There has never been a one-to-one mapping (rather one-to-many), but it's still perfectly possible to transform C into assembly by hand. It's not even that difficult.
Yes, back in the old 8 and 16 bit days, it was pretty much one-to-one for most use cases.
Nowadays not any longer if you want to write code that takes advantage of branch prediction, speculative execution, cache lines, vector units, GPGPU ...
Just watch this Going Native talk on how sometimes generating code that is 4x bigger than the direct translation can yield up to 30% performance increase.
Auto vectorization can be impressive, but it applies to hard numerical code, not so much in regular code bases. I regularly debug assembly in my job, and in the vast majority of cases it's a fairly straightforward translation of the C/C++ code. Stuff like cache locality or using the GPU is still determined by the C code.
I'll concede there's more fine variations to think about than with an 8-bit system, though, with the padding, etc.
