[turbofan][wasm][arm] Improved saturated conversions float32 to int32.

Bug: v8:10720
Change-Id: I7a05bfb3c87c4f0f5516608da5d42fdaff466536
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2304572
Commit-Queue: Richard Stotz <rstz@chromium.org>
Reviewed-by: Georg Neis <neis@chromium.org>
Reviewed-by: Andreas Haas <ahaas@chromium.org>
Cr-Commit-Position: refs/heads/master@{#69001}

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

@@ -1566,10 +1566,13 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       SwVfpRegister scratch = temps.AcquireS();
       __ vcvt_s32_f32(scratch, i.InputFloatRegister(0));
       __ vmov(i.OutputRegister(), scratch);
-      // Avoid INT32_MAX as an overflow indicator and use INT32_MIN instead,
-      // because INT32_MIN allows easier out-of-bounds detection.
-      __ cmn(i.OutputRegister(), Operand(1));
-      __ mov(i.OutputRegister(), Operand(INT32_MIN), SBit::LeaveCC, vs);
+      bool set_overflow_to_min_i32 = MiscField::decode(instr->opcode());
+      if (set_overflow_to_min_i32) {
+        // Avoid INT32_MAX as an overflow indicator and use INT32_MIN instead,
+        // because INT32_MIN allows easier out-of-bounds detection.
+        __ cmn(i.OutputRegister(), Operand(1));
+        __ mov(i.OutputRegister(), Operand(INT32_MIN), SBit::LeaveCC, vs);
+      }
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
@@ -1578,10 +1581,13 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       SwVfpRegister scratch = temps.AcquireS();
       __ vcvt_u32_f32(scratch, i.InputFloatRegister(0));
       __ vmov(i.OutputRegister(), scratch);
-      // Avoid UINT32_MAX as an overflow indicator and use 0 instead,
-      // because 0 allows easier out-of-bounds detection.
-      __ cmn(i.OutputRegister(), Operand(1));
-      __ adc(i.OutputRegister(), i.OutputRegister(), Operand::Zero());
+      bool set_overflow_to_min_u32 = MiscField::decode(instr->opcode());
+      if (set_overflow_to_min_u32) {
+        // Avoid UINT32_MAX as an overflow indicator and use 0 instead,
+        // because 0 allows easier out-of-bounds detection.
+        __ cmn(i.OutputRegister(), Operand(1));
+        __ adc(i.OutputRegister(), i.OutputRegister(), Operand::Zero());
+      }
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }

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

@@ -1471,8 +1471,6 @@ void InstructionSelector::VisitUint32Mod(Node* node) {
   V(RoundUint32ToFloat32, kArmVcvtF32U32)            \
   V(ChangeInt32ToFloat64, kArmVcvtF64S32)            \
   V(ChangeUint32ToFloat64, kArmVcvtF64U32)           \
-  V(TruncateFloat32ToInt32, kArmVcvtS32F32)          \
-  V(TruncateFloat32ToUint32, kArmVcvtU32F32)         \
   V(ChangeFloat64ToInt32, kArmVcvtS32F64)            \
   V(ChangeFloat64ToUint32, kArmVcvtU32F64)           \
   V(TruncateFloat64ToUint32, kArmVcvtU32F64)         \
@@ -3017,6 +3015,30 @@ void InstructionSelector::VisitF64x2Pmax(Node* node) {
   VisitF64x2PminOrPMax(this, kArmF64x2Pmax, node);
 }
 
+void InstructionSelector::VisitTruncateFloat32ToInt32(Node* node) {
+  ArmOperandGenerator g(this);
+
+  InstructionCode opcode = kArmVcvtS32F32;
+  TruncateKind kind = OpParameter<TruncateKind>(node->op());
+  if (kind == TruncateKind::kSetOverflowToMin) {
+    opcode |= MiscField::encode(true);
+  }
+
+  Emit(opcode, g.DefineAsRegister(node), g.UseRegister(node->InputAt(0)));
+}
+
+void InstructionSelector::VisitTruncateFloat32ToUint32(Node* node) {
+  ArmOperandGenerator g(this);
+
+  InstructionCode opcode = kArmVcvtU32F32;
+  TruncateKind kind = OpParameter<TruncateKind>(node->op());
+  if (kind == TruncateKind::kSetOverflowToMin) {
+    opcode |= MiscField::encode(true);
+  }
+
+  Emit(opcode, g.DefineAsRegister(node), g.UseRegister(node->InputAt(0)));
+}
+
 // static
 MachineOperatorBuilder::Flags
 InstructionSelector::SupportedMachineOperatorFlags() {
@@ -3041,6 +3063,7 @@ InstructionSelector::SupportedMachineOperatorFlags() {
              MachineOperatorBuilder::kFloat32RoundTiesEven |
              MachineOperatorBuilder::kFloat64RoundTiesEven;
   }
+  flags |= MachineOperatorBuilder::kSatConversionIsSafe;
   return flags;
 }
 

test/cctest/compiler/test-run-machops.cc

@@ -4192,14 +4192,14 @@ TEST(RunTruncateFloat32ToInt32) {
       } else if (i >= upper_bound) {
 #if V8_TARGET_ARCH_IA32 || V8_TARGET_ARCH_X64
         CHECK_FLOAT_EQ(std::numeric_limits<int32_t>::min(), m.Call(i));
-#elif V8_TARGET_ARCH_ARM64
+#elif V8_TARGET_ARCH_ARM64 || V8_TARGET_ARCH_ARM
         CHECK_FLOAT_EQ(std::numeric_limits<int32_t>::max(), m.Call(i));
 #endif
       } else {
         DCHECK(std::isnan(i));
 #if V8_TARGET_ARCH_IA32 || V8_TARGET_ARCH_X64
         CHECK_FLOAT_EQ(std::numeric_limits<int32_t>::min(), m.Call(i));
-#elif V8_TARGET_ARCH_ARM64
+#elif V8_TARGET_ARCH_ARM64 || V8_TARGET_ARCH_ARM
         CHECK_FLOAT_EQ(0, m.Call(i));
 #endif
       }
@@ -4218,7 +4218,7 @@ TEST(RunTruncateFloat32ToInt32) {
         DCHECK(std::isnan(i));
 #if V8_TARGET_ARCH_IA32 || V8_TARGET_ARCH_X64
         CHECK_FLOAT_EQ(std::numeric_limits<int32_t>::min(), m.Call(i));
-#elif V8_TARGET_ARCH_ARM64
+#elif V8_TARGET_ARCH_ARM64 || V8_TARGET_ARCH_ARM
         CHECK_FLOAT_EQ(0, m.Call(i));
 #endif
       }