Skip to content

Commit f89e006

Browse files
pablogsalpablogsal
committed
gh-136459: Use frame pointers in the x86_64 perf trampolines
1 parent 8b03dc1 commit f89e006

File tree

2 files changed

+82
-16
lines changed

2 files changed

+82
-16
lines changed

‎Python/asm_trampoline.S

Lines changed: 4 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -20,9 +20,10 @@ _Py_trampoline_func_start:
2020
#if defined(__CET__) && (__CET__ & 1)
2121
endbr64
2222
#endif
23-
sub $8, %rsp
24-
call *%rcx
25-
add $8, %rsp
23+
push %rbp
24+
mov %rsp, %rbp
25+
call *%rcx
26+
pop %rbp
2627
ret
2728
#endif // __x86_64__
2829
#if defined(__aarch64__) && defined(__AARCH64EL__) && !defined(__ILP32__)

‎Python/perf_jit_trampoline.c

Lines changed: 78 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -421,10 +421,12 @@ enum {
421421
DWRF_CFA_nop = 0x0, // No operation
422422
DWRF_CFA_offset_extended = 0x5, // Extended offset instruction
423423
DWRF_CFA_def_cfa = 0xc, // Define CFA rule
424+
DWRF_CFA_def_cfa_register = 0xd, // Define CFA register
424425
DWRF_CFA_def_cfa_offset = 0xe, // Define CFA offset
425426
DWRF_CFA_offset_extended_sf = 0x11, // Extended signed offset
426427
DWRF_CFA_advance_loc = 0x40, // Advance location counter
427-
DWRF_CFA_offset = 0x80 // Simple offset instruction
428+
DWRF_CFA_offset = 0x80, // Simple offset instruction
429+
DWRF_CFA_restore = 0xc0 // Restore register
428430
};
429431

430432
/* DWARF Exception Handling pointer encodings */
@@ -539,6 +541,7 @@ typedef struct ELFObjectContext {
539541
uint8_t* p; // Current write position in buffer
540542
uint8_t* startp; // Start of buffer (for offset calculations)
541543
uint8_t* eh_frame_p; // Start of EH frame data (for relative offsets)
544+
uint8_t* fde_p; // Start of FDE data (for PC-relative calculations)
542545
uint32_t code_size; // Size of the code being described
543546
} ELFObjectContext;
544547

@@ -804,7 +807,7 @@ static void elf_init_ehframe(ELFObjectContext* ctx) {
804807
*
805808
* DWRF_SECTION(FDE,
806809
* DWRF_U32((uint32_t)(p - framep)); // Offset to CIE (relative from here)
807-
* DWRF_U32(-0x30); // Initial PC-relative location of the code
810+
* DWRF_U32(pc_relative_offset); // PC-relative location of the code (calculated dynamically)
808811
* DWRF_U32(ctx->code_size); // Code range covered by this FDE
809812
* DWRF_U8(0); // Augmentation data length (none)
810813
*
@@ -873,11 +876,15 @@ static void elf_init_ehframe(ELFObjectContext* ctx) {
873876
*
874877
* The FDE describes unwinding information specific to this function.
875878
* It references the CIE and provides function-specific CFI instructions.
879+
*
880+
* The PC-relative offset is calculated after the entire EH frame is built
881+
* to ensure accurate positioning relative to the synthesized DSO layout.
876882
*/
877883
DWRF_SECTION(FDE,
878884
DWRF_U32((uint32_t)(p - framep)); // Offset to CIE (backwards reference)
879-
DWRF_U32(-0x30); // Machine code offset relative to .text
880-
DWRF_U32(ctx->code_size); // Address range covered by this FDE (code lenght)
885+
ctx->fde_p = p; // Remember where PC offset field is located for later calculation
886+
DWRF_U32(0); // Placeholder for PC-relative offset (calculated at end of elf_init_ehframe)
887+
DWRF_U32(ctx->code_size); // Address range covered by this FDE (code length)
881888
DWRF_U8(0); // Augmentation data length (none)
882889

883890
/*
@@ -888,17 +895,22 @@ static void elf_init_ehframe(ELFObjectContext* ctx) {
888895
* conventions and register usage patterns.
889896
*/
890897
#ifdef __x86_64__
891-
/* x86_64 calling convention unwinding rules */
898+
/* x86_64 calling convention unwinding rules with frame pointer */
892899
# if defined(__CET__) && (__CET__ & 1)
893-
DWRF_U8(DWRF_CFA_advance_loc | 8); // Advance location by 8 bytes when CET protection is enabled
894-
# else
895-
DWRF_U8(DWRF_CFA_advance_loc | 4); // Advance location by 4 bytes
900+
DWRF_U8(DWRF_CFA_advance_loc | 4); // Advance past endbr64 (4 bytes)
896901
# endif
897-
DWRF_U8(DWRF_CFA_def_cfa_offset); // Redefine CFA offset
898-
DWRF_UV(16); // New offset: SP + 16
899-
DWRF_U8(DWRF_CFA_advance_loc | 6); // Advance location by 6 bytes
900-
DWRF_U8(DWRF_CFA_def_cfa_offset); // Redefine CFA offset
901-
DWRF_UV(8); // New offset: SP + 8
902+
DWRF_U8(DWRF_CFA_advance_loc | 1); // Advance past push %rbp (1 byte)
903+
DWRF_U8(DWRF_CFA_def_cfa_offset); // def_cfa_offset 16
904+
DWRF_UV(16);
905+
DWRF_U8(DWRF_CFA_offset | DWRF_REG_BP); // offset r6 at cfa-16
906+
DWRF_UV(2);
907+
DWRF_U8(DWRF_CFA_advance_loc | 3); // Advance past mov %rsp,%rbp (3 bytes)
908+
DWRF_U8(DWRF_CFA_def_cfa_register); // def_cfa_register r6
909+
DWRF_UV(DWRF_REG_BP);
910+
DWRF_U8(DWRF_CFA_advance_loc | 3); // Advance past call *%rcx (2 bytes) + pop %rbp (1 byte) = 3
911+
DWRF_U8(DWRF_CFA_def_cfa); // def_cfa r7 ofs 8
912+
DWRF_UV(DWRF_REG_SP);
913+
DWRF_UV(8);
902914
#elif defined(__aarch64__) && defined(__AARCH64EL__) && !defined(__ILP32__)
903915
/* AArch64 calling convention unwinding rules */
904916
DWRF_U8(DWRF_CFA_advance_loc | 1); // Advance location by 1 instruction (stp x29, x30)
@@ -922,6 +934,58 @@ static void elf_init_ehframe(ELFObjectContext* ctx) {
922934
)
923935

924936
ctx->p = p; // Update context pointer to end of generated data
937+
938+
/* Calculate and update the PC-relative offset in the FDE
939+
*
940+
* When perf processes the jitdump, it creates a synthesized DSO with this layout:
941+
*
942+
* Synthesized DSO Memory Layout:
943+
* ┌─────────────────────────────────────────────────────────────┐ < code_start
944+
* │ Code Section │
945+
* │ (round_up(code_size, 8) bytes) │
946+
* ├─────────────────────────────────────────────────────────────┤ < start of EH frame data
947+
* │ EH Frame Data │
948+
* │ ┌─────────────────────────────────────────────────────┐ │
949+
* │ │ CIE data │ │
950+
* │ └─────────────────────────────────────────────────────┘ │
951+
* │ ┌─────────────────────────────────────────────────────┐ │
952+
* │ │ FDE Header: │ │
953+
* │ │ - CIE offset (4 bytes) │ │
954+
* │ │ - PC offset (4 bytes) <─ fde_offset_in_frame ─────┼────┼─> points to code_start
955+
* │ │ - address range (4 bytes) │ │ (this specific field)
956+
* │ │ CFI Instructions... │ │
957+
* │ └─────────────────────────────────────────────────────┘ │
958+
* ├─────────────────────────────────────────────────────────────┤ < reference_point
959+
* │ EhFrameHeader │
960+
* │ (navigation metadata) │
961+
* └─────────────────────────────────────────────────────────────┘
962+
*
963+
* The PC offset field in the FDE must contain the distance from itself to code_start:
964+
*
965+
* distance = code_start - fde_pc_field
966+
*
967+
* Where:
968+
* fde_pc_field_location = reference_point - eh_frame_size + fde_offset_in_frame
969+
* code_start_location = reference_point - eh_frame_size - round_up(code_size, 8)
970+
*
971+
* Therefore:
972+
* distance = code_start_location - fde_pc_field_location
973+
* = (ref - eh_frame_size - rounded_code_size) - (ref - eh_frame_size + fde_offset_in_frame)
974+
* = -rounded_code_size - fde_offset_in_frame
975+
* = -(round_up(code_size, 8) + fde_offset_in_frame)
976+
*
977+
* Note: fde_offset_in_frame is the offset from EH frame start to the PC offset field,
978+
*
979+
*/
980+
if (ctx->fde_p != NULL) {
981+
int32_t fde_offset_in_frame = (ctx->fde_p - ctx->startp);
982+
int32_t rounded_code_size = round_up(ctx->code_size, 8);
983+
int32_t pc_relative_offset = -(rounded_code_size + fde_offset_in_frame);
984+
985+
986+
// Update the PC-relative offset in the FDE
987+
*(int32_t*)ctx->fde_p = pc_relative_offset;
988+
}
925989
}
926990

927991
// =============================================================================
@@ -1117,6 +1181,7 @@ static void perf_map_jit_write_entry(void *state, const void *code_addr,
11171181
char buffer[1024]; // Buffer for DWARF data (1KB should be sufficient)
11181182
ctx.code_size = code_size;
11191183
ctx.startp = ctx.p = (uint8_t*)buffer;
1184+
ctx.fde_p = NULL; // Initialize to NULL, will be set when FDE is written
11201185

11211186
/* Generate EH frame (Exception Handling frame) data */
11221187
elf_init_ehframe(&ctx);

0 commit comments

Comments
 (0)