aboutsummaryrefslogtreecommitdiff
path: root/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/lib/Target/RISCV/RISCVISelLowering.cpp')
-rw-r--r--llvm/lib/Target/RISCV/RISCVISelLowering.cpp165
1 files changed, 76 insertions, 89 deletions
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index c6a8b84..c56ce3f 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -16495,6 +16495,35 @@ static SDValue expandMulToAddOrSubOfShl(SDNode *N, SelectionDAG &DAG,
return DAG.getNode(Op, DL, VT, Shift1, Shift2);
}
+static SDValue getShlAddShlAdd(SDNode *N, SelectionDAG &DAG, unsigned ShX,
+ unsigned ShY) {
+ SDLoc DL(N);
+ EVT VT = N->getValueType(0);
+ SDValue X = N->getOperand(0);
+ SDValue Mul359 = DAG.getNode(RISCVISD::SHL_ADD, DL, VT, X,
+ DAG.getConstant(ShY, DL, VT), X);
+ return DAG.getNode(RISCVISD::SHL_ADD, DL, VT, Mul359,
+ DAG.getConstant(ShX, DL, VT), Mul359);
+}
+
+static SDValue expandMulToShlAddShlAdd(SDNode *N, SelectionDAG &DAG,
+ uint64_t MulAmt) {
+ switch (MulAmt) {
+ case 5 * 3:
+ return getShlAddShlAdd(N, DAG, 2, 1);
+ case 9 * 3:
+ return getShlAddShlAdd(N, DAG, 3, 1);
+ case 5 * 5:
+ return getShlAddShlAdd(N, DAG, 2, 2);
+ case 9 * 5:
+ return getShlAddShlAdd(N, DAG, 3, 2);
+ case 9 * 9:
+ return getShlAddShlAdd(N, DAG, 3, 3);
+ default:
+ return SDValue();
+ }
+}
+
// Try to expand a scalar multiply to a faster sequence.
static SDValue expandMul(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
@@ -16524,18 +16553,17 @@ static SDValue expandMul(SDNode *N, SelectionDAG &DAG,
if (Subtarget.hasVendorXqciac() && isInt<12>(CNode->getSExtValue()))
return SDValue();
- // WARNING: The code below is knowingly incorrect with regards to undef semantics.
- // We're adding additional uses of X here, and in principle, we should be freezing
- // X before doing so. However, adding freeze here causes real regressions, and no
- // other target properly freezes X in these cases either.
- SDValue X = N->getOperand(0);
-
+ // WARNING: The code below is knowingly incorrect with regards to undef
+ // semantics. We're adding additional uses of X here, and in principle, we
+ // should be freezing X before doing so. However, adding freeze here causes
+ // real regressions, and no other target properly freezes X in these cases
+ // either.
if (Subtarget.hasShlAdd(3)) {
+ SDValue X = N->getOperand(0);
int Shift;
if (int ShXAmount = isShifted359(MulAmt, Shift)) {
// 3/5/9 * 2^N -> shl (shXadd X, X), N
SDLoc DL(N);
- SDValue X = N->getOperand(0);
// Put the shift first if we can fold a zext into the shift forming
// a slli.uw.
if (X.getOpcode() == ISD::AND && isa<ConstantSDNode>(X.getOperand(1)) &&
@@ -16554,38 +16582,8 @@ static SDValue expandMul(SDNode *N, SelectionDAG &DAG,
}
// 3/5/9 * 3/5/9 -> shXadd (shYadd X, X), (shYadd X, X)
- int ShX;
- int ShY;
- switch (MulAmt) {
- case 3 * 5:
- ShY = 1;
- ShX = 2;
- break;
- case 3 * 9:
- ShY = 1;
- ShX = 3;
- break;
- case 5 * 5:
- ShX = ShY = 2;
- break;
- case 5 * 9:
- ShY = 2;
- ShX = 3;
- break;
- case 9 * 9:
- ShX = ShY = 3;
- break;
- default:
- ShX = ShY = 0;
- break;
- }
- if (ShX) {
- SDLoc DL(N);
- SDValue Mul359 = DAG.getNode(RISCVISD::SHL_ADD, DL, VT, X,
- DAG.getConstant(ShY, DL, VT), X);
- return DAG.getNode(RISCVISD::SHL_ADD, DL, VT, Mul359,
- DAG.getConstant(ShX, DL, VT), Mul359);
- }
+ if (SDValue V = expandMulToShlAddShlAdd(N, DAG, MulAmt))
+ return V;
// If this is a power 2 + 2/4/8, we can use a shift followed by a single
// shXadd. First check if this a sum of two power of 2s because that's
@@ -16648,23 +16646,12 @@ static SDValue expandMul(SDNode *N, SelectionDAG &DAG,
}
}
- for (uint64_t Divisor : {3, 5, 9}) {
- if (MulAmt % Divisor != 0)
- continue;
- uint64_t MulAmt2 = MulAmt / Divisor;
- // 3/5/9 * 3/5/9 * 2^N - In particular, this covers multiples
- // of 25 which happen to be quite common.
- if (int ShBAmount = isShifted359(MulAmt2, Shift)) {
- SDLoc DL(N);
- SDValue Mul359A =
- DAG.getNode(RISCVISD::SHL_ADD, DL, VT, X,
- DAG.getConstant(Log2_64(Divisor - 1), DL, VT), X);
- SDValue Mul359B =
- DAG.getNode(RISCVISD::SHL_ADD, DL, VT, Mul359A,
- DAG.getConstant(ShBAmount, DL, VT), Mul359A);
- return DAG.getNode(ISD::SHL, DL, VT, Mul359B,
- DAG.getConstant(Shift, DL, VT));
- }
+ // 3/5/9 * 3/5/9 * 2^N - In particular, this covers multiples
+ // of 25 which happen to be quite common.
+ Shift = llvm::countr_zero(MulAmt);
+ if (SDValue V = expandMulToShlAddShlAdd(N, DAG, MulAmt >> Shift)) {
+ SDLoc DL(N);
+ return DAG.getNode(ISD::SHL, DL, VT, V, DAG.getConstant(Shift, DL, VT));
}
}
@@ -23946,7 +23933,7 @@ RISCVTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
.Case("{t0}", RISCV::X5)
.Case("{t1}", RISCV::X6)
.Case("{t2}", RISCV::X7)
- .Cases("{s0}", "{fp}", RISCV::X8)
+ .Cases({"{s0}", "{fp}"}, RISCV::X8)
.Case("{s1}", RISCV::X9)
.Case("{a0}", RISCV::X10)
.Case("{a1}", RISCV::X11)
@@ -23983,38 +23970,38 @@ RISCVTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
// use the ABI names in register constraint lists.
if (Subtarget.hasStdExtF()) {
unsigned FReg = StringSwitch<unsigned>(Constraint.lower())
- .Cases("{f0}", "{ft0}", RISCV::F0_F)
- .Cases("{f1}", "{ft1}", RISCV::F1_F)
- .Cases("{f2}", "{ft2}", RISCV::F2_F)
- .Cases("{f3}", "{ft3}", RISCV::F3_F)
- .Cases("{f4}", "{ft4}", RISCV::F4_F)
- .Cases("{f5}", "{ft5}", RISCV::F5_F)
- .Cases("{f6}", "{ft6}", RISCV::F6_F)
- .Cases("{f7}", "{ft7}", RISCV::F7_F)
- .Cases("{f8}", "{fs0}", RISCV::F8_F)
- .Cases("{f9}", "{fs1}", RISCV::F9_F)
- .Cases("{f10}", "{fa0}", RISCV::F10_F)
- .Cases("{f11}", "{fa1}", RISCV::F11_F)
- .Cases("{f12}", "{fa2}", RISCV::F12_F)
- .Cases("{f13}", "{fa3}", RISCV::F13_F)
- .Cases("{f14}", "{fa4}", RISCV::F14_F)
- .Cases("{f15}", "{fa5}", RISCV::F15_F)
- .Cases("{f16}", "{fa6}", RISCV::F16_F)
- .Cases("{f17}", "{fa7}", RISCV::F17_F)
- .Cases("{f18}", "{fs2}", RISCV::F18_F)
- .Cases("{f19}", "{fs3}", RISCV::F19_F)
- .Cases("{f20}", "{fs4}", RISCV::F20_F)
- .Cases("{f21}", "{fs5}", RISCV::F21_F)
- .Cases("{f22}", "{fs6}", RISCV::F22_F)
- .Cases("{f23}", "{fs7}", RISCV::F23_F)
- .Cases("{f24}", "{fs8}", RISCV::F24_F)
- .Cases("{f25}", "{fs9}", RISCV::F25_F)
- .Cases("{f26}", "{fs10}", RISCV::F26_F)
- .Cases("{f27}", "{fs11}", RISCV::F27_F)
- .Cases("{f28}", "{ft8}", RISCV::F28_F)
- .Cases("{f29}", "{ft9}", RISCV::F29_F)
- .Cases("{f30}", "{ft10}", RISCV::F30_F)
- .Cases("{f31}", "{ft11}", RISCV::F31_F)
+ .Cases({"{f0}", "{ft0}"}, RISCV::F0_F)
+ .Cases({"{f1}", "{ft1}"}, RISCV::F1_F)
+ .Cases({"{f2}", "{ft2}"}, RISCV::F2_F)
+ .Cases({"{f3}", "{ft3}"}, RISCV::F3_F)
+ .Cases({"{f4}", "{ft4}"}, RISCV::F4_F)
+ .Cases({"{f5}", "{ft5}"}, RISCV::F5_F)
+ .Cases({"{f6}", "{ft6}"}, RISCV::F6_F)
+ .Cases({"{f7}", "{ft7}"}, RISCV::F7_F)
+ .Cases({"{f8}", "{fs0}"}, RISCV::F8_F)
+ .Cases({"{f9}", "{fs1}"}, RISCV::F9_F)
+ .Cases({"{f10}", "{fa0}"}, RISCV::F10_F)
+ .Cases({"{f11}", "{fa1}"}, RISCV::F11_F)
+ .Cases({"{f12}", "{fa2}"}, RISCV::F12_F)
+ .Cases({"{f13}", "{fa3}"}, RISCV::F13_F)
+ .Cases({"{f14}", "{fa4}"}, RISCV::F14_F)
+ .Cases({"{f15}", "{fa5}"}, RISCV::F15_F)
+ .Cases({"{f16}", "{fa6}"}, RISCV::F16_F)
+ .Cases({"{f17}", "{fa7}"}, RISCV::F17_F)
+ .Cases({"{f18}", "{fs2}"}, RISCV::F18_F)
+ .Cases({"{f19}", "{fs3}"}, RISCV::F19_F)
+ .Cases({"{f20}", "{fs4}"}, RISCV::F20_F)
+ .Cases({"{f21}", "{fs5}"}, RISCV::F21_F)
+ .Cases({"{f22}", "{fs6}"}, RISCV::F22_F)
+ .Cases({"{f23}", "{fs7}"}, RISCV::F23_F)
+ .Cases({"{f24}", "{fs8}"}, RISCV::F24_F)
+ .Cases({"{f25}", "{fs9}"}, RISCV::F25_F)
+ .Cases({"{f26}", "{fs10}"}, RISCV::F26_F)
+ .Cases({"{f27}", "{fs11}"}, RISCV::F27_F)
+ .Cases({"{f28}", "{ft8}"}, RISCV::F28_F)
+ .Cases({"{f29}", "{ft9}"}, RISCV::F29_F)
+ .Cases({"{f30}", "{ft10}"}, RISCV::F30_F)
+ .Cases({"{f31}", "{ft11}"}, RISCV::F31_F)
.Default(RISCV::NoRegister);
if (FReg != RISCV::NoRegister) {
assert(RISCV::F0_F <= FReg && FReg <= RISCV::F31_F && "Unknown fp-reg");
generated by cgit v1.1 at 2025-11-04 21:27:43 +0000