[wasm][arm][ia32] Push instance only twice in lazy-compile builtin

This change already landed for x64, now come arm and ia32. The code
already existed for arm64.

The wasm instance got pushed three times in the lazy-compile builtin:
1) as part of the parameters;
2) as a parameter for the runtime function;
3) to load the jump table address after the runtime function;

The third push can be avoided by loading the jump table address after
all parameters get loaded from the stack again.

R=clemensb@chromium.org

Bug: v8:13049, v8:12926
Change-Id: Ifdbe943520c031ec5c480798694bcacc490a64bc
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3764348Reviewed-by: Clemens Backes <clemensb@chromium.org>
Commit-Queue: Andreas Haas <ahaas@chromium.org>
Cr-Commit-Position: refs/heads/main@{#81782}

src/builtins/arm/builtins-arm.cc

@@ -2695,12 +2695,7 @@ void Builtins::Generate_WasmCompileLazy(MacroAssembler* masm) {
     __ stm(db_w, sp, gp_regs);
     __ vstm(db_w, sp, lowest_fp_reg, highest_fp_reg);
 
-    // Push the Wasm instance for loading the jump table address after the
-    // runtime call.
-    __ push(kWasmInstanceRegister);
-
-    // Push the Wasm instance again as an explicit argument to the runtime
-    // function.
+    // Push the Wasm instance as an explicit argument to the runtime function.
     __ push(kWasmInstanceRegister);
     // Push the function index as second argument.
     __ push(kWasmCompileLazyFuncIndexRegister);
@@ -2710,16 +2705,18 @@ void Builtins::Generate_WasmCompileLazy(MacroAssembler* masm) {
     __ CallRuntime(Runtime::kWasmCompileLazy, 2);
     // The runtime function returns the jump table slot offset as a Smi. Use
     // that to compute the jump target in r8.
-    __ pop(kWasmInstanceRegister);
-    __ ldr(r8, MemOperand(
-                   kWasmInstanceRegister,
-                   WasmInstanceObject::kJumpTableStartOffset - kHeapObjectTag));
-    __ add(r8, r8, Operand::SmiUntag(kReturnRegister0));
-    // r8 now holds the jump table slot where we want to jump to in the end.
+    __ mov(r8, Operand::SmiUntag(kReturnRegister0));
 
     // Restore registers.
     __ vldm(ia_w, sp, lowest_fp_reg, highest_fp_reg);
     __ ldm(ia_w, sp, gp_regs);
+
+    // After the instance register has been restored, we can add the jump table
+    // start to the jump table offset already stored in r8.
+    __ ldr(r9, MemOperand(
+                   kWasmInstanceRegister,
+                   WasmInstanceObject::kJumpTableStartOffset - kHeapObjectTag));
+    __ add(r8, r8, r9);
   }
 
   // Finally, jump to the jump table slot for the function.

src/builtins/ia32/builtins-ia32.cc

@@ -2931,12 +2931,7 @@ void Builtins::Generate_WasmCompileLazy(MacroAssembler* masm) {
       offset += kSimd128Size;
     }
 
-    // Push the Wasm instance for loading the jump table address after the
-    // runtime call.
-    __ Push(kWasmInstanceRegister);
-
-    // Push the Wasm instance again as an explicit argument to the runtime
-    // function.
+    // Push the Wasm instance as an explicit argument to the runtime function.
     __ Push(kWasmInstanceRegister);
     // Push the function index as second argument.
     __ Push(kWasmCompileLazyFuncIndexRegister);
@@ -2946,13 +2941,8 @@ void Builtins::Generate_WasmCompileLazy(MacroAssembler* masm) {
     __ CallRuntime(Runtime::kWasmCompileLazy, 2);
     // The runtime function returns the jump table slot offset as a Smi. Use
     // that to compute the jump target in edi.
-    __ Pop(kWasmInstanceRegister);
-    __ mov(edi, MemOperand(kWasmInstanceRegister,
-                           WasmInstanceObject::kJumpTableStartOffset -
-                               kHeapObjectTag));
     __ SmiUntag(kReturnRegister0);
-    __ add(edi, kReturnRegister0);
-    // edi now holds the jump table slot where we want to jump to in the end.
+    __ mov(edi, kReturnRegister0);
 
     // Restore registers.
     for (DoubleRegister reg : base::Reversed(wasm::kFpParamRegisters)) {
@@ -2964,6 +2954,12 @@ void Builtins::Generate_WasmCompileLazy(MacroAssembler* masm) {
     for (Register reg : base::Reversed(wasm::kGpParamRegisters)) {
       __ Pop(reg);
     }
+
+    // After the instance register has been restored, we can add the jump table
+    // start to the jump table offset already stored in edi.
+    __ add(edi, MemOperand(kWasmInstanceRegister,
+                           WasmInstanceObject::kJumpTableStartOffset -
+                               kHeapObjectTag));
   }
 
   // Finally, jump to the jump table slot for the function.