[llvm] [RISCV] Canonicalize mask-producing SETCC for easier conversion on the consuming instruction to be VL-predicate (PR #214877)
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Mon Aug 17 10:55:45 PDT 2026
================
@@ -18750,6 +18750,89 @@ static bool narrowIndex(SDValue &N, ISD::MemIndexType IndexType, SelectionDAG &D
return true;
}
+/// Given
+/// ```
+/// %b = splat %base
+/// %s = <%offset, %offset + 1, %offset + 2, %offset + 3, ...>
+/// %a = add nuw %b, %s
+/// %N = splat %n
+/// %p = setcc ult %a, %N
+/// ```
+/// in which the SETCC's condition code could be (U)LT or (U)GT.
+/// We can turn it into
+/// ```
+/// %s = <0, 1, 2, 3, ...>
+/// %m = sub %n, min(%n, %base + %offset)
+/// %M = splat %m
+/// %p = setcc ult %s, %M
+/// ```
+/// where the SETCC's condition code is also canonicalized into (U)LT;
+/// The idea behind this canonicalization is that if %p is used as a mask
+/// in a masked.load/store, we can easily turn it to use VL-predicate later
+/// by assigning `vl = min(%m, <number of vector elements>)`.
+static SDValue canonicalizeMaskForVLPredicate(EVT MaskVT, SDValue LHS,
+ SDValue RHS, ISD::CondCode CC,
+ const SDLoc &DL,
+ SelectionDAG &DAG) {
+ using namespace SDPatternMatch;
+ if (!MaskVT.isFixedLengthVector() ||
+ !(CC == ISD::SETUGT || CC == ISD::SETGT || CC == ISD::SETULT ||
+ CC == ISD::SETLT) ||
+ !LHS.getValueType().isInteger())
+ return SDValue();
+
+ // Canonicalize the comparison.
+ if (CC == ISD::SETUGT || CC == ISD::SETGT) {
+ std::swap(LHS, RHS);
+ // Flip the 'L' and 'G' bits.
+ CC = ISD::CondCode(static_cast<unsigned>(CC) ^ 6);
----------------
topperc wrote:
Use getSetCCSwappedOperands?
https://github.com/llvm/llvm-project/pull/214877
More information about the llvm-commits
mailing list