[llvm] [BOLT] support mold linker generated PLT in disassembling (PR #115256)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Nov 6 19:37:10 PST 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-bolt
Author: Patrick Zhang (patrickphzhang)
<details>
<summary>Changes</summary>
The mold linker (https://github.com/rui314/mold/blob/v2.34.1/src/arch-x86-64.cc#L50) generates a unique format of the PLT for the X86 CET feature. The first entry of the mold-style PLT is 32 bits long, while the remaining entries are 16 bits long.
Take the existing test case --- 'plt-mold.test' --- as the example, the PLT header is:
```
00000000000013a0 <_PROCEDURE_LINKAGE_TABLE_>:
13a0: f3 0f 1e fa endbr64
13a4: 41 53 pushq %r11
13a6: ff 35 ac 21 00 00 pushq 0x21ac(%rip)
13ac: ff 25 ae 21 00 00 jmpq *0x21ae(%rip)
13b2: cc int3
13b3: cc int3
... ...
```
In the current parsing process, BOLT will successively parse the first 4 instructions in the PLT header and roll back (by mistake) to the 16th byte, which is in the middle of the 4th instruction. BOLT works fine if the latter two bytes of the 4th instruction can be disassembled. However, if those two bytes can not be disassembled, the BOLT will report
> BOLT-ERROR: unable to disassemble instruction in PLT section .plt at offset 0x10.
For the existing test case -- 'plt-mold.test' -- the last two bytes of the 4th instruction are '_0x00 0x00_', which can be disassembled into '_add [rax], al_', so it does not hit the error case.
To this end, this commit parses the first entry with a special offset limit setting under a new command option. Besides, the new test case -- 'plt-mold-header.test' -- is added to verify the above case.
---
Full diff: https://github.com/llvm/llvm-project/pull/115256.diff
5 Files Affected:
- (modified) bolt/include/bolt/Utils/CommandLineOpts.h (+1)
- (modified) bolt/lib/Rewrite/RewriteInstance.cpp (+29-1)
- (modified) bolt/lib/Utils/CommandLineOpts.cpp (+6)
- (added) bolt/test/X86/Inputs/plt-mold-header.yaml (+399)
- (added) bolt/test/X86/plt-mold-header.test (+7)
``````````diff
diff --git a/bolt/include/bolt/Utils/CommandLineOpts.h b/bolt/include/bolt/Utils/CommandLineOpts.h
index 04bf7db5de9527..3b0c0db1bd089e 100644
--- a/bolt/include/bolt/Utils/CommandLineOpts.h
+++ b/bolt/include/bolt/Utils/CommandLineOpts.h
@@ -34,6 +34,7 @@ extern llvm::cl::opt<bool> AggregateOnly;
extern llvm::cl::opt<unsigned> BucketsPerLine;
extern llvm::cl::opt<bool> DiffOnly;
extern llvm::cl::opt<bool> EnableBAT;
+extern llvm::cl::opt<bool> UseMold;
extern llvm::cl::opt<bool> EqualizeBBCounts;
extern llvm::cl::opt<bool> RemoveSymtab;
extern llvm::cl::opt<unsigned> ExecutionCountThreshold;
diff --git a/bolt/lib/Rewrite/RewriteInstance.cpp b/bolt/lib/Rewrite/RewriteInstance.cpp
index 32ec7abe8b666a..a7118be5dc263a 100644
--- a/bolt/lib/Rewrite/RewriteInstance.cpp
+++ b/bolt/lib/Rewrite/RewriteInstance.cpp
@@ -1672,7 +1672,35 @@ void RewriteInstance::disassemblePLTSectionX86(BinarySection &Section,
const uint64_t SectionAddress = Section.getAddress();
const uint64_t SectionSize = Section.getSize();
- for (uint64_t EntryOffset = 0; EntryOffset + EntrySize <= SectionSize;
+ uint64_t EntryStartOffset = 0;
+ if (opts::UseMold) {
+ // The mold linker (https://github.com/rui314/mold/blob/v2.34.1/src/arch-x86-64.cc#L50)
+ // generates a unique format for the PLT.
+ // The first entry of the mold-style PLT is 32 bytes long, while the remaining entries
+ // are 16 bytes long. We need to parse the first entry with a special offset limit setting.
+ uint64_t HeaderSize = 32;
+ outs() << "BOLT-INFO: parsing PLT header for mold\n";
+ MCInst Instruction;
+ uint64_t InstrSize, InstrOffset = EntryStartOffset;
+ while (InstrOffset < HeaderSize) {
+ disassemblePLTInstruction(Section, InstrOffset, Instruction, InstrSize);
+ if (BC->MIB->isIndirectBranch(Instruction))
+ break;
+ InstrOffset += InstrSize;
+ }
+ uint64_t TargetAddress;
+ if (!BC->MIB->evaluateMemOperandTarget(Instruction, TargetAddress,
+ SectionAddress + InstrOffset,
+ InstrSize)) {
+ errs() << "BOLT-ERROR: error evaluating PLT instruction for the mold header at offset 0x"
+ << Twine::utohexstr(SectionAddress + InstrOffset) << '\n';
+ exit(1);
+ }
+ createPLTBinaryFunction(TargetAddress, SectionAddress, HeaderSize);
+ EntryStartOffset += HeaderSize;
+ }
+
+ for (uint64_t EntryOffset = EntryStartOffset; EntryOffset + EntrySize <= SectionSize;
EntryOffset += EntrySize) {
MCInst Instruction;
uint64_t InstrSize, InstrOffset = EntryOffset;
diff --git a/bolt/lib/Utils/CommandLineOpts.cpp b/bolt/lib/Utils/CommandLineOpts.cpp
index de82420a167131..356e530c9ca361 100644
--- a/bolt/lib/Utils/CommandLineOpts.cpp
+++ b/bolt/lib/Utils/CommandLineOpts.cpp
@@ -72,6 +72,12 @@ EnableBAT("enable-bat",
cl::ZeroOrMore,
cl::cat(BoltCategory));
+cl::opt<bool> UseMold("use-mold",
+ cl::desc("the binary is generated by the mold linker"),
+ cl::init(false),
+ cl::ZeroOrMore,
+ cl::cat(BoltCategory));
+
cl::opt<bool> EqualizeBBCounts(
"equalize-bb-counts",
cl::desc("use same count for BBs that should have equivalent count (used "
diff --git a/bolt/test/X86/Inputs/plt-mold-header.yaml b/bolt/test/X86/Inputs/plt-mold-header.yaml
new file mode 100644
index 00000000000000..be6eabeccbba8f
--- /dev/null
+++ b/bolt/test/X86/Inputs/plt-mold-header.yaml
@@ -0,0 +1,399 @@
+--- !ELF
+FileHeader:
+ Class: ELFCLASS64
+ Data: ELFDATA2LSB
+ Type: ET_DYN
+ Machine: EM_X86_64
+ Entry: 0x13D0
+ProgramHeaders:
+ - Type: PT_PHDR
+ Flags: [ PF_R ]
+ VAddr: 0x40
+ Align: 0x8
+ - Type: PT_INTERP
+ Flags: [ PF_R ]
+ FirstSec: .interp
+ LastSec: .interp
+ VAddr: 0x270
+ - Type: PT_LOAD
+ Flags: [ PF_R ]
+ FirstSec: .interp
+ LastSec: .rodata.str
+ Align: 0x1000
+ - Type: PT_LOAD
+ Flags: [ PF_X, PF_R ]
+ FirstSec: .plt
+ LastSec: .text
+ VAddr: 0x13A0
+ Align: 0x1000
+ - Type: PT_LOAD
+ Flags: [ PF_W, PF_R ]
+ FirstSec: .dynamic
+ LastSec: .relro_padding
+ VAddr: 0x23F8
+ Align: 0x1000
+ - Type: PT_LOAD
+ Flags: [ PF_W, PF_R ]
+ FirstSec: .got.plt
+ LastSec: .got.plt
+ VAddr: 0x3550
+ Align: 0x1000
+ - Type: PT_DYNAMIC
+ Flags: [ PF_W, PF_R ]
+ FirstSec: .dynamic
+ LastSec: .dynamic
+ VAddr: 0x23F8
+ Align: 0x8
+ - Type: PT_GNU_EH_FRAME
+ Flags: [ PF_R ]
+ FirstSec: .eh_frame_hdr
+ LastSec: .eh_frame_hdr
+ VAddr: 0x37C
+ Align: 0x4
+ - Type: PT_GNU_STACK
+ Flags: [ PF_W, PF_R ]
+ - Type: PT_GNU_RELRO
+ Flags: [ PF_R ]
+ FirstSec: .dynamic
+ LastSec: .relro_padding
+ VAddr: 0x23F8
+Sections:
+ - Name: .interp
+ Type: SHT_PROGBITS
+ Flags: [ SHF_ALLOC ]
+ Address: 0x270
+ AddressAlign: 0x1
+ Content: 2F6C696236342F6C642D6C696E75782D7838362D36342E736F2E3200
+ - Name: .gnu.hash
+ Type: SHT_GNU_HASH
+ Flags: [ SHF_ALLOC ]
+ Address: 0x290
+ Link: .dynsym
+ AddressAlign: 0x8
+ Header:
+ SymNdx: 0x2
+ Shift2: 0x1A
+ BloomFilter: [ 0x0 ]
+ HashBuckets: [ 0x0 ]
+ HashValues: [ ]
+ - Name: .dynsym
+ Type: SHT_DYNSYM
+ Flags: [ SHF_ALLOC ]
+ Address: 0x2B0
+ Link: .dynstr
+ AddressAlign: 0x8
+ - Name: .dynstr
+ Type: SHT_STRTAB
+ Flags: [ SHF_ALLOC ]
+ Address: 0x2E0
+ AddressAlign: 0x1
+ - Name: .gnu.version
+ Type: SHT_GNU_versym
+ Flags: [ SHF_ALLOC ]
+ Address: 0x2FE
+ Link: .dynsym
+ AddressAlign: 0x2
+ Entries: [ 0, 2 ]
+ - Name: .gnu.version_r
+ Type: SHT_GNU_verneed
+ Flags: [ SHF_ALLOC ]
+ Address: 0x308
+ Link: .dynstr
+ AddressAlign: 0x8
+ Dependencies:
+ - Version: 1
+ File: libc.so.6
+ Entries:
+ - Name: GLIBC_2.2.5
+ Hash: 157882997
+ Flags: 0
+ Other: 2
+ - Name: .rela.plt
+ Type: SHT_RELA
+ Flags: [ SHF_ALLOC, SHF_INFO_LINK ]
+ Address: 0x328
+ Link: .dynsym
+ AddressAlign: 0x8
+ Info: .got.plt
+ Relocations:
+ - Offset: 0x3568
+ Symbol: printf
+ Type: R_X86_64_JUMP_SLOT
+ - Name: .eh_frame
+ Type: SHT_PROGBITS
+ Flags: [ SHF_ALLOC ]
+ Address: 0x340
+ AddressAlign: 0x8
+ Content: 1400000000000000017A5200017810011B0C0708900100001C0000001C000000701000002500000000410E108602430D06600C070800000000000000
+ - Name: .eh_frame_hdr
+ Type: SHT_PROGBITS
+ Flags: [ SHF_ALLOC ]
+ Address: 0x37C
+ AddressAlign: 0x4
+ Content: 011B033BC0FFFFFF0100000054100000DCFFFFFF
+ - Name: .rodata.str
+ Type: SHT_PROGBITS
+ Flags: [ SHF_ALLOC ]
+ Address: 0x390
+ AddressAlign: 0x1
+ Content: 48656C6C6F20776F726C64210A00
+ - Name: .plt
+ Type: SHT_PROGBITS
+ Flags: [ SHF_ALLOC, SHF_EXECINSTR ]
+ Address: 0x13A0
+ AddressAlign: 0x10
+ Content: F30F1EFA4153FF35CC3C1602FF25CE3C1602CCCCCCCCCCCCCCCCCCCCCCCCCCCCF30F1EFA41BB00000000FF2598210000
+ - Name: .text
+ Type: SHT_PROGBITS
+ Flags: [ SHF_ALLOC, SHF_EXECINSTR ]
+ Address: 0x13D0
+ AddressAlign: 0x10
+ Content: 554889E54883EC10C745FC00000000488D3DAAEFFFFFB000E8D3FFFFFF31C04883C4105DC3
+ - Name: .dynamic
+ Type: SHT_DYNAMIC
+ Flags: [ SHF_WRITE, SHF_ALLOC ]
+ Address: 0x23F8
+ Link: .dynstr
+ AddressAlign: 0x8
+ Entries:
+ - Tag: DT_NEEDED
+ Value: 0x1
+ - Tag: DT_JMPREL
+ Value: 0x328
+ - Tag: DT_PLTRELSZ
+ Value: 0x18
+ - Tag: DT_PLTREL
+ Value: 0x7
+ - Tag: DT_PLTGOT
+ Value: 0x3550
+ - Tag: DT_SYMTAB
+ Value: 0x2B0
+ - Tag: DT_SYMENT
+ Value: 0x18
+ - Tag: DT_STRTAB
+ Value: 0x2E0
+ - Tag: DT_STRSZ
+ Value: 0x1E
+ - Tag: DT_VERSYM
+ Value: 0x2FE
+ - Tag: DT_VERNEED
+ Value: 0x308
+ - Tag: DT_VERNEEDNUM
+ Value: 0x1
+ - Tag: DT_GNU_HASH
+ Value: 0x290
+ - Tag: DT_FLAGS_1
+ Value: 0x8000000
+ - Tag: DT_DEBUG
+ Value: 0x0
+ - Tag: DT_NULL
+ Value: 0x0
+ - Tag: DT_NULL
+ Value: 0x0
+ - Tag: DT_NULL
+ Value: 0x0
+ - Tag: DT_NULL
+ Value: 0x0
+ - Tag: DT_NULL
+ Value: 0x0
+ - Tag: DT_NULL
+ Value: 0x0
+ - Name: .got
+ Type: SHT_PROGBITS
+ Flags: [ SHF_WRITE, SHF_ALLOC ]
+ Address: 0x2548
+ AddressAlign: 0x8
+ Content: '0000000000000000'
+ - Name: .relro_padding
+ Type: SHT_NOBITS
+ Flags: [ SHF_WRITE, SHF_ALLOC ]
+ Address: 0x2550
+ AddressAlign: 0x1
+ Size: 0xAB0
+ - Name: .got.plt
+ Type: SHT_PROGBITS
+ Flags: [ SHF_WRITE, SHF_ALLOC ]
+ Address: 0x3550
+ AddressAlign: 0x8
+ Content: F82300000000000000000000000000000000000000000000A013000000000000
+ - Name: .rela.text
+ Type: SHT_RELA
+ Flags: [ SHF_INFO_LINK ]
+ Link: .symtab
+ AddressAlign: 0x8
+ Info: .text
+ Relocations:
+ - Offset: 0x13E2
+ Symbol: .L.str
+ Type: R_X86_64_PC32
+ Addend: -4
+ - Offset: 0x13E9
+ Symbol: printf
+ Type: R_X86_64_PLT32
+ Addend: -4
+ - Type: SectionHeaderTable
+ Sections:
+ - Name: .interp
+ - Name: .gnu.hash
+ - Name: .dynsym
+ - Name: .dynstr
+ - Name: .gnu.version
+ - Name: .gnu.version_r
+ - Name: .rela.plt
+ - Name: .eh_frame
+ - Name: .eh_frame_hdr
+ - Name: .rodata.str
+ - Name: .plt
+ - Name: .text
+ - Name: .rela.text
+ - Name: .dynamic
+ - Name: .got
+ - Name: .relro_padding
+ - Name: .got.plt
+ - Name: .symtab
+ - Name: .strtab
+ - Name: .shstrtab
+Symbols:
+ - Name: .interp
+ Type: STT_SECTION
+ Section: .interp
+ Value: 0x270
+ - Name: .gnu.hash
+ Type: STT_SECTION
+ Section: .gnu.hash
+ Value: 0x290
+ - Name: .dynsym
+ Type: STT_SECTION
+ Section: .dynsym
+ Value: 0x2B0
+ - Name: .dynstr
+ Type: STT_SECTION
+ Section: .dynstr
+ Value: 0x2E0
+ - Name: .gnu.version
+ Type: STT_SECTION
+ Section: .gnu.version
+ Value: 0x2FE
+ - Name: .gnu.version_r
+ Type: STT_SECTION
+ Section: .gnu.version_r
+ Value: 0x308
+ - Name: .rela.plt
+ Type: STT_SECTION
+ Section: .rela.plt
+ Value: 0x328
+ - Name: .eh_frame
+ Type: STT_SECTION
+ Section: .eh_frame
+ Value: 0x340
+ - Name: .eh_frame_hdr
+ Type: STT_SECTION
+ Section: .eh_frame_hdr
+ Value: 0x37C
+ - Name: .rodata.str
+ Type: STT_SECTION
+ Section: .rodata.str
+ Value: 0x390
+ - Name: .plt
+ Type: STT_SECTION
+ Section: .plt
+ Value: 0x13A0
+ - Name: .text
+ Type: STT_SECTION
+ Section: .text
+ Value: 0x13D0
+ - Name: .dynamic
+ Type: STT_SECTION
+ Section: .dynamic
+ Value: 0x23F8
+ - Name: .got
+ Type: STT_SECTION
+ Section: .got
+ Value: 0x2548
+ - Name: .relro_padding
+ Type: STT_SECTION
+ Section: .relro_padding
+ Value: 0x2550
+ - Name: .got.plt
+ Type: STT_SECTION
+ Section: .got.plt
+ Value: 0x3550
+ - Name: 'printf$plt'
+ Type: STT_FUNC
+ Section: .plt
+ Value: 0x13C0
+ - Name: hello.c
+ Type: STT_FILE
+ Index: SHN_ABS
+ - Name: .L.str
+ Type: STT_OBJECT
+ Section: .rodata.str
+ Value: 0x390
+ - Name: main
+ Type: STT_FUNC
+ Section: .text
+ Value: 0x13D0
+ Size: 0x25
+ - Name: __ehdr_start
+ Section: .interp
+ - Name: __init_array_start
+ Index: SHN_ABS
+ - Name: __init_array_end
+ Index: SHN_ABS
+ - Name: __fini_array_start
+ Index: SHN_ABS
+ - Name: __fini_array_end
+ Index: SHN_ABS
+ - Name: __preinit_array_start
+ Index: SHN_ABS
+ - Name: __preinit_array_end
+ Index: SHN_ABS
+ - Name: _DYNAMIC
+ Section: .dynamic
+ Value: 0x23F8
+ - Name: _GLOBAL_OFFSET_TABLE_
+ Section: .got.plt
+ Value: 0x3550
+ - Name: _PROCEDURE_LINKAGE_TABLE_
+ Section: .plt
+ Value: 0x13A0
+ - Name: __bss_start
+ Index: SHN_ABS
+ - Name: _end
+ Section: .got.plt
+ Value: 0x3570
+ - Name: _etext
+ Section: .text
+ Value: 0x13F5
+ - Name: _edata
+ Section: .got.plt
+ Value: 0x3570
+ - Name: __executable_start
+ Section: .interp
+ - Name: __rela_iplt_start
+ Index: SHN_ABS
+ - Name: __rela_iplt_end
+ Index: SHN_ABS
+ - Name: __GNU_EH_FRAME_HDR
+ Section: .eh_frame_hdr
+ Value: 0x37C
+ - Name: end
+ Section: .got.plt
+ Value: 0x3570
+ - Name: etext
+ Section: .text
+ Value: 0x13F5
+ - Name: edata
+ Section: .got.plt
+ Value: 0x3570
+ - Name: __dso_handle
+ Section: .interp
+ - Name: _TLS_MODULE_BASE_
+ Section: .interp
+ - Name: printf
+ Binding: STB_GLOBAL
+DynamicSymbols:
+ - Name: printf
+ Type: STT_FUNC
+ Binding: STB_GLOBAL
+...
diff --git a/bolt/test/X86/plt-mold-header.test b/bolt/test/X86/plt-mold-header.test
new file mode 100644
index 00000000000000..8cbbed8711cbce
--- /dev/null
+++ b/bolt/test/X86/plt-mold-header.test
@@ -0,0 +1,7 @@
+# RUN: yaml2obj %p/Inputs/plt-mold-header.yaml &> %t.exe
+# RUN: llvm-bolt -use-mold %t.exe --print-cfg --print-only=main.* -o %t.out | FileCheck %s
+
+## Check that llvm-bolt correctly parses PLT header created by mold linker.
+## Without the '-use-mold' option, "BOLT-ERROR: unable to disassemble instruction in PLT section .plt at offset 0x10" will be reported.
+## The only call instruction in main() should be a call to printf() in PLT.
+CHECK: callq "printf$plt
``````````
</details>
https://github.com/llvm/llvm-project/pull/115256
More information about the llvm-commits
mailing list