[llvm-commits] CVS: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
Nate Begeman
natebegeman at mac.com
Wed Oct 19 19:15:56 PDT 2005
Changes in directory llvm/lib/CodeGen/SelectionDAG:
DAGCombiner.cpp updated: 1.47 -> 1.48
---
Log message:
Move the target constant divide optimization up into the dag combiner, so
that the nodes can be folded with other nodes, and we can not duplicate
code in every backend. Alpha will probably want this too.
---
Diffs of the changes: (+263 -0)
DAGCombiner.cpp | 263 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 263 insertions(+)
Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
diff -u llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp:1.47 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp:1.48
--- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp:1.47 Tue Oct 18 01:04:22 2005
+++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp Wed Oct 19 21:15:44 2005
@@ -180,6 +180,9 @@
SDOperand N3, ISD::CondCode CC);
SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
ISD::CondCode Cond, bool foldBooleans = true);
+
+ SDOperand BuildSDIV(SDNode *N);
+ SDOperand BuildUDIV(SDNode *N);
public:
DAGCombiner(SelectionDAG &D)
: DAG(D), TLI(D.getTargetLoweringInfo()), AfterLegalize(false) {}
@@ -189,6 +192,178 @@
};
}
+struct ms {
+ int64_t m; // magic number
+ int64_t s; // shift amount
+};
+
+struct mu {
+ uint64_t m; // magic number
+ int64_t a; // add indicator
+ int64_t s; // shift amount
+};
+
+/// magic - calculate the magic numbers required to codegen an integer sdiv as
+/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
+/// or -1.
+static ms magic32(int32_t d) {
+ int32_t p;
+ uint32_t ad, anc, delta, q1, r1, q2, r2, t;
+ const uint32_t two31 = 0x80000000U;
+ struct ms mag;
+
+ ad = abs(d);
+ t = two31 + ((uint32_t)d >> 31);
+ anc = t - 1 - t%ad; // absolute value of nc
+ p = 31; // initialize p
+ q1 = two31/anc; // initialize q1 = 2p/abs(nc)
+ r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
+ q2 = two31/ad; // initialize q2 = 2p/abs(d)
+ r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
+ do {
+ p = p + 1;
+ q1 = 2*q1; // update q1 = 2p/abs(nc)
+ r1 = 2*r1; // update r1 = rem(2p/abs(nc))
+ if (r1 >= anc) { // must be unsigned comparison
+ q1 = q1 + 1;
+ r1 = r1 - anc;
+ }
+ q2 = 2*q2; // update q2 = 2p/abs(d)
+ r2 = 2*r2; // update r2 = rem(2p/abs(d))
+ if (r2 >= ad) { // must be unsigned comparison
+ q2 = q2 + 1;
+ r2 = r2 - ad;
+ }
+ delta = ad - r2;
+ } while (q1 < delta || (q1 == delta && r1 == 0));
+
+ mag.m = (int32_t)(q2 + 1); // make sure to sign extend
+ if (d < 0) mag.m = -mag.m; // resulting magic number
+ mag.s = p - 32; // resulting shift
+ return mag;
+}
+
+/// magicu - calculate the magic numbers required to codegen an integer udiv as
+/// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
+static mu magicu32(uint32_t d) {
+ int32_t p;
+ uint32_t nc, delta, q1, r1, q2, r2;
+ struct mu magu;
+ magu.a = 0; // initialize "add" indicator
+ nc = - 1 - (-d)%d;
+ p = 31; // initialize p
+ q1 = 0x80000000/nc; // initialize q1 = 2p/nc
+ r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
+ q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
+ r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
+ do {
+ p = p + 1;
+ if (r1 >= nc - r1 ) {
+ q1 = 2*q1 + 1; // update q1
+ r1 = 2*r1 - nc; // update r1
+ }
+ else {
+ q1 = 2*q1; // update q1
+ r1 = 2*r1; // update r1
+ }
+ if (r2 + 1 >= d - r2) {
+ if (q2 >= 0x7FFFFFFF) magu.a = 1;
+ q2 = 2*q2 + 1; // update q2
+ r2 = 2*r2 + 1 - d; // update r2
+ }
+ else {
+ if (q2 >= 0x80000000) magu.a = 1;
+ q2 = 2*q2; // update q2
+ r2 = 2*r2 + 1; // update r2
+ }
+ delta = d - 1 - r2;
+ } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
+ magu.m = q2 + 1; // resulting magic number
+ magu.s = p - 32; // resulting shift
+ return magu;
+}
+
+/// magic - calculate the magic numbers required to codegen an integer sdiv as
+/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
+/// or -1.
+static ms magic64(int64_t d) {
+ int64_t p;
+ uint64_t ad, anc, delta, q1, r1, q2, r2, t;
+ const uint64_t two63 = 9223372036854775808ULL; // 2^63
+ struct ms mag;
+
+ ad = llabs(d);
+ t = two63 + ((uint64_t)d >> 63);
+ anc = t - 1 - t%ad; // absolute value of nc
+ p = 63; // initialize p
+ q1 = two63/anc; // initialize q1 = 2p/abs(nc)
+ r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
+ q2 = two63/ad; // initialize q2 = 2p/abs(d)
+ r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
+ do {
+ p = p + 1;
+ q1 = 2*q1; // update q1 = 2p/abs(nc)
+ r1 = 2*r1; // update r1 = rem(2p/abs(nc))
+ if (r1 >= anc) { // must be unsigned comparison
+ q1 = q1 + 1;
+ r1 = r1 - anc;
+ }
+ q2 = 2*q2; // update q2 = 2p/abs(d)
+ r2 = 2*r2; // update r2 = rem(2p/abs(d))
+ if (r2 >= ad) { // must be unsigned comparison
+ q2 = q2 + 1;
+ r2 = r2 - ad;
+ }
+ delta = ad - r2;
+ } while (q1 < delta || (q1 == delta && r1 == 0));
+
+ mag.m = q2 + 1;
+ if (d < 0) mag.m = -mag.m; // resulting magic number
+ mag.s = p - 64; // resulting shift
+ return mag;
+}
+
+/// magicu - calculate the magic numbers required to codegen an integer udiv as
+/// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
+static mu magicu64(uint64_t d)
+{
+ int64_t p;
+ uint64_t nc, delta, q1, r1, q2, r2;
+ struct mu magu;
+ magu.a = 0; // initialize "add" indicator
+ nc = - 1 - (-d)%d;
+ p = 63; // initialize p
+ q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
+ r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
+ q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
+ r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
+ do {
+ p = p + 1;
+ if (r1 >= nc - r1 ) {
+ q1 = 2*q1 + 1; // update q1
+ r1 = 2*r1 - nc; // update r1
+ }
+ else {
+ q1 = 2*q1; // update q1
+ r1 = 2*r1; // update r1
+ }
+ if (r2 + 1 >= d - r2) {
+ if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
+ q2 = 2*q2 + 1; // update q2
+ r2 = 2*r2 + 1 - d; // update r2
+ }
+ else {
+ if (q2 >= 0x8000000000000000ull) magu.a = 1;
+ q2 = 2*q2; // update q2
+ r2 = 2*r2 + 1; // update r2
+ }
+ delta = d - 1 - r2;
+ } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
+ magu.m = q2 + 1; // resulting magic number
+ magu.s = p - 64; // resulting shift
+ return magu;
+}
+
/// MaskedValueIsZero - Return true if 'Op & Mask' is known to be zero. We use
/// this predicate to simplify operations downstream. Op and Mask are known to
/// be the same type.
@@ -608,12 +783,23 @@
if (MaskedValueIsZero(N1, SignBit, TLI) &&
MaskedValueIsZero(N0, SignBit, TLI))
return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
+ // if integer divide is expensive and we satisfy the requirements, emit an
+ // alternate sequence.
+ // FIXME: This currently opts out powers of two, since targets can often be
+ // more clever in those cases. In an idea world, we would have some way to
+ // detect that too.
+ if (N1C && !isPowerOf2_64(N1C->getSignExtended()) &&
+ (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) &&
+ TLI.isOperationLegal(ISD::MULHS, VT) && TLI.isIntDivExpensive()) {
+ return BuildSDIV(N);
+ }
return SDOperand();
}
SDOperand DAGCombiner::visitUDIV(SDNode *N) {
SDOperand N0 = N->getOperand(0);
SDOperand N1 = N->getOperand(1);
+ MVT::ValueType VT = N->getValueType(0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
@@ -626,6 +812,10 @@
return DAG.getNode(ISD::SRL, N->getValueType(0), N0,
DAG.getConstant(Log2_64(N1C->getValue()),
TLI.getShiftAmountTy()));
+ // fold (udiv x, c) -> alternate
+ if (N1C && N1C->getValue() && TLI.isOperationLegal(ISD::MULHU, VT) &&
+ TLI.isIntDivExpensive())
+ return BuildUDIV(N);
return SDOperand();
}
@@ -2356,6 +2546,79 @@
return SDOperand();
}
+/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
+/// return a DAG expression to select that will generate the same value by
+/// multiplying by a magic number. See:
+/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
+SDOperand DAGCombiner::BuildSDIV(SDNode *N) {
+ MVT::ValueType VT = N->getValueType(0);
+ assert((VT == MVT::i32 || VT == MVT::i64) &&
+ "BuildSDIV only operates on i32 or i64!");
+
+ int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
+ ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
+
+ // Multiply the numerator (operand 0) by the magic value
+ SDOperand Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
+ DAG.getConstant(magics.m, VT));
+ // If d > 0 and m < 0, add the numerator
+ if (d > 0 && magics.m < 0) {
+ Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
+ WorkList.push_back(Q.Val);
+ }
+ // If d < 0 and m > 0, subtract the numerator.
+ if (d < 0 && magics.m > 0) {
+ Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
+ WorkList.push_back(Q.Val);
+ }
+ // Shift right algebraic if shift value is nonzero
+ if (magics.s > 0) {
+ Q = DAG.getNode(ISD::SRA, VT, Q,
+ DAG.getConstant(magics.s, TLI.getShiftAmountTy()));
+ WorkList.push_back(Q.Val);
+ }
+ // Extract the sign bit and add it to the quotient
+ SDOperand T =
+ DAG.getNode(ISD::SRL, MVT::i32, Q,
+ DAG.getConstant(MVT::getSizeInBits(VT)-1,
+ TLI.getShiftAmountTy()));
+ WorkList.push_back(T.Val);
+ return DAG.getNode(ISD::ADD, VT, Q, T);
+}
+
+/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
+/// return a DAG expression to select that will generate the same value by
+/// multiplying by a magic number. See:
+/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
+SDOperand DAGCombiner::BuildUDIV(SDNode *N) {
+ MVT::ValueType VT = N->getValueType(0);
+ assert((VT == MVT::i32 || VT == MVT::i64) &&
+ "BuildUDIV only operates on i32 or i64!");
+
+ uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
+ mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
+
+ // Multiply the numerator (operand 0) by the magic value
+ SDOperand Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
+ DAG.getConstant(magics.m, VT));
+ WorkList.push_back(Q.Val);
+
+ if (magics.a == 0) {
+ return DAG.getNode(ISD::SRL, VT, Q,
+ DAG.getConstant(magics.s, TLI.getShiftAmountTy()));
+ } else {
+ SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
+ WorkList.push_back(NPQ.Val);
+ NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
+ DAG.getConstant(1, TLI.getShiftAmountTy()));
+ WorkList.push_back(NPQ.Val);
+ NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
+ WorkList.push_back(NPQ.Val);
+ return DAG.getNode(ISD::SRL, VT, NPQ,
+ DAG.getConstant(magics.s-1, TLI.getShiftAmountTy()));
+ }
+}
+
// SelectionDAG::Combine - This is the entry point for the file.
//
void SelectionDAG::Combine(bool RunningAfterLegalize) {
More information about the llvm-commits
mailing list