Program Loading and Dynamic Linking
Program Loading A number of criteria constrain the mapping of an executable file or shared object file to virtual memory segments. During mapping, the operating system may use delayed physical reads to improve performance, which necessitates that file offsets and virtual addresses are congruent, modulo the page size. Page size must be less than or equal to the operating system implemented congruency. This ABI defines 64 KB congruency as the minimum allowable. To maintain interoperability between operating system implementations, 64 KB congruency is recommended. There is historical precedence for 64 KB congruency in that there is synergy with the Power Architecture instruction set whereby low and high adjusted relocations can be easily performed using addi or addis instructions. The value of the p_align member of the program header struct must be 0x10000 or a larger power of 2. If a larger congruency size is used for large pages, p_align should match the congruency value. The following program header information illustrates an application that is mapped with a base address of 0x10000000: Program Header Example Header Member Text Segment Data Segment p_type PT_LOAD PT_LOAD p_offset 0x000000 0x000af0 p_vaddr 0x10000000 0x10010af0 p_paddr 0x10000000 0x10010af0 p_filesz 0x00af0 0x00124 p_memsz 0x00af0 0x00128 p_flags R-E RW- p_align 0x10000 0x10000
For the PT_LOAD entry describing the data segment, the p_memsz may be greater than the p_filesz. The difference is the size of the .bss section. On implementations that use virtual memory file mapping, only the portion of the file between the .data p_offset (rounded down to the nearest page) to p_offset + p_filesz (rounded up to the next page size) is included. If the distance between p_offset + p_filesz and p_offset + p_memsz crosses a page boundary, then additional memory must be allocated out of anonymous memory to include data through p_vaddr + p_memsz. demonstrates a typical mapping of file to memory segments. Memory Segment Mappings File Section Virtual Address 0x0 header 0x10000000 0x100 .text 0x10000100 0xaf0 .data 0x10010af0 Not applicable. Zero-initialized data is not stored in the file. .bss 0x10010c14 Not stored in the file. End of sections 0x10010c18
Operating systems typically enforce memory permission on a per-page granularity. This ABI maintains that the memory permissions are consistent across each memory segment when a file image is mapped to a process memory segment. The text segment and data segment require differing memory permissions. To maintain congruency of file offset to virtual address modulo the page size, the system maps the file region holding the overlapped text and data twice at different virtual addresses for each segment (see ). To increase the security attributes of this ABI, the text and certain sections of the data segment (such as the .rodata section) may be protected as read only after the pages are mapped and relocations are resolved. See for more information.
File Image to Process Memory Image Mapping
As a result of this mapping, there can be up to four pages of impure text or data in the virtual memory segments for the application as described in the following list: ELF header information, program headers, and other information will precede the .text section and reside at the beginning of the text segment. The last memory page of the text segment can contain a copy of the partial, first file-image data page as an artifact of page faulting the last file-image text page from the file image to the text segment while maintaining the required offsets as shown in . Likewise, the first memory page of the data segment may contain a copy of the partial, last file-image text page as an artifact of page faulting the first file-image data page from the file image to the data segment while maintaining the required offsets. The last faulted data-segment memory page may contain residual data from the last file-image data page that is not part of the actual file image. The system is required to zero this residual memory after that page is mapped to the data segment. If the application requires static data, the remainder of this page is used for that purpose. If the static data requirements exceed the remnant left in the last faulted memory page, additional pages shall be mapped from anonymous memory and zeroed. The handling of the contents of the first three pages is undefined by this ABI. They are unused by the executable program once started.
Addressing Models When mapping an executable file or shared object file to memory, the system can use the following addressing models. Each application is allocated its own virtual address space. Traditionally, executable files are mapped to virtual memory using an absolute addressing model, where the mapping of the sections to segments uses the section p_vaddr specified by the ELF header directly as an absolute address. The position-independent code (PIC) addressing model allows the file image text of an executable file or shared object file to be loaded into the virtual address space of a process at an arbitrary starting address chosen by the kernel loader or program interpreter (dynamic linker). Shared objects need to use the PIC addressing model so that all references to global variables go through the Global Offset Table. Position-independent executables should use the PIC addressing model.
Process Initialization To provide a standard environment for application programs, the exec system call creates an initial program machine state. That state includes the use of registers, the layout of the stack frame, and argument passing. For example, a C program might typically issue the following declaration to begin executing at the local entry point of a function named main: extern int main (int argc, char *argv[ ], char *envp[ ], void *auxv[ ]); int main(int argc, char *argv[ ], char *envp[ ], ElfW(auxv_t) *auxvec) where: argc is a nonnegative argument count. argv is an array of argument strings. It is terminated by a NULL pointer, argv[argc] == 0. envp is an array of environment strings. It is also terminated by a NULL pointer. auxv is an array of structures that contain the auxiliary vector. It is terminated by a structure entry with an a_type of AT_NULL. For more information, see . This section explains how to implement the call to main or to the entry point.
Registers The contents of most registers are not specified when a process is first entered from an exec system call. A program should not expect the operating system to set all registers to 0. If a register other than those listed in must have a specific value, the program must set it to that value during process initialization. The contents of the following registers are specified: Registers Specified during Process Initialization Register Description r1 The initial stack pointer, aligned to a quadword boundary. r2 Undefined. r3 Contains argc, the nonnegative argument count. r4 Contains argv, a pointer to the array of argument pointers in the stack. The array is immediately followed by a NULL pointer. If there are no arguments, r4 points to a NULL pointer. r5 Contains envp, a pointer to the array of environment pointers in the stack. The array is immediately followed by a NULL pointer. If no environment exists, r5 points to a NULL pointer. r6 Contains a pointer to the auxiliary vector. The auxiliary vector shall have at least one member, a terminating entry with an a_type of AT_NULL (see ). r7 Contains a termination function pointer. If r7 contains a nonzero value, the value represents a function pointer that the application should register with atexit. If r7 contains zero, no action is required. r12 Contains the address of the global entry point of the first function being invoked, which represents the start address of the executable specified in the exec call. FPSCR Contains 0, specifying “round to nearest” mode for both binary and decimal rounding modes, IEEE Mode, and the disabling of floating-point exceptions. VSCR Vector Status and Control Register. Contains 0, specifying vector Java/IEEE mode and that no saturation has occurred.
The run-time that gets control from _start is responsible for: Creating the first stack frame Initializing the first stack frame's back chain pointer to NULL Allocating and initializing TLS storage Initializing the thread control block (TCB) and dynamic thread vector (DTV) Initializing any __thread variables Setting R13 for the initial process thread. This initialization must be completed before any library initialization codes are run and before control is transferred to the main program (main( )).
Process Stack Although every process has a stack, no fixed stack address is defined by the system. In addition, a program's stack address can change from one system to another. It can even change from one process invocation to another. Thus, the process initialization code must use the stack address in general-purpose register r1. Data in the stack segment at addresses below the stack pointer contain undefined values.
Auxiliary Vector The argument and environment vectors transmit information from one application program to another. However, the auxiliary vector conveys information from the operating system to the program. This vector is an array of structures, defined as follows: typedef struct { long a_type; union { long a_val; void *a_ptr; void (*a_fcn)( ); } a_un; } auxv_t; Name Value a_un field Comment AT_NULL 0 ignored /* End of vector */ AT_PHDR 3 a_ptr /* Program headers for program */ AT_PHENT 4 a_val /* Size of program header entry */ AT_PHNUM 5 a_val /* Number of program headers */ AT_PAGESZ 6 a_val /* System page size */ AT_BASE 7 a_ptr /* Base address of interpreter */ AT_FLAGS 8 a_val /* Flags */ AT_ENTRY 9 a_ptr /* Entry point of program */ AT_UID 11 /* Real user ID (uid) */ AT_EUID 12 /* Effective user ID (euid) */ AT_GID 13 /* Real group ID (gid) */ AT_EGID 14 /* Effective group ID (egid) */ AT_PLATFORM 15 a_ptr /* String identifying platform. */ AT_HWCAP 16 a_val /* Machine-dependent hints about processor capabilities. */ AT_CLKTCK 17 /* Frequency of times( ), always 100 */ AT_DCACHEBSIZE 19 a_val /* Data cache block size */ AT_ICACHEBSIZE 20 a_val /* Instruction cache block size */ AT_UCACHEBSIZE 21 a_val /* Unified cache block size */ AT_IGNOREPPC 22 /* Ignore this entry! */ AT_SECURE 23 /* Boolean, was exec authorized to use setuid or setgid */ AT_BASE_PLATFORM 24 a_ptr /* String identifying real platforms */ AT_RANDOM 25 /* Address of 16 random bytes */ AT_HWCAP2 26 a_val /* More machine-dependent hints about processor capabilities. */ AT_EXECFN 31 /* File name of executable */ AT_SYSINFO_EHDR 33 /* In many architectures, the kernel provides a virtual dynamic shared object (VDSO) that contains a function callable from the user state. AT_SYSINFO_EHDR is the address of the VDSO header that is used by the dynamic linker to resolve function symbols with the VDSO. */ AT_L1I_CACHESIZE 40 /* Cache sizes and geometries. */ AT_L1I_CACHEGEOMETRY 41 AT_L1D_CACHESIZE 42 AT_L1D_CACHEGEOMETRY 43 AT_L2_CACHESIZE 44 AT_L2_CACHEGEOMETRY 45 AT_L3_CACHESIZE 46 AT_L3_CACHEGEOMETRY 47 AT_NULL The auxiliary vector has no fixed length; instead an entry of this type denotes the end of the vector. The corresponding value of a_un is undefined. AT_PHDR Under some conditions, the system creates the memory image of the application program before passing control to an interpreter program. When this happens, the a_ptr member of the AT_PHDR entry tells the interpreter where to find the program header table in the memory image. If the AT_PHDR entry is present, entries of types AT_PHENT, AT_PHNUM, and AT_ENTRY must also be present. See the Program Header section in Chapter 5 of the System V ABI for more information about the program header table. AT_PHENT The a_val member of this entry holds the size, in bytes, of one entry in the program header table to which the AT_PHDR entry points. AT_PHNUM The a_val member of this entry holds the number of entries in the program header table to which the AT_PHDR entry points. AT_PAGESZ If present, this entry's a_val member gives the system page size in bytes. The same information is also available through the sysconf system call. AT_BASE The a_ptr member of this entry holds the base address at which the interpreter program was loaded into memory. See the Program Header section in Chapter 5 of the System V ABI for more information about the base address. AT_FLAGS If present, the a_val member of this entry holds 1-bit flags. Bits with undefined semantics are set to zero. Other auxiliary vector types are reserved. No flags are currently defined for AT_FLAGS on the 64-bit OpenPOWER ABI Architecture. AT_ENTRY The a_ptr member of this entry holds the entry point of the application program to which the interpreter program should transfer control. AT_DCACHEBSIZE The a_val member of this entry gives the data cache block size for processors on the system on which this program is running. If the processors have unified caches, AT_DCACHEBSIZE is the same as AT_UCACHEBSIZE. AT_ICACHEBSIZE The a_val member of this entry gives the instruction cache block size for processors on the system on which this program is running. If the processors have unified caches, AT_ICACHEBSIZE is the same as AT_UCACHEBSIZE. AT_UCACHEBSIZE The a_val member of this entry is zero if the processors on the system on which this program is running do not have a unified instruction and data cache. Otherwise, it gives the cache block size. AT_PLATFORM The a_ptr member is the address of the platform name string. For virtualized systems, this may be different (that is, an older platform) than the physical machine running this environment. AT_BASE_PLATFORM The a_ptr member is the address of the platform name string for the physical machine. For virtualized systems, this will be the platform name of the real hardware. AT_HWCAP The a_val member of this entry is a bit map of hardware capabilities. Some bit mask values include: PPC_FEATURE_32 0x80000000 /* Always set for powerpc64 */ PPC_FEATURE_64 0x40000000 /* Always set for powerpc64 */ PPC_FEATURE_HAS_ALTIVEC 0x10000000 PPC_FEATURE_HAS_FPU 0x08000000 PPC_FEATURE_HAS_MMU 0x04000000 PPC_FEATURE_UNIFIED_CACHE 0x01000000 PPC_FEATURE_NO_TB 0x00100000 /* 601/403gx have no timebase */ PPC_FEATURE_POWER4 0x00080000 /* POWER4 ISA 2.00 */ PPC_FEATURE_POWER5 0x00040000 /* POWER5 ISA 2.02 */ PPC_FEATURE_POWER5_PLUS 0x00020000 /* POWER5+ ISA 2.03 */ PPC_FEATURE_CELL_BE 0x00010000 /* CELL Broadband Engine */ PPC_FEATURE_BOOKE 0x00008000 /* ISA Category Embedded */ PPC_FEATURE_SMT 0x00004000 /* Simultaneous Multi-Threading */ PPC_FEATURE_ICACHE_SNOOP 0x00002000 PPC_FEATURE_ARCH_2_05 0x00001000 /* ISA 2.05 */ PPC_FEATURE_PA6T 0x00000800 /* PA Semi 6T Core */ PPC_FEATURE_HAS_DFP 0x00000400 /* Decimal FP Unit */ PPC_FEATURE_POWER6_EXT 0x00000200 /* P6 + mffgpr/mftgpr */ PPC_FEATURE_ARCH_2_06 0x00000100 /* ISA 2.06 */ PPC_FEATURE_HAS_VSX 0x00000080 /* P7 Vector Extension. */ PPC_FEATURE_PSERIES_PERFMON_COMPAT 0x00000040 PPC_FEATURE_TRUE_LE 0x00000002 PPC_FEATURE_PPC_LE 0x00000001 Bit 0x00000004 is reserved for kernel use. AT_HWCAP2 The a_val member of this entry is a bit map of hardware capabilities. Some bit mask values include: PPC_FEATURE2_ARCH_2_07 0x80000000 /* ISA 2.07 */ PPC_FEATURE2_HAS_HTM 0x40000000 /* Hardware Transactional Memory */ PPC_FEATURE2_HAS_DSCR 0x20000000 /* Data Stream Control Register */ PPC_FEATURE2_HAS_EBB 0x10000000 /* Event Base Branching */ PPC_FEATURE2_HAS_ISEL 0x08000000 /* Integer Select */ PPC_FEATURE2_HAS_TAR 0x04000000 /* Target Address Register */ PPC_FEATURE2_HAS_VCRYPTO 0x02000000 /* The processor implements the Vector.AES category */ PPC_FEATURE2_HTM_NOSC 0x01000000 PPC_FEATURE2_ARCH_3_00 0x00800000 /* ISA 3.0 */ PPC_FEATURE2_HAS_IEEE128 0x00400000 /* VSX IEEE Binary Float 128-bit */ PPC_FEATURE2_DARN 0x00200000 /* darn instruction */ PPC_FEATURE2_SCV 0x00100000 /* scv syscall */ PPC_FEATURE2_HTM_NO_SUSPEND 0x00080000 /* TM without suspended state */ PPC_FEATURE2_ARCH_3_1 0x00040000 /* ISA 3.1 */ PPC_FEATURE2_MMA 0x00020000 /* Matrix Multiply Accumulate */ When a process starts to execute, its stack holds the arguments, environment, and auxiliary vector received from the exec call. The system makes no guarantees about the relative arrangement of argument strings, environment strings, and the auxiliary information, which appear in no defined or predictable order. Further, the system may allocate memory after the null auxiliary vector entry and before the beginning of the information block. AT_L1I_CACHESIZE The size of the level-1 instruction cache, in bytes. AT_L1I_CACHEGEOMETRY The geometry of the level-1 instruction cache. The low-order sixteen bits contain the cache associativity as a value N, where N = 1 represents a direct-mapped cache, N = 0xffff represents a fully associative cache, and any other N represents an N-way set-associative cache. The next higher-order sixteen bits contain the size of the cache line in bytes. Note that the cache line size is not necessarily the same as the cache block size. AT_L1D_CACHESIZE The size of the level-1 data cache, in bytes. AT_L1D_CACHEGEOMETRY The geometry of the level-1 data cache, defined in the same manner as for AT_L1I_CACHEGEOMETRY. AT_L2_CACHESIZE The size of the level-2 cache, in bytes. AT_L2_CACHEGEOMETRY The geometry of the level-2 cache, defined in the same manner as for AT_L1I_CACHEGEOMETRY. AT_L3_CACHESIZE The size of the level-3 cache, in bytes. AT_L3_CACHEGEOMETRY The geometry of the level-3 cache, defined in the same manner as for AT_L1I_CACHEGEOMETRY.
Dynamic Linking
Program Interpreter For dynamic linking, the standard program interpreter is /lib/ld64.so.2. It may be located in different places on different distributions.
Dynamic Section The dynamic section provides information used by the dynamic linker to manage dynamically loaded shared objects, including relocation, initialization, and termination when loaded or unloaded, resolving dependencies on other shared objects, resolving references to symbols in the shared object, and supporting debugging. The following dynamic tags are relevant to this processor-specific ABI: DT_PLTGOT The d_ptr member of this dynamic tag points to the first byte of the PLT. DT_JMPREL The d_ptr member of this dynamic tag points to the first byte of the table of relocation entries, which have a one-to-one correspondence with PLT entries. Any executable or shared object with a PLT must have DT_JMPREL. A shared object containing only data will not have a PLT and thus will not have DT_JMPREL. DT_PPC64_GLINK (DT_LOPROC + 0) The d_ptr member of this dynamic tag points to 32 bytes before the .glink lazy link symbol resolver stubs that are described in . DT_PPC64_OPT (DT_LOPROC + 3) The d_val member of this dynamic tag specifies whether various optimizations are possible. The low bit will be set to indicate that an optimized __tls_get_addr call stub is used. The next most-significant bit will be set if multiple TOCs are present.
Global Offset Table To support position-independent code, a Global Offset Table (GOT) shall be constructed by the link editor in the data segment when linking code that contains any of the various R_PPC64_GOT* relocations or when linking code that references the .TOC. address. The GOT consists of an 8-byte header that contains the TOC base (the first TOC base when multiple TOCs are present), followed by an array of 8-byte addresses. The link editor shall emit dynamic relocations as appropriate for each entry in the GOT. At runtime, the dynamic linker will apply these relocations after the addresses of all memory segments are known (and thus the addresses of all symbols). While the GOT may be appear to be an array of absolute addresses, this ABI does not preclude the GOT containing nonaddress entries and specifies the presence of nonaddress tls_index entries. Absolute addresses are generated for all GOT relocations by the dynamic linker before giving control to general application code. (However, IFUNC resolution functions may be invoked before relocation is completed, limiting the use of global variables by such functions.) The dynamic linker is free to choose different memory segment addresses for the executable or shared objects in a different process image. After the initial mapping of the process image by the dynamic linker, memory segments reside at fixed addresses for the life of a process. The symbol .TOC. may be used to access the GOT or in TOC-relative addressing to other data constructs, such as the procedure linkage table. The symbol may be offset by 0x8000 bytes, or another offset, from the start of the .got section. This offset allows the use of the full (64 KB) signed range of 16-bit displacement fields by using both positive and negative subscripts into the array of addresses, or a larger offset to afford addressing using references within ±2 GB with 32-bit displacements. The 32-bit displacements are constructed by using the addis instruction to provide a first high-order 16-bit portion of a 32-bit displacement in conjunction with an instruction to supply a low-order 16-bit portion of a 32-bit displacement. In PIC code that uses the TOC, the TOC pointer r2 points to the TOC base, enabling easy reference. For static nonrelocatable modules, the GOT address is fixed and can be directly used by code. All functions except leaf routines must load the value of the TOC base into the TOC register r2. Code may access GOT entries directly using PC-relative addressing, where available.
Function Addresses The following requirements concern function addresses. When referencing a function address, consider the following requirements: Intraobject executable or shared object function address references may be resolved by the dynamic linker to the absolute virtual address of the symbol. Function address references from within the executable file to a function defined in a shared object file are resolved by the link editor to the .text section address of the PLT call stub for that functionwithin the executable file. In a static module, when a function pointer reference is made to a function provided by a dynamically loaded shared module, the function may be resolved to the address of a PLT stub. If this resolution is made, all function pointer references must be made through the same PLT stub in the static module to ensure correct intraobject comparisons for function addresses. A function address of a nested function may also be resolved to the address of a trampoline used to call it. When comparing function addresses, consider the following requirements: The address of a function shall compare to the same value in executables and shared objects. For intraobject comparisons of function addresses within the executable or shared object, the link editor may directly compare the absolute virtual addresses. For a function address comparison where an executable references a function defined in a a shared object, the link editor will place the address of a .text section PLT call stub for that function in the corresponding dynamic symbol table entry's st_value field (see ). When the dynamic linker loads shared objects associated with an executable and resolves any GOT entry relocations into absolute addresses, it will search the dynamic symbol table of the executable for each symbol that needs to be resolved. If it finds the symbol and the st_value of the symbol table entry is nonzero, it shall use the address indicated in the st_value entry as the symbol’s address. If the dynamic linker does not find the symbol in the executable’s dynamic symbol table or the entry’s st_value member is zero, the dynamic linker may consider the symbol as undefined in the executable file.
Procedure Linkage Table When the link editor builds an executable file or shared object file, it does not know the absolute address of undefined function calls. Therefore, it cannot generate code to directly transfer execution to another shared object or executable. For each execution transfer to an undefined function call in the file image, the link editor places a relocation against an entry in the Procedure Linkage Table (PLT) of the executable or shared object that corresponds to that function call. Additionally, for all nonstatic functions with standard (nonhidden) visibility in a shared object, the link editor invokes the function through the PLT, even if the shared object defines the function. The same is not true for executables. The link editor knows the number of functions invoked through the PLT, and it reserves space for an appropriately sized .plt section. The .plt section is located in the section following the .got. It consists of an array of addresses and is initialized by the module loader. There will also be an array of R_PPC_JMP_SLOT relocations in .rela.plt, with a one-to-one correspondence between elements of each array. Each R_PPC_JMP_SLOT relocation will have r_offset pointing at the .plt word it relocates. A unique PLT is constructed by the static linker for each static module (that is, the main executable) and each dynamic shared object. The PLT is located in the data segment of the process image at object load time by the dynamic linker using the information about the .plt section stored in the file image. The individual PLT entries are populated by the dynamic linker using one of the following binding methods. Execution can then be redirected to a dependent shared object or executable.
Lazy Binding The lazy binding method is the default. It delays the resolution of a PLT entry to an absolute address until the function call is made the first time. The benefit of this method is that the application does not pay the resolution cost until the first time it needs to call the function, if at all. To implement lazy binding, the dynamic loader points each PLT entry to a lazy resolution stub at load time. After the function call is made the first time, this lazy resolution stub gets control, resolves the symbol, and updates the PLT entry to hold the final value to be used for future calls.
Immediate Binding The immediate binding method resolves the absolute addresses of all PLT entries in the executable and dependent shared objects at load time, before passing execution control to the application. The environment variable LD_BIND_NOW may be set to a nonnull value to signal the dynamic linker that immediate binding is requested at load time, before control is given to the application. For some performance-sensitive situations, it may be better to pay the resolution cost to populate the PLT entries up front rather than during execution.
Procedure Linkage Table For every call site that needs to use the PLT, the link editor constructs a call stub in the .text section and resolves the call site to use that call stub. The call stub transfers control to the address indicated in the PLT entry. These call stubs need not be adjacent to one another or unique. They can be scattered throughout the text segment so that they can be reached with a branch and link instruction. Depending on relocation information at the call site, the stub provides one of the following properties: The caller has set up r2 to hold the TOC pointer and expects the PLT call stub to save that value to the TOC save stack slot. This is the default. The caller has set up r2 to hold the TOC pointer and has already saved that value to the TOC save stack slot itself. This is indicated by the presence of a R_PPC64_TOCSAVE relocation on the nop following the call. tocsaveloc: nop ... bl target .reloc ., R_PPC64_TOCSAVE, tocsaveloc nop The caller has not set up r2 to hold the TOC pointer. This is indicated by use of a R_PPC64_REL24_NOTOC relocation (instead of R_PPC64_REL24) on the call instruction. In any scenario, the PLT call stub must transfer control to the function whose address is provided in the associated PLT entry. This address is treated as a global entry point for ABI purposes. This means that the PLT call stub loads the address into r12 before transferring control. Although the details of the call stub implementation are left to the link editor, some examples are provided. In those examples, func@plt is used to denote the address of the PLT entry for func; func@plt@toc denotes the offset of that address relative to the TOC pointer; and the @ha and @l variants denote the high-adjusted and low parts of these values as usual. Because the link editor synthesizes the PLT call stubs directly, it can determine all these values as immediate constants. The assembler is not required to support those notations. A possible implementation for case 1 looks as follows (if func@plt@toc is less than 32 KB, the call stub may be simplified to omit the addis): std r2,24(r1) addis r12,r2,func@plt@toc@ha ld r12,func@plt@toc@l(r12) mtctr r12 bctr For case 2, the same implementation as for case 1 may be used, except that the first instruction “std r2,24(r1)” is omitted: addis r12,r2,func@plt@toc@ha ld r12,func@plt@toc@l(r12) mtctr r12 bctr A possible implementation for case 3 looks as follows: mflr r0 bcl 20,31,1f 1: mflr r2 mtlr r0 addis r2,r2,(.TOC.-1b)@ha addi r2,r2,(.TOC.-1b)@l addis r12,r2,func@plt@toc@ha ld r12,func@plt@toc@l(r12) mtctr r12 bctr When generating non-PIC code for the small or medium code model, a simpler variant may alternatively be used for cases 2 or 3: lis r12,func@plt@ha ld r12,func@plt@l(r12) mtctr r12 bctr When PC-relative addressing is available, another simpler variant may alternatively be used for cases 2 or 3: pld r12, func@plt@pcrel mtctr r12 bctr To support lazy binding, the link editor also provides a set of symbol resolver stubs, one for each PLT entry. Each resolver stub consists of a single instruction, which is usually a branch to a common resolver entry point or a nop. The resolver stubs are placed in the .glink section, which is merged into the .text section of the final executable or dynamic object. The address of the resolver stubs is communicated to the dynamic loader through the DT_PPC64_GLINK dynamic section entry. The address of the symbol resolver stub associated with PLT entry N is determined by adding 4xN + 32 to the d_ptr field of the DT_PPC64_GLINK entry. When using lazy binding, the dynamic linker initializes each PLT entry at load time to that address. The resolver stubs provided by the link editor must call into the main resolver routine provided by the dynamic linker. This resolver routine must be called with r0 set to the index of the PLT entry to be resolved, r11 set to the identifier of the current dynamic object, and r12 set to the resolver entry point address (as usual when calling a global entry point). The resolver entry point address and the dynamic object identifier are installed at load time by the dynamic linker into the two doublewords immediately preceding the array of PLT entries, allowing the resolver stubs to retrieve these values from there. These two doublewords are considered part of the .plt section; the DT_PLTGOT dynamic section entry points to the first of those words. Beyond the above requirements, the implementation of the .glink resolver stubs is up to the link editor. The following shows an example implementation: # ABI note: At entry to the resolver stub: # - r12 holds the address of the res_N stub for the target routine # - all argument registers hold arguments for the target routine PLTresolve: # Determine addressability. This sequence works for both PIC # and non-PIC code and does not rely on presence of the TOC pointer. mflr r0 bcl 20,31,1f 1: mflr r11 mtlr r0 # Compute .plt section index from entry point address in r12 # .plt section index is placed into r0 as argument to the resolver sub r12,r12,r11 subi r12,r12,res_0-1b srdi r0,r12,2 # Load address of the first byte of the PLT ld r12,PLToffset-1b(r11) add r11,r12,r11 # Load resolver address and DSO identifier from the # first two doublewords of the PLT ld r12,0(r11) ld r11,8(r11) # Branch to resolver mtctr r12 bctr # ABI note: At entry to the resolver: # - r12 holds the resolver address # - r11 holds the DSO identifier # - r0 holds the PLT index of the target routine # - all argument registers hold arguments for the target routine # Constant pool holding offset to the PLT # Note that there is no actual symbol PLT; the link editor # synthesizes this value when creating the .glink section PLToffset: .quad PLT-. # A table of branches, one for each PLT entry # The idea is that the PLT call stub loads r12 with these # addresses, so (r12 - res_0) gives the PLT index × 4. res_0: b PLTresolve res_1: b PLTresolve ... After resolution, the value of a PLT entry in the PLT is the address of the function’s global entry point, unless the resolver can determine that a module-local call occurs with a shared TOC value wherein the TOC is shared between the caller and the callee.