simde/x86/avx512/fixupimm.h: work around GCC bug 121064

Use simde_memcpy instead of direct assignment to prevent GCC from
generating incorrect vshuf.w instructions on LoongArch with -Ofast.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121064

Author: HecaiYuan

simde/simde-common.h

@@ -1079,7 +1079,7 @@ HEDLEY_DIAGNOSTIC_POP
 #      if HEDLEY_GCC_VERSION_CHECK(16,0,0)
 #        define SIMDE_BUG_GCC_123807
 #      endif
-#      if defined(SIMDE_LOONGARCH_LSX_NATIVE) && \
+#      if defined(SIMDE_ARCH_LOONGARCH) && \
            ((HEDLEY_GCC_VERSION_CHECK(14,0,0) && !HEDLEY_GCC_VERSION_CHECK(14,4,0)) || \
             (HEDLEY_GCC_VERSION_CHECK(15,0,0) && !HEDLEY_GCC_VERSION_CHECK(15,2,0)))
 #        define SIMDE_BUG_GCC_121064

simde/x86/avx512/fixupimm.h

@@ -401,7 +401,14 @@ simde_mm_fixupimm_ss (simde__m128 a, simde__m128 b, simde__m128i c, int imm8)
 
   switch (((c_.i32[0] >> (select << 2)) & 15)) {
     case 0:
-      b_.f32[0] =  a_.f32[0];
+      #if defined(SIMDE_BUG_GCC_121064)
+      {
+         simde_float32 tmp = a_.f32[0];
+         simde_memcpy(&b_.f32[0], &tmp, sizeof(tmp));
+      }
+      #else
+         b_.f32[0] = a_.f32[0];
+      #endif
       break;
     case 2:
       b_.f32[0] =  SIMDE_MATH_NANF;