On 11/27/25 17:22, mengqinggang wrote:
[snip]
diff --git a/gcc/config/loongarch/loongarch-c.cc
b/gcc/config/loongarch/loongarch-c.cc
index fc031a6fe90..6e2f8f43818 100644
--- a/gcc/config/loongarch/loongarch-c.cc
+++ b/gcc/config/loongarch/loongarch-c.cc
@@ -55,6 +55,11 @@ loongarch_define_unconditional_macros (cpp_reader *pfile)
builtin_define ("__loongarch_grlen=64");
builtin_define ("__loongarch64");
}
+ else
+ builtin_define ("__loongarch_grlen=32");
+
+ if (TARGET_ABI_ILP32)
+ builtin_define ("__loongarch_ilp32");
if (TARGET_ABI_LP64)
{
It would be great to indicate LA32R with a builtin macro too, something
like "__loongarch_la32r" sounds great to me. I'm okay with spelling
"reduced" fully out too.
Background: I just found it necessary while trying to port RocksDB's
timestamp routine to LA32, which needs to handle the well-known behavior
of counter wraparound without branching:
rdcntvh.w HI1
rdcntvl.w LO
rdcntvh.w HI2
xor HI1, HI1, HI2 // HI1 != 0 -> wrapped around
masknez LO, LO, HI1 // LO = wrapped around ? 0 : LO
The problem is that "masknez" is absent in LA32R, and requires expansion:
...
sltui HI1, HI1, 1 // HI1 = wrapped around ? 0 : 1
sub HI1, $zero, HI1 // HI1 = wrapped around ? 0 : all ones
and LO, LO, HI1 // LO = wrapped around ? 0 : LO
which means 2 additional cycles on top of 5.
Given that the instruction is present in LA32S (S for "standard"), and
that in practice most if not all LA32 cores sold are LA32S-capable
[^la32s], it can be beneficial to have the "LA32R flag" exposed to
programmers, who can then bring more performance for the majority of
LA32S deployments while preserving compatibility.
[^la32s]: After all, LA32R probably only got defined for educational
purposes, because otherwise Loongson should already have made available
some extremely small LA32R-only IP cores, to promote adoption.
--
Regards,
WANG "xen0n" Xuerui
Linux/LoongArch mailing list: https://lore.kernel.org/loongarch/
Unofficial Loongson OSS community: https://github.com/loongson-community
