On 13/03/11 13:40, Tom Anderson wrote:
On Sat, 12 Mar 2011, Bruce Stephens wrote:

Tom Anderson writes:

On the flip side, the 64-bitness of the registers only gives a speed
boost to code that specifically needs to manipulate wide operands -
cryptography, scientific computing, perhaps some SIMD-based graphics
operations. For most use today, 32-bit+PAE should be faster than
64-bit. And yet, the market has moved almost entirely over to 64-bit.
Grr.

IIUC the 64 bit instruction set has lots more registers (not just
bigger ones), and that's a significant benefit for lots of applications.

It has more *architectural* registers - ones that a program can refer to
explicitly. But CPUs since the 90s have had more actual registers than
that, and have used register renaming to supply them to the program.
That has never been perfect, and it hasn't made compiler writers' lives
any easier, but it means the impact of adding more registers to an
architecture is not as great as you might think.

Someone must have done a properly-controlled benchmark of 32-bit vs
64-bit on the x86 at some point (same hardware, same source code, same
compiler). It doesn't seem like it would be hard to do. Does anyone know
of one?

Off the top of my head, I remember reading ones by Phoronix at the time.