From e765133481169b35521e3f2108a68dc8006e7741 Mon Sep 17 00:00:00 2001
From: Hannes Verschore <hv1989@gmail.com>
Date: Fri, 29 Jul 2016 16:53:49 +0200
Subject: [PATCH] Bug 1289054 - Part 16: Implement the 64bit variant of Clz and
 Ctz on arm, r=bbouvier

---
 js/src/jit/MacroAssembler.h             |  4 ++--
 js/src/jit/arm/CodeGenerator-arm.cpp    | 28 +++++++++++++++++++-----
 js/src/jit/arm/CodeGenerator-arm.h      |  2 ++
 js/src/jit/arm/MacroAssembler-arm-inl.h | 29 +++++++++++++++++++++++++
 4 files changed, 55 insertions(+), 8 deletions(-)

diff --git a/js/src/jit/MacroAssembler.h b/js/src/jit/MacroAssembler.h
index 702082ec2a14..3ba68c393aa9 100644
--- a/js/src/jit/MacroAssembler.h
+++ b/js/src/jit/MacroAssembler.h
@@ -909,8 +909,8 @@ class MacroAssembler : public MacroAssemblerSpecific
     inline void clz32(Register src, Register dest, bool knownNotZero) PER_SHARED_ARCH;
     inline void ctz32(Register src, Register dest, bool knownNotZero) PER_SHARED_ARCH;
 
-    inline void clz64(Register64 src, Register dest) DEFINED_ON(x86, x64);
-    inline void ctz64(Register64 src, Register dest) DEFINED_ON(x86, x64);
+    inline void clz64(Register64 src, Register dest) DEFINED_ON(x86, x64, arm);
+    inline void ctz64(Register64 src, Register dest) DEFINED_ON(x86, x64, arm);
 
     // On x86_shared, temp may be Invalid only if the chip has the POPCNT instruction.
     // On ARM, temp may never be Invalid.
diff --git a/js/src/jit/arm/CodeGenerator-arm.cpp b/js/src/jit/arm/CodeGenerator-arm.cpp
index f6c39af6e26d..149065be12fc 100644
--- a/js/src/jit/arm/CodeGenerator-arm.cpp
+++ b/js/src/jit/arm/CodeGenerator-arm.cpp
@@ -982,7 +982,7 @@ CodeGeneratorARM::visitClzI(LClzI* ins)
     Register input = ToRegister(ins->input());
     Register output = ToRegister(ins->output());
 
-    masm.ma_clz(input, output);
+    masm.clz32(input, output, /* knownNotZero = */ false);
 }
 
 void
@@ -990,12 +990,8 @@ CodeGeneratorARM::visitCtzI(LCtzI* ins)
 {
     Register input = ToRegister(ins->input());
     Register output = ToRegister(ins->output());
-    ScratchRegisterScope scratch(masm);
 
-    masm.ma_rsb(input, Imm32(0), scratch, SetCC);
-    masm.ma_and(input, scratch, input);
-    masm.ma_clz(input, output);
-    masm.ma_rsb(input, Imm32(0x1f), output, LeaveCC, Assembler::NotEqual);
+    masm.ctz32(input, output, /* knownNotZero = */ false);
 }
 
 void
@@ -3420,3 +3416,23 @@ CodeGeneratorARM::visitPopcntI64(LPopcntI64* lir)
     masm.popcnt64(input, output, temp);
 }
 
+void
+CodeGeneratorARM::visitClzI64(LClzI64* lir)
+{
+    Register64 input = ToRegister64(lir->getInt64Operand(0));
+    Register64 output = ToOutRegister64(lir);
+
+    masm.clz64(input, output.low);
+    masm.move32(Imm32(0), output.high);
+}
+
+void
+CodeGeneratorARM::visitCtzI64(LCtzI64* lir)
+{
+    Register64 input = ToRegister64(lir->getInt64Operand(0));
+    Register64 output = ToOutRegister64(lir);
+
+    masm.ctz64(input, output.low);
+    masm.move32(Imm32(0), output.high);
+}
+
diff --git a/js/src/jit/arm/CodeGenerator-arm.h b/js/src/jit/arm/CodeGenerator-arm.h
index 6eb0c04d085a..a000d67d7368 100644
--- a/js/src/jit/arm/CodeGenerator-arm.h
+++ b/js/src/jit/arm/CodeGenerator-arm.h
@@ -174,6 +174,8 @@ class CodeGeneratorARM : public CodeGeneratorShared
     virtual void visitAsmReinterpretFromI64(LAsmReinterpretFromI64* lir);
     virtual void visitAsmReinterpretToI64(LAsmReinterpretToI64* lir);
     virtual void visitPopcntI64(LPopcntI64* ins);
+    virtual void visitClzI64(LClzI64* ins);
+    virtual void visitCtzI64(LCtzI64* ins);
 
     // Out of line visitors.
     void visitOutOfLineBailout(OutOfLineBailout* ool);
diff --git a/js/src/jit/arm/MacroAssembler-arm-inl.h b/js/src/jit/arm/MacroAssembler-arm-inl.h
index 41f98892674b..66980c3e9a27 100644
--- a/js/src/jit/arm/MacroAssembler-arm-inl.h
+++ b/js/src/jit/arm/MacroAssembler-arm-inl.h
@@ -955,12 +955,41 @@ MacroAssembler::clz32(Register src, Register dest, bool knownNotZero)
     ma_clz(src, dest);
 }
 
+void
+MacroAssembler::clz64(Register64 src, Register dest)
+{
+    ScratchRegisterScope scratch(*this);
+    ma_clz(src.high, scratch);
+    ma_cmp(scratch, Imm32(32));
+    ma_mov(scratch, dest, LeaveCC, NotEqual);
+    ma_clz(src.low, dest, Equal);
+    ma_add(Imm32(32), dest, LeaveCC, Equal);
+}
+
 void
 MacroAssembler::ctz32(Register src, Register dest, bool knownNotZero)
 {
     ma_ctz(src, dest);
 }
 
+void
+MacroAssembler::ctz64(Register64 src, Register dest)
+{
+    Label done, high;
+
+    ma_cmp(src.low, Imm32(0));
+    ma_b(&high, Equal);
+
+    ctz32(src.low, dest, /* knownNotZero = */ true);
+    ma_b(&done);
+
+    bind(&high);
+    ctz32(src.high, dest, /* knownNotZero = */ false);
+    ma_add(Imm32(32), dest);
+
+    bind(&done);
+}
+
 void
 MacroAssembler::popcnt32(Register input,  Register output, Register tmp)
 {