PPC: Use builtin D-Cache/I-Cache sync code.

src/buildvm_ppc.dasc

@@ -2533,6 +2533,31 @@ static void build_subroutines(BuildCtx *ctx)
   |//-- Miscellaneous functions --------------------------------------------
   |//-----------------------------------------------------------------------
   |
+  |// void lj_vm_cachesync(void *start, void *end)
+  |// Flush D-Cache and invalidate I-Cache. Assumes 32 byte cache line size.
+  |// This is a good lower bound, except for very ancient PPC models.
+  |->vm_cachesync:
+  |  // Compute start of first cache line and number of cache lines.
+  |  rlwinm CARG1, CARG1, 0, 0, 26
+  |  sub CARG2, CARG2, CARG1
+  |  addi CARG2, CARG2, 31
+  |  rlwinm. CARG2, CARG2, 27, 5, 31
+  |  beqlr
+  |  mtctr CARG2
+  |  mr CARG3, CARG1
+  |1:  // Flush D-Cache.
+  |  dcbst r0, CARG1
+  |  addi CARG1, CARG1, 32
+  |  bdnz <1
+  |  sync
+  |  mtctr CARG2
+  |1:  // Invalidate I-Cache.
+  |  icbi r0, CARG3
+  |  addi CARG3, CARG3, 32
+  |  bdnz <1
+  |  isync
+  |  blr
+  |
   |//-----------------------------------------------------------------------
   |//-- FFI helper functions -----------------------------------------------
   |//-----------------------------------------------------------------------

src/lj_mcode.c

@@ -29,15 +29,6 @@
 void sys_icache_invalidate(void *start, size_t len);
 #endif
 
-#if LJ_TARGET_LINUX && LJ_TARGET_PPC
-#include <dlfcn.h>
-static void (*mcode_sync_ppc)(void *start, void *end);
-static void mcode_sync_dummy(void *start, void *end)
-{
-    UNUSED(start); UNUSED(end);
-}
-#endif
-
 /* Synchronize data/instruction cache. */
 void lj_mcode_sync(void *start, void *end)
 {
@@ -48,14 +39,9 @@ void lj_mcode_sync(void *start, void *end)
   UNUSED(start); UNUSED(end);
 #elif LJ_TARGET_OSX
   sys_icache_invalidate(start, (char *)end-(char *)start);
-#elif LJ_TARGET_LINUX && LJ_TARGET_PPC
-  if (!mcode_sync_ppc) {
-    void *vdso = dlopen("linux-vdso32.so.1", RTLD_LAZY);
-    if (!vdso || !(mcode_sync_ppc = dlsym(vdso, "__kernel_sync_dicache")))
-      mcode_sync_ppc = mcode_sync_dummy;
-  }
-  mcode_sync_ppc(start, end);
-#elif defined(__GNUC__) && !LJ_TARGET_PPC
+#elif LJ_TARGET_PPC
+  lj_vm_cachesync(start, end);
+#elif defined(__GNUC__)
   __clear_cache(start, end);
 #else
 #error "Missing builtin to flush instruction cache"

src/lj_vm.h

@@ -27,6 +27,9 @@ LJ_ASMF void lj_vm_unwind_rethrow(void);
 #if LJ_TARGET_X86ORX64
 LJ_ASMF int lj_vm_cpuid(uint32_t f, uint32_t res[4]);
 #endif
+#if LJ_TARGET_PPC
+void lj_vm_cachesync(void *start, void *end);
+#endif
 LJ_ASMF double lj_vm_foldarith(double x, double y, int op);
 #if LJ_HASJIT
 LJ_ASMF double lj_vm_foldfpm(double x, int op);