[RISCV] Fold Zba-expanded (mul (shr exact X, C1), C2) (llvm#168019)

Author: pfusik

llvm/include/llvm/CodeGen/SDPatternMatch.h

@@ -903,6 +903,11 @@ template <typename LHS, typename RHS>
 inline BinaryOpc_match<LHS, RHS> m_Srl(const LHS &L, const RHS &R) {
   return BinaryOpc_match<LHS, RHS>(ISD::SRL, L, R);
 }
+template <typename LHS, typename RHS>
+inline auto m_ExactSr(const LHS &L, const RHS &R) {
+  return m_AnyOf(BinaryOpc_match<LHS, RHS>(ISD::SRA, L, R, SDNodeFlags::Exact),
+                 BinaryOpc_match<LHS, RHS>(ISD::SRL, L, R, SDNodeFlags::Exact));
+}
 
 template <typename LHS, typename RHS>
 inline BinaryOpc_match<LHS, RHS> m_Rotl(const LHS &L, const RHS &R) {

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -16798,9 +16798,7 @@ static SDValue expandMulToAddOrSubOfShl(SDNode *N, SelectionDAG &DAG,
     // because X is exact (Y >> M + 2).
     uint64_t ShAmt = Log2_64(MulAmtLowBit) + 2;
     using namespace SDPatternMatch;
-    return sd_match(X, m_AnyOf(m_Sra(m_Value(), m_SpecificInt(ShAmt)),
-                               m_Srl(m_Value(), m_SpecificInt(ShAmt)))) &&
-           X->getFlags().hasExact();
+    return sd_match(X, m_ExactSr(m_Value(), m_SpecificInt(ShAmt)));
   };
   if (isPowerOf2_64(MulAmt - MulAmtLowBit) && !(CanSub && PreferSub())) {
     Op = ISD::ADD;
@@ -16825,10 +16823,13 @@ static SDValue getShlAddShlAdd(SDNode *N, SelectionDAG &DAG, unsigned ShX,
   SDLoc DL(N);
   EVT VT = N->getValueType(0);
   SDValue X = N->getOperand(0);
-  // Put the shift first if we can fold a zext into the shift forming a slli.uw.
+  // Put the shift first if we can fold:
+  // a. a zext into the shift forming a slli.uw
+  // b. an exact shift right forming one shorter shift or no shift at all
   using namespace SDPatternMatch;
   if (Shift != 0 &&
-      sd_match(X, m_And(m_Value(), m_SpecificInt(UINT64_C(0xffffffff))))) {
+      sd_match(X, m_AnyOf(m_And(m_Value(), m_SpecificInt(UINT64_C(0xffffffff))),
+                          m_ExactSr(m_Value(), m_ConstInt())))) {
     X = DAG.getNode(ISD::SHL, DL, VT, X, DAG.getConstant(Shift, DL, VT));
     Shift = 0;
   }

llvm/test/CodeGen/RISCV/rv64zba.ll

@@ -5016,3 +5016,74 @@ define ptr @shl_add_knownbits(ptr %p, i64 %i) {
   %r = getelementptr i8, ptr %p, i64 %shr
   ret ptr %r
 }
+
+define i64 @exactashr1mul6(i64 %a) {
+; RV64I-LABEL: exactashr1mul6:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    slli a1, a0, 1
+; RV64I-NEXT:    add a0, a1, a0
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: exactashr1mul6:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    sh1add a0, a0, a0
+; RV64ZBA-NEXT:    ret
+;
+; RV64XANDESPERF-LABEL: exactashr1mul6:
+; RV64XANDESPERF:       # %bb.0:
+; RV64XANDESPERF-NEXT:    nds.lea.h a0, a0, a0
+; RV64XANDESPERF-NEXT:    ret
+  %c = ashr exact i64 %a, 1
+  %d = mul i64 %c, 6
+  ret i64 %d
+}
+
+define i64 @exactlshr3mul22(i64 %a) {
+; RV64I-LABEL: exactlshr3mul22:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    srli a0, a0, 3
+; RV64I-NEXT:    li a1, 22
+; RV64I-NEXT:    mul a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: exactlshr3mul22:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    srli a0, a0, 2
+; RV64ZBA-NEXT:    sh2add a1, a0, a0
+; RV64ZBA-NEXT:    sh1add a0, a1, a0
+; RV64ZBA-NEXT:    ret
+;
+; RV64XANDESPERF-LABEL: exactlshr3mul22:
+; RV64XANDESPERF:       # %bb.0:
+; RV64XANDESPERF-NEXT:    srli a0, a0, 2
+; RV64XANDESPERF-NEXT:    nds.lea.w a1, a0, a0
+; RV64XANDESPERF-NEXT:    nds.lea.h a0, a0, a1
+; RV64XANDESPERF-NEXT:    ret
+  %c = lshr exact i64 %a, 3
+  %d = mul i64 %c, 22
+  ret i64 %d
+}
+
+define i64 @exactashr1mul36(i64 %a) {
+; RV64I-LABEL: exactashr1mul36:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    slli a1, a0, 1
+; RV64I-NEXT:    slli a0, a0, 4
+; RV64I-NEXT:    add a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: exactashr1mul36:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    slli a0, a0, 1
+; RV64ZBA-NEXT:    sh3add a0, a0, a0
+; RV64ZBA-NEXT:    ret
+;
+; RV64XANDESPERF-LABEL: exactashr1mul36:
+; RV64XANDESPERF:       # %bb.0:
+; RV64XANDESPERF-NEXT:    slli a0, a0, 1
+; RV64XANDESPERF-NEXT:    nds.lea.d a0, a0, a0
+; RV64XANDESPERF-NEXT:    ret
+  %c = ashr exact i64 %a, 1
+  %d = mul i64 %c, 36
+  ret i64 %d
+}