test: Avoid -ffinite-math-only on floating point comparisons

With `-ffinite-math-only` (implied by `-ffast-math` and -Ofast),
GCC considers all comparisons against infinities to be false
and can optimize them away. This breaks several tests that rely
on infinities to be compared correctly. To avoid this, we add
GCC-specific optimize attribute that disables `-ffinite-math-only`
optimization for floating point comarisions used in assertions.

Signed-off-by: Ivan A. Melnikov <iv@altlinux.org>

Author: iv-m

test/test.h

@@ -126,6 +126,10 @@ simde_test_debug_printf_(const char* format, ...) {
 HEDLEY_DIAGNOSTIC_PUSH
 SIMDE_DIAGNOSTIC_DISABLE_FLOAT_EQUAL_
 
+#if defined(SIMDE_FAST_MATH) && defined(HEDLEY_GCC_VERSION) && HEDLEY_GCC_VERSION_CHECK(4,4,0)
+__attribute__((optimize("-fno-finite-math-only")))
+__attribute__((noinline))
+#endif
 static int
 simde_test_equal_f32(simde_float32 a, simde_float32 b, simde_float32 slop) {
   if (simde_math_isnan(a)) {
@@ -156,6 +160,10 @@ simde_test_equal_f16(simde_float16 a, simde_float16 b, simde_float16 slop) {
   return simde_test_equal_f32(af, bf, slopf);
 }
 
+#if defined(SIMDE_FAST_MATH) && defined(HEDLEY_GCC_VERSION) && HEDLEY_GCC_VERSION_CHECK(4,4,0)
+__attribute__((optimize("-fno-finite-math-only")))
+__attribute__((noinline))
+#endif
 static int
 simde_test_equal_f64(simde_float64 a, simde_float64 b, simde_float64 slop) {
   if (simde_math_isnan(a)) {