3 ms·

> Unfortunately, LEA became very slow in the recent generations of CPUs. Intel doesn’t suggest this will change. lea is a pretty common instruction even in non

by rwmj 3y ago

> Unfortunately, LEA became very slow in the recent generations of CPUs. Intel doesn’t suggest this will change. lea is a pretty common instruction even in non-OCaml code so this is annoying. I wonder if this has changed since 2014? $ objdump -d /usr/bin/ls | grep '\blea\b' | wc -l 908 $ objdump -d /usr/bin/ls | wc -l 21398 4.2% of all instructions, and that's an underestimate because the raw objdump -d output contains blank lines, nops, etc. For an OCaml 4.14 program (/usr/bin/virt-v2v), I got 38886 / 432937 which is 9.0%. For OCaml 5 using the flambda optimizer, 36531 / 480357 = 7.6%

orra 3y ago
> I wonder if this has changed since 2014? Good question. Probably not. LLVM reduced the number of LEAs in 2017. https://reviews.llvm.org/D32277 https://reviews.llvm.org/D32277 Also, Golang is gradually reducing the number of slow LEAs. https://github.com/golang/go/issues/21735 https://github.com/golang/go/issues/21735 > grep '\blea\b' | wc -l This maybe isn't sophisticated enough? The other links seem to explain two operand LEA isn't so problematic.
- rwmj 3y ago
  Looks like there have been proposals to eliminate use of 3 operand lea in OCaml code (not accepted sadly): https://github.com/ocaml/ocaml/pull/8531 https://github.com/ocaml/ocaml/pull/8531 https://github.com/ocaml/ocaml/issues/9850 https://github.com/ocaml/ocaml/issues/9850
  - asddubs 3y ago
    not outright rejected either though, the PRs just both died
saagarjha 3y ago
Not all leas, just the three-operand version described in the post. The less complex ones are heavily used for address construction and (less commonly) just normal arithmetic since they can help with port usage.
devit 3y ago
Yes, Zen 2+ can execute 4 two-operand LEAs per cycles (according to Agner Fog), and apparently Sapphire Rapids can execute 5 LEAs of any 32/64-bit kind per cycle (according to Intel).
rollcat 3y ago
Common in program text != common in hot code paths. To estimate the actual impact, you'd have to run the code in a profiler.

by rwmj 3y ago

orra 3y ago

> I wonder if this has changed since 2014? Good question. Probably not. LLVM reduced the number of LEAs in 2017. https://reviews.llvm.org/D32277 https://reviews.llvm.org/D32277 Also, Golang is gradually reducing the number of slow LEAs. https://github.com/golang/go/issues/21735 https://github.com/golang/go/issues/21735 > grep '\blea\b' | wc -l This maybe isn't sophisticated enough? The other links seem to explain two operand LEA isn't so problematic.

rwmj 3y ago
Looks like there have been proposals to eliminate use of 3 operand lea in OCaml code (not accepted sadly): https://github.com/ocaml/ocaml/pull/8531 https://github.com/ocaml/ocaml/pull/8531 https://github.com/ocaml/ocaml/issues/9850 https://github.com/ocaml/ocaml/issues/9850
- asddubs 3y ago
  not outright rejected either though, the PRs just both died

saagarjha 3y ago

Not all leas, just the three-operand version described in the post. The less complex ones are heavily used for address construction and (less commonly) just normal arithmetic since they can help with port usage.

devit 3y ago

Yes, Zen 2+ can execute 4 two-operand LEAs per cycles (according to Agner Fog), and apparently Sapphire Rapids can execute 5 LEAs of any 32/64-bit kind per cycle (according to Intel).

rollcat 3y ago

Common in program text != common in hot code paths. To estimate the actual impact, you'd have to run the code in a profiler.