Port of optimized ICs for external and pixel arrays from ia32 to ARM.

src/arm/assembler-arm.cc

Index: src/arm/assembler-arm.cc
===================================================================
--- src/arm/assembler-arm.cc	(revision 4136)
+++ src/arm/assembler-arm.cc	(working copy)
@@ -1313,11 +1313,28 @@
   // Vdst(15-12) | 1011(11-8) | offset
   ASSERT(CpuFeatures::IsEnabled(VFP3));
   ASSERT(offset % 4 == 0);
+  ASSERT((offset / 4) < 256);
   emit(cond | 0xD9*B20 | base.code()*B16 | dst.code()*B12 |
        0xB*B8 | ((offset / 4) & 255));
 }
 
 
+void Assembler::vldr(const SwVfpRegister dst,
+                     const Register base,
+                     int offset,
+                     const Condition cond) {
+  // Sdst = MEM(Rbase + offset).
+  // Instruction details available in ARM DDI 0406A, A8-628.
+  // cond(31-28) | 1101(27-24)| 1001(23-20) | Rbase(19-16) |
+  // Vdst(15-12) | 1010(11-8) | offset
+  ASSERT(CpuFeatures::IsEnabled(VFP3));
+  ASSERT(offset % 4 == 0);
+  ASSERT((offset / 4) < 256);
+  emit(cond | 0xD9*B20 | base.code()*B16 | dst.code()*B12 |
+       0xA*B8 | ((offset / 4) & 255));
+}
+
+
 void Assembler::vstr(const DwVfpRegister src,
                      const Register base,
                      int offset,
@@ -1328,6 +1345,7 @@
   // Vsrc(15-12) | 1011(11-8) | (offset/4)
   ASSERT(CpuFeatures::IsEnabled(VFP3));
   ASSERT(offset % 4 == 0);
+  ASSERT((offset / 4) < 256);
   emit(cond | 0xD8*B20 | base.code()*B16 | src.code()*B12 |
        0xB*B8 | ((offset / 4) & 255));
 }
@@ -1391,34 +1409,175 @@
 }
 
 
-void Assembler::vcvt(const DwVfpRegister dst,
-                     const SwVfpRegister src,
-                     const Condition cond) {
-  // Dd = Sm (integer in Sm converted to IEEE 64-bit doubles in Dd).
-  // Instruction details available in ARM DDI 0406A, A8-576.
-  // cond(31-28) | 11101(27-23)| D=?(22) | 11(21-20) | 1(19) | opc2=000(18-16) |
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | op(7)=1 | 1(6) | M=?(5) | 0(4) | Vm(3-0)
+// Type of data to read from or write to VFP register.
+// Used as specifier in generic vcvt instruction.
+enum VFPType { S32, U32, F32, F64 };
+
+
+static bool IsSignedVFPType(VFPType type) {
+  switch (type) {
+    case S32:
+      return true;
+    case U32:
+      return false;
+    default:
+      UNREACHABLE();
+      return false;
+  }
+}
+
+
+static bool IsIntegerVFPType(VFPType type) {
+  switch (type) {
+    case S32:
+    case U32:
+      return true;
+    case F32:
+    case F64:
+      return false;
+    default:
+      UNREACHABLE();
+      return false;
+  }
+}
+
+
+static bool IsDoubleVFPType(VFPType type) {
+  switch (type) {
+    case F32:
+      return false;
+    case F64:
+      return true;
+    default:
+      UNREACHABLE();
+      return false;
+  }
+}
+
+
+// Depending on split_last_bit split binary representation of reg_code into Vm:M
+// or M:Vm form (where M is single bit).
+static void SplitRegCode(bool split_last_bit,
+                         int reg_code,
+                         int* vm,
+                         int* m) {
+  if (split_last_bit) {
+    *m  = reg_code & 0x1;
+    *vm = reg_code >> 1;
+  } else {
+    *m  = (reg_code & 0x10) >> 4;
+    *vm = reg_code & 0x0F;
+  }
+}
+
+
+// Encode vcvt.src_type.dst_type instruction.
+static Instr EncodeVCVT(const VFPType dst_type,
+                        const int dst_code,
+                        const VFPType src_type,
+                        const int src_code,
+                        const Condition cond) {
+  if (IsIntegerVFPType(dst_type) || IsIntegerVFPType(src_type)) {
+    // Conversion between IEEE floating point and 32-bit integer.
+    // Instruction details available in ARM DDI 0406B, A8.6.295.
+    // cond(31-28) | 11101(27-23)| D(22) | 11(21-20) | 1(19) | opc2(18-16) |
+    // Vd(15-12) | 101(11-9) | sz(8) | op(7) | 1(6) | M(5) | 0(4) | Vm(3-0)
+    ASSERT(!IsIntegerVFPType(dst_type) || !IsIntegerVFPType(src_type));
+
+    int sz, opc2, D, Vd, M, Vm, op;
+
+    if (IsIntegerVFPType(dst_type)) {
+      opc2 = IsSignedVFPType(dst_type) ? 0x5 : 0x4;
+      sz = IsDoubleVFPType(src_type) ? 0x1 : 0x0;
+      op = 1;  // round towards zero
+      SplitRegCode(!IsDoubleVFPType(src_type), src_code, &Vm, &M);
+      SplitRegCode(true, dst_code, &Vd, &D);
+    } else {
+      ASSERT(IsIntegerVFPType(src_type));
+
+      opc2 = 0x0;
+      sz = IsDoubleVFPType(dst_type) ? 0x1 : 0x0;
+      op = IsSignedVFPType(src_type) ? 0x1 : 0x0;
+      SplitRegCode(true, src_code, &Vm, &M);
+      SplitRegCode(!IsDoubleVFPType(dst_type), dst_code, &Vd, &D);
+    }
+
+    return (cond | 0xE*B24 | B23 | D*B22 | 0x3*B20 | B19 | opc2*B16 |
+            Vd*B12 | 0x5*B9 | sz*B8 | op*B7 | B6 | M*B5 | Vm);
+  } else {
+    // Conversion between IEEE double and single precision.
+    // Instruction details available in ARM DDI 0406B, A8.6.298.
+    // cond(31-28) | 11101(27-23)| D(22) | 11(21-20) | 0111(19-16) |
+    // Vd(15-12) | 101(11-9) | sz(8) | 1(7) | 1(6) | M(5) | 0(4) | Vm(3-0)
+    int sz, D, Vd, M, Vm;
+
+    ASSERT(IsDoubleVFPType(dst_type) != IsDoubleVFPType(src_type));
+    sz = IsDoubleVFPType(src_type) ? 0x1 : 0x0;
+    SplitRegCode(IsDoubleVFPType(src_type), dst_code, &Vd, &D);
+    SplitRegCode(!IsDoubleVFPType(src_type), src_code, &Vm, &M);
+
+    return (cond | 0xE*B24 | B23 | D*B22 | 0x3*B20 | 0x7*B16 |
+            Vd*B12 | 0x5*B9 | sz*B8 | B7 | B6 | M*B5 | Vm);
+  }
+}
+
+
+void Assembler::vcvt_f64_s32(const DwVfpRegister dst,
+                             const SwVfpRegister src,
+                             const Condition cond) {
   ASSERT(CpuFeatures::IsEnabled(VFP3));
-  emit(cond | 0xE*B24 | B23 | 0x3*B20 | B19 |
-       dst.code()*B12 | 0x5*B9 | B8 | B7 | B6 |
-       (0x1 & src.code())*B5 | (src.code() >> 1));
+  emit(EncodeVCVT(F64, dst.code(), S32, src.code(), cond));
 }
 
 
-void Assembler::vcvt(const SwVfpRegister dst,
-                     const DwVfpRegister src,
-                     const Condition cond) {
-  // Sd = Dm (IEEE 64-bit doubles in Dm converted to 32 bit integer in Sd).
-  // Instruction details available in ARM DDI 0406A, A8-576.
-  // cond(31-28) | 11101(27-23)| D=?(22) | 11(21-20) | 1(19) | opc2=101(18-16)|
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | op(7)=? | 1(6) | M=?(5) | 0(4) | Vm(3-0)
+void Assembler::vcvt_f32_s32(const SwVfpRegister dst,
+                             const SwVfpRegister src,
+                             const Condition cond) {
   ASSERT(CpuFeatures::IsEnabled(VFP3));
-  emit(cond | 0xE*B24 | B23 |(0x1 & dst.code())*B22 |
-       0x3*B20 | B19 | 0x5*B16 | (dst.code() >> 1)*B12 |
-       0x5*B9 | B8 | B7 | B6 | src.code());
+  emit(EncodeVCVT(F32, dst.code(), S32, src.code(), cond));
 }
 
 
+void Assembler::vcvt_f64_u32(const DwVfpRegister dst,
+                             const SwVfpRegister src,
+                             const Condition cond) {
+  ASSERT(CpuFeatures::IsEnabled(VFP3));
+  emit(EncodeVCVT(F64, dst.code(), U32, src.code(), cond));
+}
+
+
+void Assembler::vcvt_s32_f64(const SwVfpRegister dst,
+                             const DwVfpRegister src,
+                             const Condition cond) {
+  ASSERT(CpuFeatures::IsEnabled(VFP3));
+  emit(EncodeVCVT(S32, dst.code(), F64, src.code(), cond));
+}
+
+
+void Assembler::vcvt_u32_f64(const SwVfpRegister dst,
+                             const DwVfpRegister src,
+                             const Condition cond) {
+  ASSERT(CpuFeatures::IsEnabled(VFP3));
+  emit(EncodeVCVT(U32, dst.code(), F64, src.code(), cond));
+}
+
+
+void Assembler::vcvt_f64_f32(const DwVfpRegister dst,
+                             const SwVfpRegister src,
+                             const Condition cond) {
+  ASSERT(CpuFeatures::IsEnabled(VFP3));
+  emit(EncodeVCVT(F64, dst.code(), F32, src.code(), cond));
+}
+
+
+void Assembler::vcvt_f32_f64(const SwVfpRegister dst,
+                             const DwVfpRegister src,
+                             const Condition cond) {
+  ASSERT(CpuFeatures::IsEnabled(VFP3));
+  emit(EncodeVCVT(F32, dst.code(), F64, src.code(), cond));
+}
+
+
 void Assembler::vadd(const DwVfpRegister dst,
                      const DwVfpRegister src1,
                      const DwVfpRegister src2,