ARM: Make use of d16-d31 when available.

src/arm/assembler-arm.h

 // Copyright (c) 1994-2006 Sun Microsystems Inc.
 // All Rights Reserved.
 //
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions
 // are met:
 //
 // - Redistributions of source code must retain the above copyright notice,
 // this list of conditions and the following disclaimer.
 //
@@ skipping 29 matching lines @@
 #ifndef V8_ARM_ASSEMBLER_ARM_H_
 #define V8_ARM_ASSEMBLER_ARM_H_
 #include <stdio.h>
 #include "assembler.h"
 #include "constants-arm.h"
 #include "serialize.h"
 
 namespace v8 {
 namespace internal {
 
+// CpuFeatures keeps track of which features are supported by the target CPU.
+// Supported features must be enabled by a Scope before use.
+class CpuFeatures : public AllStatic {
+ public:
+  // Detect features of the target CPU. Set safe defaults if the serializer
+  // is enabled (snapshots must be portable).
+  static void Probe();
+
+  // Check whether a feature is supported by the target CPU.
+  static bool IsSupported(CpuFeature f) {
+    ASSERT(initialized_);
+    if (f == VFP3 && !FLAG_enable_vfp3) return false;
+    if (f == VFP2 && !FLAG_enable_vfp2) return false;
+    if (f == SUDIV && !FLAG_enable_sudiv) return false;
+    if (f == UNALIGNED_ACCESSES && !FLAG_enable_unaligned_accesses) {
+      return false;
+    }
+    return (supported_ & (1u << f)) != 0;
+  }
+
+#ifdef DEBUG
+  // Check whether a feature is currently enabled.
+  static bool IsEnabled(CpuFeature f) {
+    ASSERT(initialized_);
+    Isolate* isolate = Isolate::UncheckedCurrent();
+    if (isolate == NULL) {
+      // When no isolate is available, work as if we're running in
+      // release mode.
+      return IsSupported(f);
+    }
+    unsigned enabled = static_cast<unsigned>(isolate->enabled_cpu_features());
+    return (enabled & (1u << f)) != 0;
+  }
+#endif
+
+  // Enable a specified feature within a scope.
+  class Scope BASE_EMBEDDED {
+#ifdef DEBUG
+
+   public:
+    explicit Scope(CpuFeature f) {
+      unsigned mask = 1u << f;
+      // VFP2 and ARMv7 are implied by VFP3.
+      if (f == VFP3) mask |= 1u << VFP2 | 1u << ARMv7;
+      ASSERT(CpuFeatures::IsSupported(f));
+      ASSERT(!Serializer::enabled() ||
+             (CpuFeatures::found_by_runtime_probing_ & mask) == 0);
+      isolate_ = Isolate::UncheckedCurrent();
+      old_enabled_ = 0;
+      if (isolate_ != NULL) {
+        old_enabled_ = static_cast<unsigned>(isolate_->enabled_cpu_features());
+        isolate_->set_enabled_cpu_features(old_enabled_ | mask);
+      }
+    }
+    ~Scope() {
+      ASSERT_EQ(Isolate::UncheckedCurrent(), isolate_);
+      if (isolate_ != NULL) {
+        isolate_->set_enabled_cpu_features(old_enabled_);
+      }
+    }
+
+   private:
+    Isolate* isolate_;
+    unsigned old_enabled_;
+#else
+
+   public:
+    explicit Scope(CpuFeature f) {}
+#endif
+  };
+
+  class TryForceFeatureScope BASE_EMBEDDED {
+   public:
+    explicit TryForceFeatureScope(CpuFeature f)
+        : old_supported_(CpuFeatures::supported_) {
+      if (CanForce()) {
+        CpuFeatures::supported_ |= (1u << f);
+      }
+    }
+
+    ~TryForceFeatureScope() {
+      if (CanForce()) {
+        CpuFeatures::supported_ = old_supported_;
+      }
+    }
+
+   private:
+    static bool CanForce() {
+      // It's only safe to temporarily force support of CPU features
+      // when there's only a single isolate, which is guaranteed when
+      // the serializer is enabled.
+      return Serializer::enabled();
+    }
+
+    const unsigned old_supported_;
+  };
+
+ private:
+#ifdef DEBUG
+  static bool initialized_;
+#endif
+  static unsigned supported_;
+  static unsigned found_by_runtime_probing_;
+
+  DISALLOW_COPY_AND_ASSIGN(CpuFeatures);
+};
+
+
 // CPU Registers.
 //
 // 1) We would prefer to use an enum, but enum values are assignment-
 // compatible with int, which has caused code-generation bugs.
 //
 // 2) We would prefer to use a class instead of a struct but we don't like
 // the register initialization to depend on the particular initialization
 // order (which appears to be different on OS X, Linux, and Windows for the
 // installed versions of C++ we tried). Using a struct permits C-style
 // "initialization". Also, the Register objects cannot be const as this
@@ skipping 123 matching lines @@
     *m = code_ & 0x1;
     *vm = code_ >> 1;
   }
 
   int code_;
 };
 
 
 // Double word VFP register.
 struct DwVfpRegister {
-  static const int kNumRegisters = 16;
+  static const int kNumRegisters = 32;
   // A few double registers are reserved: one as a scratch register and one to
   // hold 0.0, that does not fit in the immediate field of vmov instructions.
   //  d14: 0.0
   //  d15: scratch register.
   static const int kNumReservedRegisters = 2;
-  static const int kNumAllocatableRegisters = kNumRegisters -
-      kNumReservedRegisters;
+  static int NumAvailableRegisters() {
+    return CpuFeatures::IsSupported(VFP32DREGS) ? 32 : 16;
+  }
+  static int NumAllocatableRegisters() {
+    return NumAvailableRegisters() - kNumReservedRegisters;
+  }
 
   inline static int ToAllocationIndex(DwVfpRegister reg);
 
   static DwVfpRegister FromAllocationIndex(int index) {
-    ASSERT(index >= 0 && index < kNumAllocatableRegisters);
+    ASSERT(index >= 0 && index < NumAllocatableRegisters());
+    if (index >= 14)

[ulan, 2012/12/17 10:44:52]

Here and below we can use kDoubleRegZero.code() instead of 14.

[hans, 2012/12/17 12:31:44]

I don't think that would work, since kDoubleRegZero hasn't been defined at this
point, and it expands to d14 which also hasn't been defined yet (and can't be
defined before this struct).

I guess I could just declare FromAllocationIndex here, and then put the
definition further below, when d14 and kDoubleRegZero are available. Would you
prefer that?

[ulan, 2012/12/17 13:03:36]

I see. Yep, putting FromAllocationIndex near the ToAllocationIndex would make it
more readable.

On 2012/12/17 12:31:44, hans wrote:

[hans, 2012/12/17 13:53:21]

Cool. I've moved it and AllocationIndexToString.

+      return from_code(index + kNumReservedRegisters);
     return from_code(index);
   }
 
   static const char* AllocationIndexToString(int index) {
-    ASSERT(index >= 0 && index < kNumAllocatableRegisters);
+    ASSERT(index >= 0 && index < NumAllocatableRegisters());
+    if (index >= 14)
+      index += kNumReservedRegisters;
     const char* const names[] = {
       "d0",
       "d1",
       "d2",
       "d3",
       "d4",
       "d5",
       "d6",
       "d7",
       "d8",
       "d9",
       "d10",
       "d11",
       "d12",
-      "d13"
+      "d13",
+      "d14",
+      "d15",
+      "d16",
+      "d17",
+      "d18",
+      "d19",
+      "d20",
+      "d21",
+      "d22",
+      "d23",
+      "d24",
+      "d25",
+      "d26",
+      "d27",
+      "d28",
+      "d29",
+      "d30",
+      "d31"
     };
     return names[index];
   }
 
   static DwVfpRegister from_code(int code) {
     DwVfpRegister r = { code };
     return r;
   }
 
-  // Supporting d0 to d15, can be later extended to d31.
-  bool is_valid() const { return 0 <= code_ && code_ < 16; }
+  bool is_valid() const {
+    return 0 <= code_ && code_ < NumAvailableRegisters();
+  }
   bool is(DwVfpRegister reg) const { return code_ == reg.code_; }
   SwVfpRegister low() const {
+    ASSERT(code_ < 16);
     SwVfpRegister reg;
     reg.code_ = code_ * 2;
 
     ASSERT(reg.is_valid());
     return reg;
   }
   SwVfpRegister high() const {
+    ASSERT(code_ < 16);
     SwVfpRegister reg;
     reg.code_ = (code_ * 2) + 1;
 
     ASSERT(reg.is_valid());
     return reg;
   }
   int code() const {
     ASSERT(is_valid());
     return code_;
   }
@@ skipping 59 matching lines @@
 const DwVfpRegister d6  = {  6 };
 const DwVfpRegister d7  = {  7 };
 const DwVfpRegister d8  = {  8 };
 const DwVfpRegister d9  = {  9 };
 const DwVfpRegister d10 = { 10 };
 const DwVfpRegister d11 = { 11 };
 const DwVfpRegister d12 = { 12 };
 const DwVfpRegister d13 = { 13 };
 const DwVfpRegister d14 = { 14 };
 const DwVfpRegister d15 = { 15 };
+const DwVfpRegister d16 = { 16 };
+const DwVfpRegister d17 = { 17 };
+const DwVfpRegister d18 = { 18 };
+const DwVfpRegister d19 = { 19 };
+const DwVfpRegister d20 = { 20 };
+const DwVfpRegister d21 = { 21 };
+const DwVfpRegister d22 = { 22 };
+const DwVfpRegister d23 = { 23 };
+const DwVfpRegister d24 = { 24 };
+const DwVfpRegister d25 = { 25 };
+const DwVfpRegister d26 = { 26 };
+const DwVfpRegister d27 = { 27 };
+const DwVfpRegister d28 = { 28 };
+const DwVfpRegister d29 = { 29 };
+const DwVfpRegister d30 = { 30 };
+const DwVfpRegister d31 = { 31 };
 
-// Aliases for double registers.  Defined using #define instead of
+// Aliases for double registers. Defined using #define instead of
 // "static const DwVfpRegister&" because Clang complains otherwise when a
 // compilation unit that includes this header doesn't use the variables.
 #define kFirstCalleeSavedDoubleReg d8
 #define kLastCalleeSavedDoubleReg d15
 #define kDoubleRegZero d14
 #define kScratchDoubleReg d15
 
 
 // Coprocessor register
 struct CRegister {
@@ skipping 154 matching lines @@
   Register rn_;  // base
   Register rm_;  // register offset
   int32_t offset_;  // valid if rm_ == no_reg
   ShiftOp shift_op_;
   int shift_imm_;  // valid if rm_ != no_reg && rs_ == no_reg
   AddrMode am_;  // bits P, U, and W
 
   friend class Assembler;
 };
 
-// CpuFeatures keeps track of which features are supported by the target CPU.
-// Supported features must be enabled by a Scope before use.
-class CpuFeatures : public AllStatic {
- public:
-  // Detect features of the target CPU. Set safe defaults if the serializer
-  // is enabled (snapshots must be portable).
-  static void Probe();
-
-  // Check whether a feature is supported by the target CPU.
-  static bool IsSupported(CpuFeature f) {
-    ASSERT(initialized_);
-    if (f == VFP3 && !FLAG_enable_vfp3) return false;
-    if (f == VFP2 && !FLAG_enable_vfp2) return false;
-    if (f == SUDIV && !FLAG_enable_sudiv) return false;
-    if (f == UNALIGNED_ACCESSES && !FLAG_enable_unaligned_accesses) {
-      return false;
-    }
-    return (supported_ & (1u << f)) != 0;
-  }
-
-#ifdef DEBUG
-  // Check whether a feature is currently enabled.
-  static bool IsEnabled(CpuFeature f) {
-    ASSERT(initialized_);
-    Isolate* isolate = Isolate::UncheckedCurrent();
-    if (isolate == NULL) {
-      // When no isolate is available, work as if we're running in
-      // release mode.
-      return IsSupported(f);
-    }
-    unsigned enabled = static_cast<unsigned>(isolate->enabled_cpu_features());
-    return (enabled & (1u << f)) != 0;
-  }
-#endif
-
-  // Enable a specified feature within a scope.
-  class Scope BASE_EMBEDDED {
-#ifdef DEBUG
-
-   public:
-    explicit Scope(CpuFeature f) {
-      unsigned mask = 1u << f;
-      // VFP2 and ARMv7 are implied by VFP3.
-      if (f == VFP3) mask |= 1u << VFP2 | 1u << ARMv7;
-      ASSERT(CpuFeatures::IsSupported(f));
-      ASSERT(!Serializer::enabled() ||
-             (CpuFeatures::found_by_runtime_probing_ & mask) == 0);
-      isolate_ = Isolate::UncheckedCurrent();
-      old_enabled_ = 0;
-      if (isolate_ != NULL) {
-        old_enabled_ = static_cast<unsigned>(isolate_->enabled_cpu_features());
-        isolate_->set_enabled_cpu_features(old_enabled_ | mask);
-      }
-    }
-    ~Scope() {
-      ASSERT_EQ(Isolate::UncheckedCurrent(), isolate_);
-      if (isolate_ != NULL) {
-        isolate_->set_enabled_cpu_features(old_enabled_);
-      }
-    }
-
-   private:
-    Isolate* isolate_;
-    unsigned old_enabled_;
-#else
-
-   public:
-    explicit Scope(CpuFeature f) {}
-#endif
-  };
-
-  class TryForceFeatureScope BASE_EMBEDDED {
-   public:
-    explicit TryForceFeatureScope(CpuFeature f)
-        : old_supported_(CpuFeatures::supported_) {
-      if (CanForce()) {
-        CpuFeatures::supported_ |= (1u << f);
-      }
-    }
-
-    ~TryForceFeatureScope() {
-      if (CanForce()) {
-        CpuFeatures::supported_ = old_supported_;
-      }
-    }
-
-   private:
-    static bool CanForce() {
-      // It's only safe to temporarily force support of CPU features
-      // when there's only a single isolate, which is guaranteed when
-      // the serializer is enabled.
-      return Serializer::enabled();
-    }
-
-    const unsigned old_supported_;
-  };
-
- private:
-#ifdef DEBUG
-  static bool initialized_;
-#endif
-  static unsigned supported_;
-  static unsigned found_by_runtime_probing_;
-
-  DISALLOW_COPY_AND_ASSIGN(CpuFeatures);
-};
-
-
 extern const Instr kMovLrPc;
 extern const Instr kLdrPCMask;
 extern const Instr kLdrPCPattern;
 extern const Instr kBlxRegMask;
 extern const Instr kBlxRegPattern;
 extern const Instr kBlxIp;
 
 extern const Instr kMovMvnMask;
 extern const Instr kMovMvnPattern;
 extern const Instr kMovMvnFlip;
@@ skipping 368 matching lines @@
            LFlag l = Short, Condition cond = al);
   void ldc(Coprocessor coproc, CRegister crd, Register base, int option,
            LFlag l = Short, Condition cond = al);
 
   void ldc2(Coprocessor coproc, CRegister crd, const MemOperand& src,
             LFlag l = Short);  // v5 and above
   void ldc2(Coprocessor coproc, CRegister crd, Register base, int option,
             LFlag l = Short);  // v5 and above
 
   // Support for VFP.
-  // All these APIs support S0 to S31 and D0 to D15.
-  // Currently these APIs do not support extended D registers, i.e, D16 to D31.
-  // However, some simple modifications can allow
-  // these APIs to support D16 to D31.
+  // All these APIs support S0 to S31 and D0 to D31.
 
   void vldr(const DwVfpRegister dst,
             const Register base,
             int offset,
             const Condition cond = al);
   void vldr(const DwVfpRegister dst,
             const MemOperand& src,
             const Condition cond = al);
 
   void vldr(const SwVfpRegister dst,
@@ skipping 48 matching lines @@
             double imm,
             const Register scratch = no_reg,
             const Condition cond = al);
   void vmov(const SwVfpRegister dst,
             const SwVfpRegister src,
             const Condition cond = al);
   void vmov(const DwVfpRegister dst,
             const DwVfpRegister src,
             const Condition cond = al);
   void vmov(const DwVfpRegister dst,
+            int x,

[ulan, 2012/12/17 10:44:52]

I wonder if there is more descriptive name than 'x'.

[hans, 2012/12/17 12:31:44]

How about "index"?

[ulan, 2012/12/17 13:03:36]

Yep, I like it more.

[hans, 2012/12/17 13:53:21]

Done.

+            const Register src,
+            const Condition cond = al);
+  void vmov(const DwVfpRegister dst,
             const Register src1,
             const Register src2,
             const Condition cond = al);
   void vmov(const Register dst1,
             const Register dst2,
             const DwVfpRegister src,
             const Condition cond = al);
   void vmov(const SwVfpRegister dst,
             const Register src,
             const Condition cond = al);
@@ skipping 409 matching lines @@
  public:
   explicit EnsureSpace(Assembler* assembler) {
     assembler->CheckBuffer();
   }
 };
 
 
 } }  // namespace v8::internal
 
 #endif  // V8_ARM_ASSEMBLER_ARM_H_