[A64] Skip software denormal flushing when FPCR.FZ handles inputs

Depends on hardware support, detected and runtime and controllable
via a64_extension_mask bit 1.

src/xenia/base/platform_arm64.cc

@@ -7,6 +7,10 @@
  ******************************************************************************
  */
 
+#include <cfenv>
+#include <cmath>
+#include <cstring>
+
 #include "xenia/base/cvar.h"
 #include "xenia/base/platform.h"
 #define XBYAK_NO_OP_NAMES
@@ -17,6 +21,7 @@ DEFINE_int64(a64_extension_mask, -1LL,
              "features.\n"
              "    0 = armv8.0\n"
              "    1 = Large System Extensions(LSE) atomic operations\n"
+             "    2 = FPCR.FZ flushes denormal inputs (skip software flush)\n"
              "   -1 = Detect and utilize all possible processor features\n",
              "a64");
 namespace xe {
@@ -43,6 +48,43 @@ void InitFeatureFlags() {
                       Xbyak_aarch64::util::XBYAK_AARCH64_HWCAP_ATOMIC);
 #undef TEST_EMIT_FEATURE
   }
+
+  // Detect whether FPCR.FZ flushes denormal float32 inputs to zero.
+  // The ARM spec says input flushing is implementation-defined.
+  // Modern cores (Cortex-A76+, Apple M1+) flush inputs; older ones may not.
+  if ((cvars::a64_extension_mask & kA64FZFlushesInputs) ==
+      kA64FZFlushesInputs) {
+    // Build a denormal float32: smallest positive denormal = 0x00000001.
+    uint32_t denorm_bits = 1;
+    float denorm;
+    std::memcpy(&denorm, &denorm_bits, 4);
+
+    // Save FPCR, enable FZ, add two denormals, check result.
+    uint64_t saved_fpcr;
+#if XE_COMPILER_MSVC
+    saved_fpcr = _ReadStatusReg(ARM64_FPCR);
+    _WriteStatusReg(ARM64_FPCR, saved_fpcr | (1ULL << 24));
+#else
+    asm volatile("mrs %0, fpcr" : "=r"(saved_fpcr));
+    uint64_t fz_fpcr = saved_fpcr | (1ULL << 24);
+    asm volatile("msr fpcr, %0" ::"r"(fz_fpcr));
+#endif
+
+    volatile float a = denorm;
+    volatile float b = denorm;
+    volatile float result = a + b;
+
+#if XE_COMPILER_MSVC
+    _WriteStatusReg(ARM64_FPCR, saved_fpcr);
+#else
+    asm volatile("msr fpcr, %0" ::"r"(saved_fpcr));
+#endif
+
+    if (result == 0.0f) {
+      feature_flags_ |= kA64FZFlushesInputs;
+    }
+  }
+
   g_feature_flags = feature_flags_;
   g_did_initialize_feature_flags = true;
 }

src/xenia/base/platform_arm64.h

@@ -15,7 +15,7 @@ namespace xe {
 namespace arm64 {
 enum A64FeatureFlags : uint64_t {
   kA64EmitLSE = 1 << 0,
-
+  kA64FZFlushesInputs = 1 << 1,
 };
 
 XE_NOALIAS