Even intrinsics didn't even necessarily work well. MSVC, in particular, was really, really bad back then with SIMD intrinsics -- any use of MMX or SSE intrinsic

Even intrinsics didn't even necessarily work well. MSVC, in particular, was really, really bad back then with SIMD intrinsics -- any use of MMX or SSE intrinsics in VC6 would result in more than two-thirds of the generated code being move instructions, with a single value sometimes moved two or three times between ALU instructions. It was trivial to beat the compiler with hand-written assembly. MMX intrinsics were never fixed and SSE intrinsics weren't fixed until VS2010. For scalar code, it was more that the CPUs got better, as out-of-order execution starting with the Pentium Pro made instruction scheduling less important. The original Pentium CPU was an in-order design with two pipes where the second V pipe had significant restrictions, which was harder for compilers to deal with than the PPro/PII and its decoding pattern.