Merge pull request #9695 from JosJuice/jitarm64-fres

JitArm64: Implement fres and frsqrte

4 files changed, 154 insertions, 2 deletions

diff --git a/Source/UnitTests/Core/CMakeLists.txt b/Source/UnitTests/Core/CMakeLists.txt
index 6dd344e638..134f7da78b 100644
--- a/Source/UnitTests/Core/CMakeLists.txt
+++ b/Source/UnitTests/Core/CMakeLists.txt
@@ -25,6 +25,8 @@ elseif(_M_ARM_64)
     PowerPC/DivUtilsTest.cpp
     PowerPC/JitArm64/ConvertSingleDouble.cpp
     PowerPC/JitArm64/FPRF.cpp
+    PowerPC/JitArm64/Fres.cpp
+    PowerPC/JitArm64/Frsqrte.cpp
     PowerPC/JitArm64/MovI2R.cpp
   )
 else()
diff --git a/Source/UnitTests/Core/PowerPC/JitArm64/Fres.cpp b/Source/UnitTests/Core/PowerPC/JitArm64/Fres.cpp
new file mode 100644
index 0000000000..027a3ac89f
--- /dev/null
+++ b/Source/UnitTests/Core/PowerPC/JitArm64/Fres.cpp
@@ -0,0 +1,66 @@
+// Copyright 2021 Dolphin Emulator Project
+// Licensed under GPLv2+
+// Refer to the license.txt file included.
+
+#include <functional>
+
+#include "Common/Arm64Emitter.h"
+#include "Common/BitUtils.h"
+#include "Common/CommonTypes.h"
+#include "Core/PowerPC/Interpreter/Interpreter_FPUtils.h"
+#include "Core/PowerPC/JitArm64/Jit.h"
+#include "Core/PowerPC/PowerPC.h"
+
+#include "../TestValues.h"
+
+#include <gtest/gtest.h>
+
+namespace
+{
+using namespace Arm64Gen;
+
+class TestFres : public JitArm64
+{
+public:
+  TestFres()
+  {
+    AllocCodeSpace(4096);
+
+    const u8* raw_fres = GetCodePtr();
+    GenerateFres();
+
+    fres = Common::BitCast<u64 (*)(u64)>(GetCodePtr());
+    MOV(ARM64Reg::X15, ARM64Reg::X30);
+    MOV(ARM64Reg::X14, PPC_REG);
+    MOVP2R(PPC_REG, &PowerPC::ppcState);
+    MOV(ARM64Reg::X1, ARM64Reg::X0);
+    m_float_emit.FMOV(ARM64Reg::D0, ARM64Reg::X0);
+    m_float_emit.FRECPE(ARM64Reg::D0, ARM64Reg::D0);
+    BL(raw_fres);
+    MOV(ARM64Reg::X30, ARM64Reg::X15);
+    MOV(PPC_REG, ARM64Reg::X14);
+    RET();
+  }
+
+  std::function<u64(u64)> fres;
+};
+
+}  // namespace
+
+TEST(JitArm64, Fres)
+{
+  TestFres test;
+
+  for (const u64 ivalue : double_test_values)
+  {
+    const double dvalue = Common::BitCast<double>(ivalue);
+
+    const u64 expected = Common::BitCast<u64>(Common::ApproximateReciprocal(dvalue));
+    const u64 actual = test.fres(ivalue);
+
+    if (expected != actual)
+      fmt::print("{:016x} -> {:016x} == {:016x}\n", ivalue, actual, expected);
+
+    EXPECT_EQ(expected, actual);
+  }
+}
diff --git a/Source/UnitTests/Core/PowerPC/JitArm64/Frsqrte.cpp b/Source/UnitTests/Core/PowerPC/JitArm64/Frsqrte.cpp
new file mode 100644
index 0000000000..749b147dcb
--- /dev/null
+++ b/Source/UnitTests/Core/PowerPC/JitArm64/Frsqrte.cpp
@@ -0,0 +1,66 @@
+// Copyright 2021 Dolphin Emulator Project
+// Licensed under GPLv2+
+// Refer to the license.txt file included.
+
+#include <functional>
+
+#include "Common/Arm64Emitter.h"
+#include "Common/BitUtils.h"
+#include "Common/CommonTypes.h"
+#include "Core/PowerPC/Interpreter/Interpreter_FPUtils.h"
+#include "Core/PowerPC/JitArm64/Jit.h"
+#include "Core/PowerPC/PowerPC.h"
+
+#include "../TestValues.h"
+
+#include <gtest/gtest.h>
+
+namespace
+{
+using namespace Arm64Gen;
+
+class TestFrsqrte : public JitArm64
+{
+public:
+  TestFrsqrte()
+  {
+    AllocCodeSpace(4096);
+
+    const u8* raw_frsqrte = GetCodePtr();
+    GenerateFrsqrte();
+
+    frsqrte = Common::BitCast<u64 (*)(u64)>(GetCodePtr());
+    MOV(ARM64Reg::X15, ARM64Reg::X30);
+    MOV(ARM64Reg::X14, PPC_REG);
+    MOVP2R(PPC_REG, &PowerPC::ppcState);
+    MOV(ARM64Reg::X1, ARM64Reg::X0);
+    m_float_emit.FMOV(ARM64Reg::D0, ARM64Reg::X0);
+    m_float_emit.FRSQRTE(ARM64Reg::D0, ARM64Reg::D0);
+    BL(raw_frsqrte);
+    MOV(ARM64Reg::X30, ARM64Reg::X15);
+    MOV(PPC_REG, ARM64Reg::X14);
+    RET();
+  }
+
+  std::function<u64(u64)> frsqrte;
+};
+
+}  // namespace
+
+TEST(JitArm64, Frsqrte)
+{
+  TestFrsqrte test;
+
+  for (const u64 ivalue : double_test_values)
+  {
+    const double dvalue = Common::BitCast<double>(ivalue);
+
+    const u64 expected = Common::BitCast<u64>(Common::ApproximateReciprocalSquareRoot(dvalue));
+    const u64 actual = test.frsqrte(ivalue);
+
+    if (expected != actual)
+      fmt::print("{:016x} -> {:016x} == {:016x}\n", ivalue, actual, expected);
+
+    EXPECT_EQ(expected, actual);
+  }
+}
diff --git a/Source/UnitTests/Core/PowerPC/TestValues.h b/Source/UnitTests/Core/PowerPC/TestValues.h
index 20fbbe913c..a1df5b0ff5 100644
--- a/Source/UnitTests/Core/PowerPC/TestValues.h
+++ b/Source/UnitTests/Core/PowerPC/TestValues.h
@@ -8,7 +8,7 @@
 
 #include "Common/CommonTypes.h"
 
-constexpr std::array<u64, 49> double_test_values{
+constexpr std::array<u64, 57> double_test_values{
     // Special values
     0x0000'0000'0000'0000,  // positive zero
     0x0000'0000'0000'0001,  // smallest positive denormal
@@ -54,13 +54,25 @@ constexpr std::array<u64, 49> double_test_values{
     0x3680'1234'5678'9ABC, 0x36A0'1234'5678'9ABC, 0x36B0'1234'5678'9ABC, 0xB680'1234'5678'9ABC,
     0xB6A0'1234'5678'9ABC, 0xB6B0'1234'5678'9ABC,
 
+    // (exp > 1148) Boundary case for fres
+    0x47C0'0000'0000'0000,  // 2^125 = fres result is non-zero
+    0x47D0'0000'0000'0000,  // 2^126 = fres result is zero
+    0xC7C0'0000'0000'0000,  // -2^125 = fres result is non-zero
+    0xC7D0'0000'0000'0000,  // -2^126 = fres result is zero
+
+    // (exp < 895) Boundary case for fres
+    0x37F0'0000'0000'0000,  // 2^(-128) = fres result is non-max
+    0x37E0'0000'0000'0000,  // 2^(-129) = fres result is max
+    0xB7F0'0000'0000'0000,  // -2^(-128) = fres result is non-max
+    0xB7E0'0000'0000'0000,  // -2^(-129) = fres result is max
+
     // Some typical numbers
     0x3FF8'0000'0000'0000,  // 1.5
     0x408F'4000'0000'0000,  // 1000
     0xC008'0000'0000'0000,  // -3
 };
 
-constexpr std::array<u32, 29> single_test_values{
+constexpr std::array<u32, 33> single_test_values{
     // Special values
     0x0000'0000,  // positive zero
     0x0000'0001,  // smallest positive denormal
@@ -89,6 +101,12 @@ constexpr std::array<u32, 29> single_test_values{
     0xFFC0'0000,  // first negative QNaN
     0xFFFF'FFFF,  // last negative QNaN
 
+    // (exp > 252) Boundary case for fres
+    0x7E00'0000,  // 2^125 = fres result is non-zero
+    0x7E80'0000,  // 2^126 = fres result is zero
+    0xC7C0'0000,  // -2^125 = fres result is non-zero
+    0xC7D0'0000,  // -2^126 = fres result is zero
+
     // Some typical numbers
     0x3FC0'0000,  // 1.5
     0x447A'0000,  // 1000