[turbofan] Don't assume that Word32AtomicPairBinops has a projection-0

The instruction selector assumed for Word32AtomicPairBinop nodes that if
there exists a Projection(1) user, then there also exists a
Projection(0) user. This, however, is not the case, because TurboFan
eliminates unreachable nodes. The missing projection node lead to a
failed DCHECK in the register allocator.

With this CL we allocate the right registers for the existing
projections, and allocate the other needed registers as temp registers.

R=gdeepti@chromium.org

Bug: v8:10140
Change-Id: I22331cae58f933e89dac6993fe3b21ff6502838a
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2011829Reviewed-by: Deepti Gandluri <gdeepti@chromium.org>
Commit-Queue: Andreas Haas <ahaas@chromium.org>
Cr-Commit-Position: refs/heads/master@{#65941}

src/compiler/backend/arm/code-generator-arm.cc

@@ -589,12 +589,20 @@ void AdjustStackPointerForTailCall(
 bool VerifyOutputOfAtomicPairInstr(ArmOperandConverter* converter,
                                    const Instruction* instr, Register low,
                                    Register high) {
-  if (instr->OutputCount() > 0) {
-    if (converter->OutputRegister(0) != low) return false;
-    if (instr->OutputCount() == 2 && converter->OutputRegister(1) != high)
-      return false;
-  }
-  return true;
+  DCHECK_GE(instr->OutputCount() + instr->TempCount(), 2);
+  if (instr->OutputCount() == 2) {
+    return (converter->OutputRegister(0) == low &&
+            converter->OutputRegister(1) == high);
+  }
+  if (instr->OutputCount() == 1) {
+    return (converter->OutputRegister(0) == low &&
+            converter->TempRegister(instr->TempCount() - 1) == high) ||
+           (converter->OutputRegister(0) == high &&
+            converter->TempRegister(instr->TempCount() - 1) == low);
+  }
+  DCHECK_EQ(instr->OutputCount(), 0);
+  return (converter->TempRegister(instr->TempCount() - 2) == low &&
+          converter->TempRegister(instr->TempCount() - 1) == high);
 }
 #endif
 

src/compiler/backend/arm/instruction-selector-arm.cc

@@ -427,33 +427,32 @@ void VisitPairAtomicBinOp(InstructionSelector* selector, Node* node,
   Node* value = node->InputAt(2);
   Node* value_high = node->InputAt(3);
   AddressingMode addressing_mode = kMode_Offset_RR;
+  InstructionCode code = opcode | AddressingModeField::encode(addressing_mode);
   InstructionOperand inputs[] = {
       g.UseUniqueRegister(value), g.UseUniqueRegister(value_high),
       g.UseUniqueRegister(base), g.UseUniqueRegister(index)};
+  InstructionOperand outputs[2];
+  size_t output_count = 0;
+  InstructionOperand temps[6];
+  size_t temp_count = 0;
+  temps[temp_count++] = g.TempRegister();
+  temps[temp_count++] = g.TempRegister(r6);
+  temps[temp_count++] = g.TempRegister(r7);
+  temps[temp_count++] = g.TempRegister();
   Node* projection0 = NodeProperties::FindProjection(node, 0);
   Node* projection1 = NodeProperties::FindProjection(node, 1);
-  InstructionCode code = opcode | AddressingModeField::encode(addressing_mode);
+  if (projection0) {
+    outputs[output_count++] = g.DefineAsFixed(projection0, r2);
+  } else {
+    temps[temp_count++] = g.TempRegister(r2);
+  }
   if (projection1) {
-    InstructionOperand outputs[] = {g.DefineAsFixed(projection0, r2),
-                                    g.DefineAsFixed(projection1, r3)};
-    InstructionOperand temps[] = {g.TempRegister(), g.TempRegister(r6),
-                                  g.TempRegister(r7), g.TempRegister()};
-    selector->Emit(code, arraysize(outputs), outputs, arraysize(inputs), inputs,
-                   arraysize(temps), temps);
-  } else if (projection0) {
-    InstructionOperand outputs[] = {g.DefineAsFixed(projection0, r2)};
-    InstructionOperand temps[] = {g.TempRegister(), g.TempRegister(r6),
-                                  g.TempRegister(r7), g.TempRegister(),
-                                  g.TempRegister(r3)};
-    selector->Emit(code, arraysize(outputs), outputs, arraysize(inputs), inputs,
-                   arraysize(temps), temps);
+    outputs[output_count++] = g.DefineAsFixed(projection1, r3);
   } else {
-    InstructionOperand temps[] = {g.TempRegister(),   g.TempRegister(r6),
-                                  g.TempRegister(r7), g.TempRegister(),
-                                  g.TempRegister(r2), g.TempRegister(r3)};
-    selector->Emit(code, 0, nullptr, arraysize(inputs), inputs,
-                   arraysize(temps), temps);
+    temps[temp_count++] = g.TempRegister(r3);
   }
+  selector->Emit(code, output_count, outputs, arraysize(inputs), inputs,
+                 temp_count, temps);
 }
 
 }  // namespace

src/compiler/backend/ia32/code-generator-ia32.cc

@@ -563,12 +563,19 @@ void AdjustStackPointerForTailCall(TurboAssembler* tasm,
 #ifdef DEBUG
 bool VerifyOutputOfAtomicPairInstr(IA32OperandConverter* converter,
                                    const Instruction* instr) {
-  if (instr->OutputCount() > 0) {
-    if (converter->OutputRegister(0) != eax) return false;
-    if (instr->OutputCount() == 2 && converter->OutputRegister(1) != edx)
-      return false;
-  }
-  return true;
+  if (instr->OutputCount() == 2) {
+    return (converter->OutputRegister(0) == eax &&
+            converter->OutputRegister(1) == edx);
+  }
+  if (instr->OutputCount() == 1) {
+    return (converter->OutputRegister(0) == eax &&
+            converter->TempRegister(0) == edx) ||
+           (converter->OutputRegister(0) == edx &&
+            converter->TempRegister(0) == eax);
+  }
+  DCHECK_EQ(instr->OutputCount(), 0);
+  return (converter->TempRegister(0) == eax &&
+          converter->TempRegister(1) == edx);
 }
 #endif
 

src/compiler/backend/ia32/instruction-selector-ia32.cc

@@ -1514,23 +1514,22 @@ void VisitPairAtomicBinOp(InstructionSelector* selector, Node* node,
   InstructionCode code = opcode | AddressingModeField::encode(addressing_mode);
   Node* projection0 = NodeProperties::FindProjection(node, 0);
   Node* projection1 = NodeProperties::FindProjection(node, 1);
+  InstructionOperand outputs[2];
+  size_t output_count = 0;
+  InstructionOperand temps[2];
+  size_t temp_count = 0;
+  if (projection0) {
+    outputs[output_count++] = g.DefineAsFixed(projection0, eax);
+  } else {
+    temps[temp_count++] = g.TempRegister(eax);
+  }
   if (projection1) {
-    InstructionOperand outputs[] = {g.DefineAsFixed(projection0, eax),
-                                    g.DefineAsFixed(projection1, edx)};
-    selector->Emit(code, arraysize(outputs), outputs, arraysize(inputs), inputs,
-                   0, {});
-  } else if (projection0) {
-    InstructionOperand outputs[] = {g.DefineAsFixed(projection0, eax)};
-    InstructionOperand temps[] = {g.TempRegister(edx)};
-    const int num_temps = arraysize(temps);
-    selector->Emit(code, arraysize(outputs), outputs, arraysize(inputs), inputs,
-                   num_temps, temps);
+    outputs[output_count++] = g.DefineAsFixed(projection1, edx);
   } else {
-    InstructionOperand temps[] = {g.TempRegister(eax), g.TempRegister(edx)};
-    const int num_temps = arraysize(temps);
-    selector->Emit(code, 0, nullptr, arraysize(inputs), inputs, num_temps,
-                   temps);
+    temps[temp_count++] = g.TempRegister(edx);
   }
+  selector->Emit(code, output_count, outputs, arraysize(inputs), inputs,
+                 temp_count, temps);
 }
 
 }  // namespace

test/cctest/wasm/test-run-wasm-atomics64.cc

@@ -723,6 +723,35 @@ WASM_EXEC_TEST(I64AtomicLoadUseOnlyHighWord) {
   CHECK_EQ(0x12345678, r.Call());
 }
 
+WASM_EXEC_TEST(I64AtomicAddUseOnlyLowWord) {
+  EXPERIMENTAL_FLAG_SCOPE(threads);
+  WasmRunner<uint32_t> r(execution_tier);
+  uint64_t* memory =
+      r.builder().AddMemoryElems<uint64_t>(kWasmPageSize / sizeof(uint64_t));
+  memory[1] = 0x1234567890abcdeful;
+  r.builder().SetHasSharedMemory();
+  // Test that we can use just the low word of an I64AtomicLoad.
+  BUILD(r, WASM_I32_CONVERT_I64(
+               WASM_ATOMICS_BINOP(kExprI64AtomicAdd, WASM_I32V(8), WASM_I64V(1),
+                                  MachineRepresentation::kWord64)));
+  CHECK_EQ(0x90abcdef, r.Call());
+}
+
+WASM_EXEC_TEST(I64AtomicAddUseOnlyHighWord) {
+  EXPERIMENTAL_FLAG_SCOPE(threads);
+  WasmRunner<uint32_t> r(execution_tier);
+  uint64_t* memory =
+      r.builder().AddMemoryElems<uint64_t>(kWasmPageSize / sizeof(uint64_t));
+  memory[1] = 0x1234567890abcdeful;
+  r.builder().SetHasSharedMemory();
+  // Test that we can use just the high word of an I64AtomicLoad.
+  BUILD(r, WASM_I32_CONVERT_I64(WASM_I64_ROR(
+               WASM_ATOMICS_BINOP(kExprI64AtomicAdd, WASM_I32V(8), WASM_I64V(1),
+                                  MachineRepresentation::kWord64),
+               WASM_I64V(32))));
+  CHECK_EQ(0x12345678, r.Call());
+}
+
 }  // namespace test_run_wasm_atomics_64
 }  // namespace wasm
 }  // namespace internal