You may be right, but from observation, it crops up rather frequently
with the proposed "Fast "x mod constant = 0" optimisation
<https://gitlab.com/freepascal.org/fpc/source/-/merge_requests/45>"
merge request and occasionally with some divisions by constants. The
argument there though would be to simply encode the optimisation
directly into the code generation.
Outside of RISC processors like aarch64, it is generally an x86_64-only
thing though because a lot of instructions like ADD and CMP don't
support 64-bit immediates.
Gareth aka. Kit
On 17/10/2021 14:52, Florian Klämpfl via fpc-devel wrote:
Am 17.10.2021 um 13:25 schrieb J. Gareth Moreton via fpc-devel
<fpc-devel@lists.freepascal.org>:
Hi everyone,
While reading up on some algorithms, I came across a recommendation of using a
shorter arithmetic function to change the value of a constant in a register
rather than loading the new value directly. However, the algorithm assumes a
RISC-like processor, so I'm not sure if it applies to an Intel x86-64
processor. Consider the following:
movq $0xaaaaaaaaaaaaaaab,%rax
imulq %rax,%rcx
movq $0x5555555555555555,%rax
cmpq %rax,%rcx
setle %al
This algorithm sets %al to 1 if %rcx is divisible by 3, and 0 if it's not, and
was compiled from the following Pascal code (under -O3, but -O1 produces almost
exactly the same):
function IsDivisible3(Numerator: QWord): Boolean;
begin
Result := (Numerator * $AAAAAAAAAAAAAAAB) <= $5555555555555555;
end;
(One of my merge requests produces this code from "Result := (x mod 3) = 0")
My question is this: can "movq $0x5555555555555555,%rax" be replaced with "shrq
$0x1,%rax" without incurring an additional pipeline stall? The MOV instruction takes 10 bytes to store,
while "SHR 1" takes only 3. Given that %rax is used beforehand and the CMP instruction has to wait
until the IMUL instruction has finished executing, logic tells me that I can get away with it here, but I'm
not sure if the metric to go by is the execution speed of IMUL (i.e. the IMUL instruction is the limiting
factor before CMP can be executed), or the simple fact that the previous value of %rax was used and will be
loaded with $AAAAAAAAAAAAAAAB by the time it comes to load it with a new value.
I’d expect that the shl is executed in parallel on with the imul on most modern
out of order architectures. So no real issue. OTOH, this is a very rare case so
it is questionable if it is useful to check for this situation.
_______________________________________________
fpc-devel maillist - fpc-devel@lists.freepascal.org
https://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
--
This email has been checked for viruses by Avast antivirus software.
https://www.avast.com/antivirus
_______________________________________________
fpc-devel maillist - fpc-devel@lists.freepascal.org
https://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
