[x64] Transform setcc + movzxbl into xorl + setcc

xorl + setcc is more efficient than setcc + movzxbl and encodes one byte
shorter.

Change-Id: Ib6679ce9ab0ac0d34701daba5c3d4d8bb57a8fc6
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3492946Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Commit-Queue: Hao A Xu <hao.a.xu@intel.com>
Cr-Commit-Position: refs/heads/main@{#79377}

src/compiler/backend/x64/code-generator-x64.cc

@@ -1175,12 +1175,37 @@ void CodeGenerator::BailoutIfDeoptimized() {
           RelocInfo::CODE_TARGET, not_zero);
 }
 
+bool ShouldClearOutputRegisterBeforeInstruction(CodeGenerator* g,
+                                                Instruction* instr) {
+  X64OperandConverter i(g, instr);
+  FlagsMode mode = FlagsModeField::decode(instr->opcode());
+  if (mode == kFlags_set) {
+    FlagsCondition condition = FlagsConditionField::decode(instr->opcode());
+    if (condition != kUnorderedEqual && condition != kUnorderedNotEqual) {
+      Register reg = i.OutputRegister(instr->OutputCount() - 1);
+      // Do not clear output register when it is also input register.
+      for (size_t index = 0; index < instr->InputCount(); ++index) {
+        if (HasRegisterInput(instr, index) && reg == i.InputRegister(index))
+          return false;
+      }
+      return true;
+    }
+  }
+  return false;
+}
+
 // Assembles an instruction after register allocation, producing machine code.
 CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
     Instruction* instr) {
   X64OperandConverter i(this, instr);
   InstructionCode opcode = instr->opcode();
   ArchOpcode arch_opcode = ArchOpcodeField::decode(opcode);
+  if (ShouldClearOutputRegisterBeforeInstruction(this, instr)) {
+    // Transform setcc + movzxbl into xorl + setcc to avoid register stall and
+    // encode one byte shorter.
+    Register reg = i.OutputRegister(instr->OutputCount() - 1);
+    __ xorl(reg, reg);
+  }
   switch (arch_opcode) {
     case kArchCallCodeObject: {
       if (HasImmediateInput(instr, 0)) {
@@ -4473,7 +4498,9 @@ void CodeGenerator::AssembleArchBoolean(Instruction* instr,
   }
   __ bind(&check);
   __ setcc(FlagsConditionToCondition(condition), reg);
-  __ movzxbl(reg, reg);
+  if (!ShouldClearOutputRegisterBeforeInstruction(this, instr)) {
+    __ movzxbl(reg, reg);
+  }
   __ bind(&done);
 }