amd/rtld: update the ELF representation of LDS symbols
authorNicolai Hähnle <nicolai.haehnle@amd.com>
Sun, 16 Jun 2019 23:24:29 +0000 (01:24 +0200)
committerMarek Olšák <marek.olsak@amd.com>
Tue, 25 Jun 2019 01:04:10 +0000 (21:04 -0400)
The initial prototype used a processor-specific symbol type, but
feedback suggests that an approach using processor-specific section
name that encodes the alignment analogous to SHN_COMMON symbols is
preferred.

This patch keeps both variants around for now to reduce problems
with LLVM compatibility as we switch branches around.

This also cleans up the error reporting in this function.

Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
src/amd/common/ac_rtld.c

index 57d6b0151b42ab3235ac366cb394a60d20d38c42..ebf64d9165885ec332c64e68e2dc7dadc90e14e2 100644 (file)
 #define MY_EM_AMDGPU 224
 
 #ifndef STT_AMDGPU_LDS
-#define STT_AMDGPU_LDS 13
+#define STT_AMDGPU_LDS 13 // this is deprecated -- remove
+#endif
+
+#ifndef SHN_AMDGPU_LDS
+#define SHN_AMDGPU_LDS 0xff00
 #endif
 
 #ifndef R_AMDGPU_NONE
@@ -176,13 +180,20 @@ static bool read_private_lds_symbols(struct ac_rtld_binary *binary,
                                     Elf_Scn *section,
                                     uint32_t *lds_end_align)
 {
-#define report_elf_if(cond) \
+#define report_if(cond) \
        do { \
                if ((cond)) { \
                        report_errorf(#cond); \
                        return false; \
                } \
        } while (false)
+#define report_elf_if(cond) \
+       do { \
+               if ((cond)) { \
+                       report_elf_errorf(#cond); \
+                       return false; \
+               } \
+       } while (false)
 
        struct ac_rtld_part *part = &binary->parts[part_idx];
        Elf64_Shdr *shdr = elf64_getshdr(section);
@@ -194,15 +205,21 @@ static bool read_private_lds_symbols(struct ac_rtld_binary *binary,
        size_t num_symbols = symbols_data->d_size / sizeof(Elf64_Sym);
 
        for (size_t j = 0; j < num_symbols; ++j, ++symbol) {
-               if (ELF64_ST_TYPE(symbol->st_info) != STT_AMDGPU_LDS)
+               struct ac_rtld_symbol s = {};
+
+               if (ELF64_ST_TYPE(symbol->st_info) == STT_AMDGPU_LDS) {
+                       /* old-style LDS symbols from initial prototype -- remove eventually */
+                       s.align = MIN2(1u << (symbol->st_other >> 3), 1u << 16);
+               } else if (symbol->st_shndx == SHN_AMDGPU_LDS) {
+                       s.align = MIN2(symbol->st_value, 1u << 16);
+                       report_if(!util_is_power_of_two_nonzero(s.align));
+               } else
                        continue;
 
-               report_elf_if(symbol->st_size > 1u << 29);
+               report_if(symbol->st_size > 1u << 29);
 
-               struct ac_rtld_symbol s = {};
                s.name = elf_strptr(part->elf, strtabidx, symbol->st_name);
                s.size = symbol->st_size;
-               s.align = MIN2(1u << (symbol->st_other >> 3), 1u << 16);
                s.part_idx = part_idx;
 
                if (!strcmp(s.name, "__lds_end")) {
@@ -224,6 +241,7 @@ static bool read_private_lds_symbols(struct ac_rtld_binary *binary,
 
        return true;
 
+#undef report_if
 #undef report_elf_if
 }
 
@@ -522,7 +540,9 @@ static bool resolve_symbol(const struct ac_rtld_upload_info *u,
                           unsigned part_idx, const Elf64_Sym *sym,
                           const char *name, uint64_t *value)
 {
-       if (sym->st_shndx == SHN_UNDEF) {
+       /* TODO: properly disentangle the undef and the LDS cases once
+        * STT_AMDGPU_LDS is retired. */
+       if (sym->st_shndx == SHN_UNDEF || sym->st_shndx == SHN_AMDGPU_LDS) {
                const struct ac_rtld_symbol *lds_sym =
                        find_symbol(&u->binary->lds_symbols, name, part_idx);