Use a shorter insruction on x86 for loading unsigned 32-bit immediates into 64-bit registers.

[valoran-M]

Writing into a 32-bit register erases the upper 32 bits of a 64-bit register. This is the shortest instruction for loading unsigned 32-bit immediates. The length of those instructions is:

• movl: 5 bytes
• movq: 7 bytes
• movabsq: 10 bytes

Note: Most assemblers will choose the proper instruction by themselves; this commit just makes this explicit.

[silene]

For the record, this issue showed up when improving the code generation for division by a constant. The algorithm loads a 64-bit immediate between 2^31 and 2^32-1, which was thus emitted as movabsq by CompCert. But GCC/GAS honors this choice (to account for self-modifying code?), which means that the machine code would use the 10-byte instruction where the 5-byte one would have been sufficient. This can have adverse effect on instruction decoding and caching.

[xavierleroy]

Well spotted! I was not aware that movl produces shorter code than movq even in 64 bit mode. And there's the additional benefit for constants in the 2^31.. 2^32-1 range. Nice!