================
@@ -24483,6 +24483,26 @@ ISD::NodeType
RISCVTargetLowering::getExtendForAtomicCmpSwapArg() const {
return Subtarget.hasStdExtZacas() ? ISD::ANY_EXTEND : ISD::SIGN_EXTEND;
}
+ISD::NodeType RISCVTargetLowering::getExtendForAtomicRMWArg(unsigned Op) const
{
+ // Zaamo will use amo<op>.w which does not require extension.
+ if (Subtarget.hasStdExtZaamo() || Subtarget.hasForcedAtomics())
+ return ISD::ANY_EXTEND;
+
+ // Zalasr pseudo expansions with comparison
+ assert(Subtarget.hasStdExtZalrsc());
+ switch (Op) {
+ case ISD::ATOMIC_LOAD_MIN:
+ case ISD::ATOMIC_LOAD_MAX:
+ return ISD::SIGN_EXTEND;
+ case ISD::ATOMIC_LOAD_UMIN:
+ case ISD::ATOMIC_LOAD_UMAX:
+ return ISD::ZERO_EXTEND;
----------------
slachowsky wrote:
Good idea.
The LR.W used in the ExpandAtomicPseudos is always sign-extending the value
read from memory, so the idea is to shrink the sequence from:
```
atomicrmw_umax_i32_monotonic: # @atomicrmw_umax_i32_monotonic
# %bb.0:
slli a1, a1, 32
srli a2, a1, 32
.LBB0_1: # =>This Inner Loop Header: Depth=1
lr.w a1, (a0)
slli a3, a1, 32
srli a3, a3, 32
bgeu a3, a2, .LBB0_3
# %bb.2: # in Loop: Header=BB0_1 Depth=1
mv a3, a2
.LBB0_3: # in Loop: Header=BB0_1 Depth=1
sc.w a3, a3, (a0)
bnez a3, .LBB0_1
# %bb.4:
mv a0, a1
ret
```
To something like:
```
atomicrmw_umax_i32_monotonic: # @atomicrmw_umax_i32_monotonic
# %bb.0:
sext.w a2, a1
.LBB0_1: # =>This Inner Loop Header: Depth=1
lr.w a1, (a0)
bgeu a3, a2, .LBB0_3
# %bb.2: # in Loop: Header=BB0_1 Depth=1
mv a3, a2
.LBB0_3: # in Loop: Header=BB0_1 Depth=1
sc.w a3, a3, (a0)
bnez a3, .LBB0_1
# %bb.4:
mv a0, a1
ret
```
https://github.com/llvm/llvm-project/pull/163672
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
