Implement FMADD in terms of std::fma

Author: mrexodia

include/remill/Arch/Runtime/sysroot/cmath

@@ -80,6 +80,19 @@ constexpr long double sqrt(long double num) {
   return __builtin_sqrtl(num);
 }
 
+// https://en.cppreference.com/w/cpp/numeric/math/fma.html
+constexpr float fma(float x, float y, float z) {
+  return __builtin_fmaf(x, y, z);
+}
+
+constexpr double fma(double x, double y, double z) {
+  return __builtin_fma(x, y, z);
+}
+
+constexpr long double fma(long double x, long double y, long double z) {
+  return __builtin_fmal(x, y, z);
+}
+
 // https://en.cppreference.com/w/cpp/numeric/math/isunordered
 constexpr bool isunordered(float x, float y) {
   return __builtin_isunordered(x, y);

lib/Arch/AArch64/Semantics/BINARY.cpp

@@ -274,8 +274,7 @@ DEF_SEM(FMADD_S, V128W dst, V32 src1, V32 src2, V32 src3) {
 
   __remill_fpu_exception_clear(kFPUExceptionAll);
   BarrierReorder();
-  auto prod = FMul32(factor1, factor2);
-  auto res = FAdd32(prod, add);
+  auto res = std::fma(factor1, factor2, add);
   BarrierReorder();
   auto new_except = __remill_fpu_exception_test(kFPUExceptionAll);
   BarrierReorder();
@@ -301,8 +300,7 @@ DEF_SEM(FMADD_D, V128W dst, V64 src1, V64 src2, V64 src3) {
 
   __remill_fpu_exception_clear(kFPUExceptionAll);
   BarrierReorder();
-  auto prod = FMul64(factor1, factor2);
-  auto res = FAdd64(prod, add);
+  auto res = std::fma(factor1, factor2, add);
   BarrierReorder();
   auto except_new = __remill_fpu_exception_test(kFPUExceptionAll);
   SetFPSRStatusFlags(state, except_new);

tests/AArch64/Run.cpp

@@ -708,6 +708,11 @@ static void RunWithFlags(const test::TestInfo *info, NZCV flags,
   auto lifted_state = reinterpret_cast<State *>(&gLiftedState);
   auto native_state = reinterpret_cast<State *>(&gNativeState);
 
+  // CRITICAL: Establish consistent FPU environment for both tests
+  std::fesetenv(FE_DFL_ENV);
+  std::feclearexcept(FE_ALL_EXCEPT);
+  std::fesetround(FE_TONEAREST);  // Explicit rounding mode
+
   // Set up the run's info.
   gTestToRun = info->test_begin;
   gStackSwitcher = &(gLiftedStack._redzone2[0]);
@@ -719,6 +724,12 @@ static void RunWithFlags(const test::TestInfo *info, NZCV flags,
   auto native_test_faulted = false;
   if (!sigsetjmp(gJmpBuf, true)) {
     gInNativeTest = true;
+
+    // Reset FPU environment before native test
+    std::fesetenv(FE_DFL_ENV);
+    std::feclearexcept(FE_ALL_EXCEPT);
+    std::fesetround(FE_TONEAREST);
+
     asm("msr nzcv, %0" : : "r"(flags));
     InvokeTestCase(arg1, arg2, arg3);
   } else {
@@ -740,7 +751,11 @@ static void RunWithFlags(const test::TestInfo *info, NZCV flags,
   // native program state recorded before executing the native testcase,
   // but after swapping execution to operate on `gStack`.
   if (!sigsetjmp(gJmpBuf, true)) {
+    // Reset FPU environment before lifted test (same as native)
     std::fesetenv(FE_DFL_ENV);
+    std::feclearexcept(FE_ALL_EXCEPT);
+    std::fesetround(FE_TONEAREST);
+
     gInNativeTest = false;
     (void) lifted_func(*lifted_state, lifted_state->gpr.pc.aword, nullptr);
   } else {