https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126022
Bug ID: 126022
Summary: For RISC-V, prefer li+slli rather than lui+add for
some constants
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: law at gcc dot gnu.org
Target Milestone: ---
There are a variety of methods to synthesize constants. Those methods have
primarily been focused on minimizing the instruction count (as a rough proxy
for performance). However, there are cases where we can produce the same
instruction count, but generate sequences that potentially compress better.
Let's consider this case for -O2 -march=rv64gcbv_zicond
unsigned long fubar(unsigned long x) { return 3072UL; }
The naive code generation for this would look like:
li a0,4096
addi a0,a0,-1024
However the "addi" instruction is a 4 byte instruction.
Instead we could generate:
li a0, 3
slli a0, a0, 10
That has the advantage that it compresses better (the slli compresses to 2
bytes). Of course lui+addi is the poster child for fusion, so on designs which
support fusion, lui+addi may still be the better choice. But certainly for
-Os, and on designs without fusion support the li+slli sequence is better.
I suspect we just need some special casing in the constant synthesis code. We
can query fusion status in there as well so we can likely tune this however we
want.
