diff --git a/src/ffi/platforms/ppc64.cc b/src/ffi/platforms/ppc64.cc
index a0a53cfdb44363..78cd914405618c 100644
--- a/src/ffi/platforms/ppc64.cc
+++ b/src/ffi/platforms/ppc64.cc
@@ -134,13 +134,21 @@ extern "C" bool node_ffi_create_fast_trampoline(
   }
 
   // Load the target address from the literal pool into r12, then branch through
-  // CTR. ELFv2 functions can use r12 to establish their TOC on global entry.
-  Emit32(&cursor, Bl(1));              // bl .+4
-  Emit32(&cursor, Mfspr(12, 8));       // mflr r12
-  Emit32(&cursor, Ld(12, 12, 20));     // ld r12, literal-mflr(r12)
-  Emit32(&cursor, Mtspr(9, 12));       // mtctr r12
-  Emit32(&cursor, 0x4e800420);         // bctr
-  Emit32(&cursor, 0x60000000);         // nop; align literal to 8 bytes
+  // CTR. Save the caller's link register before using `bl` to obtain the
+  // trampoline's address, and restore it before the tail branch so the native
+  // target returns to V8 rather than back into the trampoline. ELFv2 functions
+  // can use r12 to establish their TOC on global entry.
+  Emit32(&cursor, Mfspr(0, 8));   // mflr r0
+  Emit32(&cursor, Bl(1));         // bl .+4
+  Emit32(&cursor, Mfspr(12, 8));  // mflr r12
+  Emit32(&cursor, Mtspr(8, 0));   // mtlr r0
+  const unsigned literal_offset = gp_count % 2 == 0 ? 24 : 20;
+  Emit32(&cursor, Ld(12, 12, literal_offset));
+  Emit32(&cursor, Mtspr(9, 12));  // mtctr r12
+  Emit32(&cursor, 0x4e800420);    // bctr
+  if (gp_count % 2 == 0) {
+    Emit32(&cursor, 0x60000000);  // nop; align literal to 8 bytes
+  }
   Emit64(&cursor, reinterpret_cast<uintptr_t>(target));
 
   const size_t written = reinterpret_cast<uint8_t*>(cursor) -