[DYNAREC] Implement perf map (#2212)

* [DYNAREC] Implement perf map

* [DYNAREC] Use function name as the Symbol

docs/USAGE.md

@@ -245,6 +245,11 @@ The GDBJIT debugging support, only available with the compilation option GDBJIT=
 * 1 : Dynarec will generate GDBJIT debuginfo.
 * 2 : Dynarec will generate detailed GDBJIT debuginfo with internal state.
 
+#### BOX64_DYNAREC_PERFMAP *
+Dynarec generate map file for Linux perf tool.
+* 0 : Dynarec will not generate perf map. (Default)
+* 1 : Dynarec will generate perf map.
+
 #### BOX64_DYNAREC_MISSING *
 Dynarec print the missing opcodes
 * 0 : not print the missing opcode (Default, unless DYNAREC_LOG>=1 or DYNAREC_DUMP>=1 is used)

docs/box64.pod

@@ -375,6 +375,13 @@ The GDBJIT debugging support, only available with the compilation option GDBJIT=
     * 1 : Dynarec will generate GDBJIT debuginfo.
     * 2 : Dynarec will generate detailed GDBJIT debuginfo with internal state.
 
+=item B<BOX64_DYNAREC_PERFMAP>=I<0|1>
+
+Dynarec generate map file for Linux perf tool.
+
+    * 0 : Dynarec will not generate perf map. (Default)
+    * 1 : Dynarec will generate perf map.
+
 =item B<BOX64_SSE_FLUSHTO0>=I<0|1>
 
 Handling of SSE Flush to 0 flags

src/core.c

@@ -17,6 +17,9 @@
 #include <linux/auxvec.h>
 #include <asm/hwcap.h>
 #endif
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
 #endif
 
 #include "build_info.h"
@@ -99,6 +102,8 @@ uintptr_t box64_dynarec_test_start = 0;
 uintptr_t box64_dynarec_test_end = 0;
 int box64_dynarec_gdbjit = 0;
 int box64_dynarec_df = 1;
+int box64_dynarec_perf_map = 0;
+int box64_dynarec_perf_map_fd = -1;
 #ifdef ARM64
 int arm64_asimd = 0;
 int arm64_aes = 0;
@@ -912,6 +917,15 @@ void LoadLogEnv()
         if (box64_dynarec_gdbjit)
             printf_log(LOG_INFO, "Dynarec will generate debuginfo for gdbjit\n");
     }
+    p = getenv("BOX64_DYNAREC_PERFMAP");
+    if (p) {
+        if (strlen(p) == 1) {
+            if (p[0] >= '0' && p[0] <= '1')
+                box64_dynarec_perf_map = p[0] - '0';
+        }
+        if (box64_dynarec_perf_map)
+            printf_log(LOG_INFO, "Dynarec will generate map file for Linux perf tool\n");
+    }
     p = getenv("BOX64_DYNAREC_DF");
     if(p) {
         if(strlen(p)==1) {
@@ -2526,6 +2540,14 @@ int initialize(int argc, const char **argv, char** env, x64emu_t** emulator, elf
 
     *emulator = emu;
 
+#ifdef DYNAREC
+    if (box64_dynarec_perf_map) {
+        char pathname[32];
+        snprintf(pathname, sizeof(pathname), "/tmp/perf-%d.map", getpid());
+        box64_dynarec_perf_map_fd = open(pathname, O_CREAT | O_RDWR | O_APPEND, S_IRUSR | S_IWUSR);
+    }
+#endif
+
     return 0;
 }
 
@@ -2565,5 +2587,12 @@ int emulate(x64emu_t* emu, elfheader_t* elf_header)
     }
 #endif
 
+#ifdef DYNAREC
+    if (box64_dynarec_perf_map && box64_dynarec_perf_map_fd != -1) {
+        close(box64_dynarec_perf_map_fd);
+        box64_dynarec_perf_map_fd = -1;
+    }
+#endif
+
     return ret;
 }

src/dynarec/arm64/dynarec_arm64_functions.c

@@ -734,7 +734,7 @@ static register_mapping_t register_mappings[] = {
 
 void inst_name_pass3(dynarec_native_t* dyn, int ninst, const char* name, rex_t rex)
 {
-    if (!box64_dynarec_dump && !box64_dynarec_gdbjit) return;
+    if (!box64_dynarec_dump && !box64_dynarec_gdbjit && !box64_dynarec_perf_map) return;
 
     static char buf[512];
     int length = sprintf(buf, "barrier=%d state=%d/%d/%d(%d:%d->%d:%d), %s=%X/%X, use=%X, need=%X/%X, sm=%d(%d/%d)",
@@ -849,6 +849,9 @@ void inst_name_pass3(dynarec_native_t* dyn, int ninst, const char* name, rex_t r
         }
         dyn->gdbjit_block = GdbJITBlockAddLine(dyn->gdbjit_block, (dyn->native_start + dyn->insts[ninst].address), inst_name);
     }
+    if (box64_dynarec_perf_map && box64_dynarec_perf_map_fd != -1) {
+        writePerfMap(dyn->insts[ninst].x64.addr, dyn->native_start + dyn->insts[ninst].address, dyn->insts[ninst].size / 4);
+    }
 }
 
 void print_opcode(dynarec_native_t* dyn, int ninst, uint32_t opcode)

src/dynarec/dynarec_native.c

@@ -841,3 +841,13 @@ void* FillBlock64(dynablock_t* block, uintptr_t addr, int alternate, int is32bit
     //block->done = 1;
     return (void*)block;
 }
+
+void writePerfMap(uintptr_t func_addr, uintptr_t code_addr, size_t code_size)
+{
+    char pbuf[128];
+    uint64_t sz = 0;
+    uintptr_t start = 0;
+    const char* symbname = FindNearestSymbolName(FindElfAddress(my_context, func_addr), (void*)func_addr, &start, &sz);
+    snprintf(pbuf, sizeof(pbuf), "0x%lx %ld %s\n", code_addr, code_size, symbname);
+    write(box64_dynarec_perf_map_fd, pbuf, strlen(pbuf));
+}

src/dynarec/la64/dynarec_la64_functions.c

@@ -26,6 +26,7 @@
 #include "custommem.h"
 #include "bridge.h"
 #include "gdbjit.h"
+#include "elfloader.h"
 
 #define XMM0 0
 #define XMM8 16
@@ -331,7 +332,7 @@ static register_mapping_t register_mappings[] = {
 
 void inst_name_pass3(dynarec_native_t* dyn, int ninst, const char* name, rex_t rex)
 {
-    if (!box64_dynarec_dump && !box64_dynarec_gdbjit) return;
+    if (!box64_dynarec_dump && !box64_dynarec_gdbjit && !box64_dynarec_perf_map) return;
 
     static char buf[512];
     int length = sprintf(buf, "barrier=%d state=%d/%d(%d), %s=%X/%X, use=%X, need=%X/%X, fuse=%d, sm=%d(%d/%d)",
@@ -397,6 +398,9 @@ void inst_name_pass3(dynarec_native_t* dyn, int ninst, const char* name, rex_t r
         }
         dyn->gdbjit_block = GdbJITBlockAddLine(dyn->gdbjit_block, (dyn->native_start + dyn->insts[ninst].address), inst_name);
     }
+    if (box64_dynarec_perf_map && box64_dynarec_perf_map_fd != -1) {
+        writePerfMap(dyn->insts[ninst].x64.addr, dyn->native_start + dyn->insts[ninst].address, dyn->insts[ninst].size / 4);
+    }
 }
 
 // will go badly if address is unaligned
@@ -522,4 +526,4 @@ void get_free_scratch(dynarec_la64_t* dyn, int ninst, uint8_t* tmp1, uint8_t* tm
     *tmp1 = tmp[0];
     *tmp2 = tmp[1];
     *tmp3 = tmp[2];
-}
+}

src/dynarec/rv64/dynarec_rv64_functions.c

@@ -698,7 +698,7 @@ static register_mapping_t register_mappings[] = {
 
 void inst_name_pass3(dynarec_native_t* dyn, int ninst, const char* name, rex_t rex)
 {
-    if (!box64_dynarec_dump && !box64_dynarec_gdbjit) return;
+    if (!box64_dynarec_dump && !box64_dynarec_gdbjit && !box64_dynarec_perf_map) return;
 
     static char buf[512];
     int length = sprintf(buf, "barrier=%d state=%d/%d(%d), %s=%X/%X, use=%X, need=%X/%X, fuse=%d, sm=%d(%d/%d), sew@entry=%d, sew@exit=%d",
@@ -772,6 +772,9 @@ void inst_name_pass3(dynarec_native_t* dyn, int ninst, const char* name, rex_t r
         }
         dyn->gdbjit_block = GdbJITBlockAddLine(dyn->gdbjit_block, (dyn->native_start + dyn->insts[ninst].address), inst_name);
     }
+    if (box64_dynarec_perf_map && box64_dynarec_perf_map_fd != -1) {
+        writePerfMap(dyn->insts[ninst].x64.addr, dyn->native_start + dyn->insts[ninst].address, dyn->insts[ninst].size / 4);
+    }
 }
 
 void print_opcode(dynarec_native_t* dyn, int ninst, uint32_t opcode)

src/include/debug.h

@@ -43,6 +43,8 @@ extern int box64_dynarec_missing;
 extern int box64_dynarec_aligned_atomics;
 extern int box64_dynarec_nativeflags;
 extern int box64_dynarec_df;
+extern int box64_dynarec_perf_map;
+extern int box64_dynarec_perf_map_fd;
 #ifdef ARM64
 extern int arm64_asimd;
 extern int arm64_aes;

src/include/dynarec_native.h

@@ -26,4 +26,6 @@ void addInst(instsize_t* insts, size_t* size, int x64_size, int native_size);
 void CancelBlock64(int need_lock);
 void* FillBlock64(dynablock_t* block, uintptr_t addr, int alternate, int is32bits);
 
-#endif //__DYNAREC_ARM_H_
+void writePerfMap(uintptr_t func_addr, uintptr_t code_addr, size_t code_size);
+
+#endif //__DYNAREC_ARM_H_