and read back successfully through llvm-ar.
# RUN: llvm-ar t %t.a | FileCheck %s --check-prefix=LIST
# RUN: llvm-ar p %t.a test.txt | FileCheck %s --check-prefix=CONTENT
-# Check raw bytes:
-# - The first 8 bytes are the z/OS archive magic in EBCDIC.
-# - Bytes 66..67 are the first member header terminator (`\n in EBCDIC => 79 15)
+## Check raw bytes:
+## - The first 8 bytes are the z/OS archive magic in EBCDIC.
+## - Bytes 66..67 are the first member header terminator (`\n in EBCDIC => 79 15)
# RUN: %python -c "d=open(r'%t.a','rb').read(); print(d[:8].hex()); print(d[66:68].hex())" | FileCheck %s --check-prefix=BYTES
# LIST: test.txt
@@ -19,14 +19,14 @@
# BYTES: 5a4c819983886e15
# BYTES-NEXT: 7915
-# Odd-sized members should be padded with EBCDIC newline (0x15).
+## Odd-sized members should be padded with EBCDIC newline (0x15).
# RUN: printf 'abc' > %t.dir/odd.txt
# RUN: llvm-ar rcSD --format=zos %t.odd.a %t.dir/odd.txt
# RUN: %python -c "d=open(r'%t.odd.a','rb').read(); print(f'{d[-1]:02x}')" | FileCheck %s --check-prefix=PAD
# PAD: 15
-# Long member names should written and read back correctly.
+## Long member names should be written and read back correctly.
# RUN: printf 'xyz' > %t.dir/very_long_member_name.txt
# RUN: llvm-ar rcSD --format=zos %t.long.a %t.dir/very_long_member_name.txt
# RUN: llvm-ar t %t.long.a | FileCheck %s --check-prefix=LONG-LIST
@@ -34,3 +34,21 @@
# LONG-LIST: very_long_member_name.txt
# LONG-CONTENT: xyz
+
+## A z/OS archive with real symbols should write a readable symbol table.
+# RUN: printf 'define i32 @mytest() {\nentry:\n ret i32 87\n}\n' > %t.dir/mytest.ll
+# RUN: llvm-as %t.dir/mytest.ll -o %t.dir/mytest.bc
+# RUN: llvm-ar rcD --format=zos %t.sym.a %t.dir/mytest.bc
+# RUN: llvm-nm --print-armap %t.sym.a | FileCheck %s --check-prefix=SYMTAB
+
+# SYMTAB: Archive map
+# SYMTAB-NEXT: mytest in mytest.bc
+
+## A z/OS archive with no real symbols should still write the dummy blank symbol.
+# RUN: printf 'source_filename = "blank.ll"\n' > %t.dir/blank.ll
+# RUN: llvm-as %t.dir/blank.ll -o %t.dir/blank.bc
+# RUN: llvm-ar rcD --format=zos %t.blank-sym.a %t.dir/blank.bc
+# RUN: llvm-nm --print-armap %t.blank-sym.a | FileCheck %s --check-prefix=EMPTY-SYMTAB
+
+# EMPTY-SYMTAB: Archive map
+# EMPTY-SYMTAB-NEXT: {{^ in blank.bc$}}
>From d5918eaac127d742861c9943fa50efa637cdc45f Mon Sep 17 00:00:00 2001
From: Amy Kwan <amy.kwan1 at ibm.com>
Date: Sat, 20 Jun 2026 23:39:42 -0400
Subject: [PATCH 3/9] Address review comments.
---
llvm/lib/Object/ArchiveWriter.cpp | 2 +-
llvm/test/tools/llvm-ar/zos-write.test | 21 +++++++++++++++++----
2 files changed, 18 insertions(+), 5 deletions(-)
diff --git a/llvm/lib/Object/ArchiveWriter.cpp b/llvm/lib/Object/ArchiveWriter.cpp
index 3bfb1d06c7056..b16f4d8c341d3 100644
--- a/llvm/lib/Object/ArchiveWriter.cpp
+++ b/llvm/lib/Object/ArchiveWriter.cpp
@@ -420,7 +420,7 @@ static uint64_t computeSymbolTableSize(object::Archive::Kind Kind,
uint64_t Size = OffsetSize; // Number of entries
// Each symbol table entry consists of a member offset.
// For BSD, each entry also includes a string table offset.
- // For z/OS, each entry also includes a flag field.
+ // For z/OS, each entry instead also includes a flag field.
if (isBSDLike(Kind) || isZOSArchive(Kind))
Size += NumSyms * OffsetSize * 2; // Table
else
diff --git a/llvm/test/tools/llvm-ar/zos-write.test b/llvm/test/tools/llvm-ar/zos-write.test
index d26b6d092b961..571964ff566bb 100644
--- a/llvm/test/tools/llvm-ar/zos-write.test
+++ b/llvm/test/tools/llvm-ar/zos-write.test
@@ -1,7 +1,9 @@
## Test writing a z/OS archive.
-## Create a z/OS archive from a plain text file, without a symbol table.
# RUN: rm -rf %t.dir && mkdir -p %t.dir
+# RUN: split-file %s %t.dir
+
+## Create a z/OS archive from a plain text file with symbol table writing disabled.
# RUN: printf 'abcd' > %t.dir/test.txt
# RUN: llvm-ar rcSD --format=zos %t.a %t.dir/test.txt
@@ -11,7 +13,7 @@
## Check raw bytes:
## - The first 8 bytes are the z/OS archive magic in EBCDIC.
-## - Bytes 66..67 are the first member header terminator (`\n in EBCDIC => 79 15)
+## - Bytes 66..67 are the first member header terminator (`\n in EBCDIC => 79 15).
# RUN: %python -c "d=open(r'%t.a','rb').read(); print(d[:8].hex()); print(d[66:68].hex())" | FileCheck %s --check-prefix=BYTES
# LIST: test.txt
@@ -35,20 +37,31 @@
# LONG-LIST: very_long_member_name.txt
# LONG-CONTENT: xyz
+## Use LLVM bitcode here so the test remains target-independent. The archive
+## writer only needs a symbolic input in order to emit the archive symbol table.
## A z/OS archive with real symbols should write a readable symbol table.
-# RUN: printf 'define i32 @mytest() {\nentry:\n ret i32 87\n}\n' > %t.dir/mytest.ll
# RUN: llvm-as %t.dir/mytest.ll -o %t.dir/mytest.bc
# RUN: llvm-ar rcD --format=zos %t.sym.a %t.dir/mytest.bc
# RUN: llvm-nm --print-armap %t.sym.a | FileCheck %s --check-prefix=SYMTAB
# SYMTAB: Archive map
# SYMTAB-NEXT: mytest in mytest.bc
+# SYMTAB-NOT: {{^ in mytest.bc$}}
## A z/OS archive with no real symbols should still write the dummy blank symbol.
-# RUN: printf 'source_filename = "blank.ll"\n' > %t.dir/blank.ll
# RUN: llvm-as %t.dir/blank.ll -o %t.dir/blank.bc
# RUN: llvm-ar rcD --format=zos %t.blank-sym.a %t.dir/blank.bc
# RUN: llvm-nm --print-armap %t.blank-sym.a | FileCheck %s --check-prefix=EMPTY-SYMTAB
# EMPTY-SYMTAB: Archive map
# EMPTY-SYMTAB-NEXT: {{^ in blank.bc$}}
+# EMPTY-SYMTAB-NOT: {{^.+ in blank.bc$}}
+
+#--- mytest.ll
+define i32 @mytest() {
+entry:
+ ret i32 87
+}
+
+#--- blank.ll
+source_filename = "blank.ll"
>From 79efc8e8898eab8b5dc1313aaba8b10b57419812 Mon Sep 17 00:00:00 2001
From: Amy Kwan <amy.kwan1 at ibm.com>
Date: Thu, 25 Jun 2026 15:19:35 -0400
Subject: [PATCH 4/9] Remove regex from checks.
---
llvm/test/tools/llvm-ar/zos-write.test | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/llvm/test/tools/llvm-ar/zos-write.test b/llvm/test/tools/llvm-ar/zos-write.test
index 571964ff566bb..d09d1214a328c 100644
--- a/llvm/test/tools/llvm-ar/zos-write.test
+++ b/llvm/test/tools/llvm-ar/zos-write.test
@@ -46,7 +46,7 @@
# SYMTAB: Archive map
# SYMTAB-NEXT: mytest in mytest.bc
-# SYMTAB-NOT: {{^ in mytest.bc$}}
+# SYMTAB-NOT: in mytest.bc
## A z/OS archive with no real symbols should still write the dummy blank symbol.
# RUN: llvm-as %t.dir/blank.ll -o %t.dir/blank.bc
@@ -54,8 +54,8 @@
# RUN: llvm-nm --print-armap %t.blank-sym.a | FileCheck %s --check-prefix=EMPTY-SYMTAB
# EMPTY-SYMTAB: Archive map
-# EMPTY-SYMTAB-NEXT: {{^ in blank.bc$}}
-# EMPTY-SYMTAB-NOT: {{^.+ in blank.bc$}}
+# EMPTY-SYMTAB-NEXT: in blank.bc
+# EMPTY-SYMTAB-NOT: in blank.bc
#--- mytest.ll
define i32 @mytest() {
>From daa1514cb7873fa86b004f6f62627ce18f1b8aa7 Mon Sep 17 00:00:00 2001
From: Amy Kwan <amy.kwan1 at ibm.com>
Date: Fri, 10 Jul 2026 16:46:03 -0400
Subject: [PATCH 5/9] Fix printZOSMemberHeader and handle ConverterEBCDIC error
codes.
---
llvm/lib/Object/ArchiveWriter.cpp | 23 ++++++++++++++++-------
1 file changed, 16 insertions(+), 7 deletions(-)
diff --git a/llvm/lib/Object/ArchiveWriter.cpp b/llvm/lib/Object/ArchiveWriter.cpp
index b16f4d8c341d3..283f2bf0fbd07 100644
--- a/llvm/lib/Object/ArchiveWriter.cpp
+++ b/llvm/lib/Object/ArchiveWriter.cpp
@@ -261,7 +261,7 @@ printBSDMemberHeader(raw_ostream &Out, uint64_t Pos, StringRef Name,
}
static void
-printZOSMemberHeader(raw_ostream &Out, uint64_t Pos, StringRef Name,
+printZOSMemberHeader(raw_ostream &Out, StringRef Name,
const sys::TimePoint<std::chrono::seconds> &ModTime,
unsigned UID, unsigned GID, unsigned Perms,
uint64_t Size) {
@@ -271,10 +271,17 @@ printZOSMemberHeader(raw_ostream &Out, uint64_t Pos, StringRef Name,
printWithSpacePadding(AOut, Twine(Name), 16);
printRestOfMemberHeader(AOut, ModTime, UID, GID, Perms, Size);
} else {
- printBSDMemberHeader(AOut, Pos, Name, ModTime, UID, GID, Perms, Size);
+ // z/OS ar stores the exact name length inline with no extra alignment
+ // padding, unlike the BSD format which pads to an 8-byte boundary.
+ printWithSpacePadding(AOut, Twine("#1/") + Twine(Name.size()), 16);
+ printRestOfMemberHeader(AOut, ModTime, UID, GID, Perms,
+ Name.size() + Size);
+ AOut << Name;
}
SmallString<256> EHeader;
- ConverterEBCDIC::convertToEBCDIC(AHeader, EHeader);
+ if (std::error_code EC = ConverterEBCDIC::convertToEBCDIC(AHeader, EHeader))
+ report_fatal_error(Twine("failed to convert z/OS member header to EBCDIC: ") +
+ EC.message());
Out << EHeader.str();
}
@@ -332,7 +339,7 @@ printMemberHeader(raw_ostream &Out, uint64_t Pos, raw_ostream &StringTable,
return printBSDMemberHeader(Out, Pos, MemberName, ModTime, M.UID, M.GID,
M.Perms, Size);
if (isZOSArchive(Kind))
- return printZOSMemberHeader(Out, Pos, MemberName, ModTime, M.UID, M.GID,
+ return printZOSMemberHeader(Out, MemberName, ModTime, M.UID, M.GID,
M.Perms, Size);
if (!useStringTable(Thin, MemberName))
return printGNUSmallMemberHeader(Out, MemberName, ModTime, M.UID, M.GID,
@@ -487,8 +494,7 @@ static void writeSymbolTableHeader(raw_ostream &Out, object::Archive::Kind Kind,
PrevMemberOffset, NextMemberOffset);
} else if (isZOSArchive(Kind)) {
const char *Name = "__.SYMDEF";
- printZOSMemberHeader(Out, Out.tell(), Name, now(Deterministic), 0, 0, 0,
- Size);
+ printZOSMemberHeader(Out, Name, now(Deterministic), 0, 0, 0, Size);
} else {
const char *Name = is64BitKind(Kind) ? "/SYM64" : "";
printGNUSmallMemberHeader(Out, Name, now(Deterministic), 0, 0, 0, Size);
@@ -684,7 +690,10 @@ static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
printNBits(Out, Kind, StringTable.size());
if (isZOSArchive(Kind)) {
SmallString<256> EStringTable;
- ConverterEBCDIC::convertToEBCDIC(StringTable, EStringTable);
+ if (std::error_code EC =
+ ConverterEBCDIC::convertToEBCDIC(StringTable, EStringTable))
+ report_fatal_error(Twine("failed to convert z/OS symbol table to EBCDIC: ") +
+ EC.message());
Out << EStringTable.str();
} else {
Out << StringTable;
>From 4e954480fd409129e6d7d89fd65eb819edc480f0 Mon Sep 17 00:00:00 2001
From: Amy Kwan <amy.kwan1 at ibm.com>
Date: Fri, 10 Jul 2026 16:49:57 -0400
Subject: [PATCH 6/9] Subtract padding prior to writing symbol header, add
FIXME for symbol flags.
---
llvm/lib/Object/ArchiveWriter.cpp | 12 ++++++++----
1 file changed, 8 insertions(+), 4 deletions(-)
diff --git a/llvm/lib/Object/ArchiveWriter.cpp b/llvm/lib/Object/ArchiveWriter.cpp
index 283f2bf0fbd07..efa5a2e311f49 100644
--- a/llvm/lib/Object/ArchiveWriter.cpp
+++ b/llvm/lib/Object/ArchiveWriter.cpp
@@ -653,12 +653,15 @@ static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
uint32_t Pad;
uint64_t Size = computeSymbolTableSize(Kind, NumSyms, OffsetSize,
StringTable.size(), &Pad);
- writeSymbolTableHeader(Out, Kind, Deterministic, Size, PrevMemberOffset,
- NextMemberOffset);
+
// Padding size is not included in the Size field of the z/OS symbol table
// header.
+ int64_t HeaderSize = Size;
if (isZOSArchive(Kind))
- Size -= Pad;
+ HeaderSize -= Pad;
+
+ writeSymbolTableHeader(Out, Kind, Deterministic, HeaderSize, PrevMemberOffset,
+ NextMemberOffset);
if (isBSDLike(Kind))
printNBits(Out, Kind, NumSyms * 2 * OffsetSize);
@@ -679,8 +682,9 @@ static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
if (isBSDLike(Kind))
printNBits(Out, Kind, StringOffset);
printNBits(Out, Kind, Pos); // member offset
+ // FIXME: Properly handle symbol attributes for z/OS archives.
if (isZOSArchive(Kind))
- printNBits(Out, Kind, Pos); // symbol flags
+ printNBits(Out, Kind, 0); // symbol flags
}
Pos += M.Header.size() + M.Data.size() + M.Padding.size();
}
>From 33951b936c6450a49bdb51e14d99c0deb61d49e2 Mon Sep 17 00:00:00 2001
From: Amy Kwan <amy.kwan1 at ibm.com>
Date: Sun, 12 Jul 2026 23:54:00 -0400
Subject: [PATCH 7/9] Update test case for archive clean up
---
llvm/test/tools/llvm-ar/zos-write.test | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/test/tools/llvm-ar/zos-write.test b/llvm/test/tools/llvm-ar/zos-write.test
index d09d1214a328c..a3209a85db296 100644
--- a/llvm/test/tools/llvm-ar/zos-write.test
+++ b/llvm/test/tools/llvm-ar/zos-write.test
@@ -1,6 +1,6 @@
## Test writing a z/OS archive.
-# RUN: rm -rf %t.dir && mkdir -p %t.dir
+# RUN: rm -rf %t.dir %t.a %t.odd.a %t.long.a %t.sym.a %t.blank-sym.a && mkdir -p %t.dir
# RUN: split-file %s %t.dir
## Create a z/OS archive from a plain text file with symbol table writing disabled.
>From 4ffc1f51a0f3f1b379d23cd3f45ef3ad526f5e04 Mon Sep 17 00:00:00 2001
From: Amy Kwan <amy.kwan1 at ibm.com>
Date: Mon, 13 Jul 2026 13:34:03 -0400
Subject: [PATCH 8/9] Update long name handling.
---
llvm/lib/Object/ArchiveWriter.cpp | 15 ++++++++++-----
llvm/test/tools/llvm-ar/zos-write.test | 7 ++++++-
2 files changed, 16 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/Object/ArchiveWriter.cpp b/llvm/lib/Object/ArchiveWriter.cpp
index efa5a2e311f49..9a8999dcc9250 100644
--- a/llvm/lib/Object/ArchiveWriter.cpp
+++ b/llvm/lib/Object/ArchiveWriter.cpp
@@ -432,7 +432,6 @@ static uint64_t computeSymbolTableSize(object::Archive::Kind Kind,
Size += NumSyms * OffsetSize * 2; // Table
else
Size += NumSyms * OffsetSize; // Table
-
if (isBSDLike(Kind))
Size += OffsetSize; // byte count
Size += StringTableSize;
@@ -924,6 +923,7 @@ computeMemberData(raw_ostream &StringTable, raw_ostream &SymNames,
if (!SymFileOrErr)
return createFileError(M.MemberName, SymFileOrErr.takeError());
D.SymFile = std::move(*SymFileOrErr);
+
if (SymMap && D.SymFile.get()) {
auto COFFObj = dyn_cast<COFFObjectFile>(D.SymFile.get());
std::optional<MemoryBufferRef> HybridView;
@@ -1007,13 +1007,18 @@ computeMemberData(raw_ostream &StringTable, raw_ostream &SymNames,
// is happy with archives that we generate.
unsigned MemberPadding =
isDarwin(Kind) ? offsetToAlignment(D.Data.size(), Align(8)) : 0;
- unsigned TailPadding =
- offsetToAlignment(D.Data.size() + MemberPadding, Align(2));
- D.Padding = StringRef(isZOSArchive(Kind) ? ZOSPaddingData : PaddingData,
- MemberPadding + TailPadding);
StringRef MemberName = D.HybridName.size() ? D.HybridName : M->MemberName;
+ // z/OS stores long member names inline using their exact byte length.
+ // Include the inline name when computing alignment.
+ uint64_t PaddingBase = D.Data.size() + MemberPadding;
+ if (isZOSArchive(Kind) && MemberName.size() > 16)
+ PaddingBase += MemberName.size();
+ unsigned TailPadding = offsetToAlignment(PaddingBase, Align(2));
+ D.Padding = StringRef(isZOSArchive(Kind) ? ZOSPaddingData : PaddingData,
+ MemberPadding + TailPadding);
+
sys::TimePoint<std::chrono::seconds> ModTime;
if (UniqueTimestamps)
// Increment timestamp for each file of a given name.
diff --git a/llvm/test/tools/llvm-ar/zos-write.test b/llvm/test/tools/llvm-ar/zos-write.test
index a3209a85db296..0e390183f8d00 100644
--- a/llvm/test/tools/llvm-ar/zos-write.test
+++ b/llvm/test/tools/llvm-ar/zos-write.test
@@ -28,14 +28,19 @@
# PAD: 15
-## Long member names should be written and read back correctly.
+## Long member names should use their exact byte length and be read back
+## correctly. The 25-byte name plus the 3-byte contents is already even-sized,
+## so the archive should not contain a trailing padding byte.
# RUN: printf 'xyz' > %t.dir/very_long_member_name.txt
# RUN: llvm-ar rcSD --format=zos %t.long.a %t.dir/very_long_member_name.txt
# RUN: llvm-ar t %t.long.a | FileCheck %s --check-prefix=LONG-LIST
# RUN: llvm-ar p %t.long.a very_long_member_name.txt | FileCheck %s --check-prefix=LONG-CONTENT
+# RUN: %python -c "d=open(r'%t.long.a','rb').read(); print(d[8:24].decode('cp037').strip()); print(len(d))" | FileCheck %s --check-prefix=LONG-BYTES
# LONG-LIST: very_long_member_name.txt
# LONG-CONTENT: xyz
+# LONG-BYTES: #1/25
+# LONG-BYTES-NEXT: 96
## Use LLVM bitcode here so the test remains target-independent. The archive
## writer only needs a symbolic input in order to emit the archive symbol table.
>From 27c245b5c993176f854d8b2db2295f9b857f8a51 Mon Sep 17 00:00:00 2001
From: Amy Kwan <amy.kwan1 at ibm.com>
Date: Mon, 13 Jul 2026 21:05:48 +0000
Subject: [PATCH 9/9] Apply clang format.
---
llvm/lib/Object/ArchiveWriter.cpp | 17 +++++++++--------
1 file changed, 9 insertions(+), 8 deletions(-)
diff --git a/llvm/lib/Object/ArchiveWriter.cpp b/llvm/lib/Object/ArchiveWriter.cpp
index 9a8999dcc9250..9af8e0efe1872 100644
--- a/llvm/lib/Object/ArchiveWriter.cpp
+++ b/llvm/lib/Object/ArchiveWriter.cpp
@@ -274,14 +274,14 @@ printZOSMemberHeader(raw_ostream &Out, StringRef Name,
// z/OS ar stores the exact name length inline with no extra alignment
// padding, unlike the BSD format which pads to an 8-byte boundary.
printWithSpacePadding(AOut, Twine("#1/") + Twine(Name.size()), 16);
- printRestOfMemberHeader(AOut, ModTime, UID, GID, Perms,
- Name.size() + Size);
+ printRestOfMemberHeader(AOut, ModTime, UID, GID, Perms, Name.size() + Size);
AOut << Name;
}
SmallString<256> EHeader;
if (std::error_code EC = ConverterEBCDIC::convertToEBCDIC(AHeader, EHeader))
- report_fatal_error(Twine("failed to convert z/OS member header to EBCDIC: ") +
- EC.message());
+ report_fatal_error(
+ Twine("failed to convert z/OS member header to EBCDIC: ") +
+ EC.message());
Out << EHeader.str();
}
@@ -339,8 +339,8 @@ printMemberHeader(raw_ostream &Out, uint64_t Pos, raw_ostream &StringTable,
return printBSDMemberHeader(Out, Pos, MemberName, ModTime, M.UID, M.GID,
M.Perms, Size);
if (isZOSArchive(Kind))
- return printZOSMemberHeader(Out, MemberName, ModTime, M.UID, M.GID,
- M.Perms, Size);
+ return printZOSMemberHeader(Out, MemberName, ModTime, M.UID, M.GID, M.Perms,
+ Size);
if (!useStringTable(Thin, MemberName))
return printGNUSmallMemberHeader(Out, MemberName, ModTime, M.UID, M.GID,
M.Perms, Size);
@@ -695,8 +695,9 @@ static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
SmallString<256> EStringTable;
if (std::error_code EC =
ConverterEBCDIC::convertToEBCDIC(StringTable, EStringTable))
- report_fatal_error(Twine("failed to convert z/OS symbol table to EBCDIC: ") +
- EC.message());
+ report_fatal_error(
+ Twine("failed to convert z/OS symbol table to EBCDIC: ") +
+ EC.message());
Out << EStringTable.str();
} else {
Out << StringTable;
More information about the llvm-commits
mailing list