Bug 1518957 - Implement ARM64 truncation and remove unused emitRoundD…

…ouble(). r=nbp

--HG--
extra : amend_source : 6d7336b98d91c2f252431d38823731b394e30dd7

js/src/jit/arm/CodeGenerator-arm.cpp

@@ -1285,18 +1285,6 @@ void CodeGenerator::visitTruncF(LTruncF* lir) {
   bailoutFrom(&bail, lir->snapshot());
 }
 
-void CodeGeneratorARM::emitRoundDouble(FloatRegister src, Register dest,
-                                       Label* fail) {
-  ScratchDoubleScope scratch(masm);
-  ScratchRegisterScope scratchReg(masm);
-
-  masm.ma_vcvt_F64_I32(src, scratch);
-  masm.ma_vxfer(scratch, dest);
-  masm.ma_cmp(dest, Imm32(0x7fffffff), scratchReg);
-  masm.ma_cmp(dest, Imm32(0x80000000), scratchReg, Assembler::NotEqual);
-  masm.ma_b(fail, Assembler::Equal);
-}
-
 void CodeGenerator::visitTruncateDToInt32(LTruncateDToInt32* ins) {
   emitTruncateDouble(ToFloatRegister(ins->input()), ToRegister(ins->output()),
                      ins->mir());

js/src/jit/arm/CodeGenerator-arm.h

@@ -68,8 +68,6 @@ class CodeGeneratorARM : public CodeGeneratorShared {
 
   bool generateOutOfLineCode();
 
-  void emitRoundDouble(FloatRegister src, Register dest, Label* fail);
-
   // Emits a branch that directs control flow to the true block if |cond| is
   // true, and the false block if |cond| is false.
   void emitBranch(Assembler::Condition cond, MBasicBlock* ifTrue,

js/src/jit/arm64/CodeGenerator-arm64.cpp

@@ -1056,7 +1056,7 @@ void CodeGenerator::visitRoundF(LRoundF* lir) {
       masm.Fcmp(input32, 0.0f);
       bailoutIf(Assembler::Overflow, lir->snapshot());
 
-      // Move all 64 bits of the input into a scratch register to check for -0.
+      // Move all 32 bits of the input into a scratch register to check for -0.
       vixl::UseScratchRegisterScope temps(&masm.asVIXL());
       const ARMRegister scratchGPR32 = temps.AcquireW();
       masm.Fmov(scratchGPR32, input32);
@@ -1098,9 +1098,83 @@ void CodeGenerator::visitRoundF(LRoundF* lir) {
   masm.bind(&done);
 }
 
-void CodeGenerator::visitTrunc(LTrunc* lir) { MOZ_CRASH("visitTrunc"); }
+void CodeGenerator::visitTrunc(LTrunc* lir) {
+  const FloatRegister input = ToFloatRegister(lir->input());
+  const ARMFPRegister input64(input, 64);
+  const Register output = ToRegister(lir->output());
+  const ARMRegister output32(output, 32);
+
+  Label done, zeroCase;
 
-void CodeGenerator::visitTruncF(LTruncF* lir) { MOZ_CRASH("visitTruncF"); }
+  // Convert scalar to signed 32-bit fixed-point, rounding toward zero.
+  // In the case of overflow, the output is saturated.
+  // In the case of NaN and -0, the output is zero.
+  masm.Fcvtzs(output32, input64);
+
+  // If the output was zero, worry about special cases.
+  masm.branch32(Assembler::Equal, output, Imm32(0), &zeroCase);
+
+  // Bail on overflow cases.
+  bailoutCmp32(Assembler::Equal, output, Imm32(INT_MAX), lir->snapshot());
+  bailoutCmp32(Assembler::Equal, output, Imm32(INT_MIN), lir->snapshot());
+
+  // If the output was non-zero and wasn't saturated, just return it.
+  masm.jump(&done);
+
+  // Handle the case of a zero output:
+  // 1. The input may have been NaN, requiring a bail.
+  // 2. The input may have been in (-1,-0], requiring a bail.
+  {
+    masm.bind(&zeroCase);
+
+    // If input is a negative number that truncated to zero, the real
+    // output should be the non-integer -0.
+    // The use of "lt" instead of "lo" also catches unordered NaN input.
+    masm.Fcmp(input64, 0.0);
+    bailoutIf(vixl::lt, lir->snapshot());
+  }
+
+  masm.bind(&done);
+}
+
+void CodeGenerator::visitTruncF(LTruncF* lir) {
+  const FloatRegister input = ToFloatRegister(lir->input());
+  const ARMFPRegister input32(input, 32);
+  const Register output = ToRegister(lir->output());
+  const ARMRegister output32(output, 32);
+
+  Label done, zeroCase;
+
+  // Convert scalar to signed 32-bit fixed-point, rounding toward zero.
+  // In the case of overflow, the output is saturated.
+  // In the case of NaN and -0, the output is zero.
+  masm.Fcvtzs(output32, input32);
+
+  // If the output was zero, worry about special cases.
+  masm.branch32(Assembler::Equal, output, Imm32(0), &zeroCase);
+
+  // Bail on overflow cases.
+  bailoutCmp32(Assembler::Equal, output, Imm32(INT_MAX), lir->snapshot());
+  bailoutCmp32(Assembler::Equal, output, Imm32(INT_MIN), lir->snapshot());
+
+  // If the output was non-zero and wasn't saturated, just return it.
+  masm.jump(&done);
+
+  // Handle the case of a zero output:
+  // 1. The input may have been NaN, requiring a bail.
+  // 2. The input may have been in (-1,-0], requiring a bail.
+  {
+    masm.bind(&zeroCase);
+
+    // If input is a negative number that truncated to zero, the real
+    // output should be the non-integer -0.
+    // The use of "lt" instead of "lo" also catches unordered NaN input.
+    masm.Fcmp(input32, 0.0f);
+    bailoutIf(vixl::lt, lir->snapshot());
+  }
+
+  masm.bind(&done);
+}
 
 void CodeGenerator::visitClzI(LClzI* lir) {
   ARMRegister input = toWRegister(lir->input());
@@ -1114,11 +1188,6 @@ void CodeGenerator::visitCtzI(LCtzI* lir) {
   masm.ctz32(input, output, /* knownNotZero = */ false);
 }
 
-void CodeGeneratorARM64::emitRoundDouble(FloatRegister src, Register dest,
-                                         Label* fail) {
-  MOZ_CRASH("CodeGeneratorARM64::emitRoundDouble");
-}
-
 void CodeGenerator::visitTruncateDToInt32(LTruncateDToInt32* ins) {
   emitTruncateDouble(ToFloatRegister(ins->input()), ToRegister(ins->output()),
                      ins->mir());

js/src/jit/arm64/CodeGenerator-arm64.h

@@ -60,8 +60,6 @@ class CodeGeneratorARM64 : public CodeGeneratorShared {
 
   bool generateOutOfLineCode();
 
-  void emitRoundDouble(FloatRegister src, Register dest, Label* fail);
-
   // Emits a branch that directs control flow to the true block if |cond| is
   // true, and the false block if |cond| is false.
   void emitBranch(Assembler::Condition cond, MBasicBlock* ifTrue,