summaryrefslogtreecommitdiff
path: root/llvm/lib/Target/Hexagon/HexagonPatterns.td
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/lib/Target/Hexagon/HexagonPatterns.td')
-rw-r--r--llvm/lib/Target/Hexagon/HexagonPatterns.td70
1 files changed, 56 insertions, 14 deletions
diff --git a/llvm/lib/Target/Hexagon/HexagonPatterns.td b/llvm/lib/Target/Hexagon/HexagonPatterns.td
index cf711058823cd..cc10627955fb0 100644
--- a/llvm/lib/Target/Hexagon/HexagonPatterns.td
+++ b/llvm/lib/Target/Hexagon/HexagonPatterns.td
@@ -362,6 +362,16 @@ def Rol: pf2<rotl>;
// --(1) Immediate -------------------------------------------------------
//
+def Imm64Lo: SDNodeXForm<imm, [{
+ return CurDAG->getTargetConstant(int32_t (N->getSExtValue()),
+ SDLoc(N), MVT::i32);
+}]>;
+def Imm64Hi: SDNodeXForm<imm, [{
+ return CurDAG->getTargetConstant(int32_t (N->getSExtValue()>>32),
+ SDLoc(N), MVT::i32);
+}]>;
+
+
def SDTHexagonCONST32
: SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisPtrTy<0>]>;
@@ -389,7 +399,10 @@ def: Pat<(HexagonCP tconstpool:$A), (A2_tfrsi imm:$A)>;
def: Pat<(i1 0), (PS_false)>;
def: Pat<(i1 1), (PS_true)>;
-def: Pat<(i64 imm:$v), (CONST64 imm:$v)>;
+def: Pat<(i64 imm:$v), (CONST64 imm:$v)>,
+ Requires<[UseSmallData,NotOptTinyCore]>;
+def: Pat<(i64 imm:$v),
+ (Combinew (A2_tfrsi (Imm64Hi $v)), (A2_tfrsi (Imm64Lo $v)))>;
def ftoi : SDNodeXForm<fpimm, [{
APInt I = N->getValueAPF().bitcastToAPInt();
@@ -923,6 +936,13 @@ let AddedComplexity = 100 in {
defm: MinMax_pats<F2_sfmax, F2_sfmin, select, setole, i1, F32>;
}
+let AddedComplexity = 100, Predicates = [HasV67] in {
+ defm: MinMax_pats<F2_dfmin, F2_dfmax, select, setogt, i1, F64>;
+ defm: MinMax_pats<F2_dfmin, F2_dfmax, select, setoge, i1, F64>;
+ defm: MinMax_pats<F2_dfmax, F2_dfmin, select, setolt, i1, F64>;
+ defm: MinMax_pats<F2_dfmax, F2_dfmin, select, setole, i1, F64>;
+}
+
defm: MinMax_pats<A2_vminb, A2_vmaxb, vselect, setgt, v8i1, V8I8>;
defm: MinMax_pats<A2_vminb, A2_vmaxb, vselect, setge, v8i1, V8I8>;
defm: MinMax_pats<A2_vminh, A2_vmaxh, vselect, setgt, v4i1, V4I16>;
@@ -1075,7 +1095,7 @@ def Divu64_8: SDNodeXForm<imm, [{
// Special cases:
let AddedComplexity = 100 in {
def: Pat<(fshl I32:$Rs, I32:$Rt, (i32 16)),
- (A2_combine_hl I32:$Rs, I32:$Rt)>;
+ (A2_combine_lh I32:$Rs, I32:$Rt)>;
def: Pat<(fshl I64:$Rs, I64:$Rt, IsMul8_U3:$S),
(S2_valignib I64:$Rs, I64:$Rt, (Divu64_8 $S))>;
}
@@ -1109,7 +1129,7 @@ def FShr64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
// Special cases:
let AddedComplexity = 100 in {
def: Pat<(fshr I32:$Rs, I32:$Rt, (i32 16)),
- (A2_combine_hl I32:$Rs, I32:$Rt)>;
+ (A2_combine_lh I32:$Rs, I32:$Rt)>;
def: Pat<(fshr I64:$Rs, I64:$Rt, IsMul8_U3:$S),
(S2_valignib I64:$Rs, I64:$Rt, (Divu8 $S))>;
}
@@ -1231,7 +1251,7 @@ class OpshIRI_pat<InstHexagon MI, PatFrag Op, PatFrag ShOp,
: Pat<(Op anyimm:$u8, (ShOp RegPred:$Rs, ImmPred:$U5)),
(MI anyimm:$u8, RegPred:$Rs, imm:$U5)>;
-let AddedComplexity = 200 in {
+let AddedComplexity = 200, Predicates = [UseCompound] in {
def: OpshIRI_pat<S4_addi_asl_ri, Add, Su<Shl>, I32, u5_0ImmPred>;
def: OpshIRI_pat<S4_addi_lsr_ri, Add, Su<Srl>, I32, u5_0ImmPred>;
def: OpshIRI_pat<S4_subi_asl_ri, Sub, Su<Shl>, I32, u5_0ImmPred>;
@@ -1408,6 +1428,26 @@ let Predicates = [HasV66] in {
def: OpR_RR_pat<F2_dfsub, pf2<fsub>, f64, F64>;
}
+def DfMpy: OutPatFrag<(ops node:$Rs, node:$Rt),
+ (F2_dfmpyhh
+ (F2_dfmpylh
+ (F2_dfmpylh
+ (F2_dfmpyll $Rs, $Rt),
+ $Rs, $Rt),
+ $Rt, $Rs),
+ $Rs, $Rt)>;
+
+let Predicates = [HasV67,UseUnsafeMath], AddedComplexity = 50 in {
+ def: Pat<(fmul F64:$Rs, F64:$Rt), (DfMpy $Rs, $Rt)>;
+}
+let Predicates = [HasV67] in {
+ def: OpR_RR_pat<F2_dfmin, pf2<fminnum>, f64, F64>;
+ def: OpR_RR_pat<F2_dfmax, pf2<fmaxnum>, f64, F64>;
+
+ def: Pat<(fmul F64:$Rs, F64:$Rt), (DfMpy (F2_dfmpyfix $Rs, $Rt),
+ (F2_dfmpyfix $Rt, $Rs))>;
+}
+
// In expressions like a0*b0 + a1*b1 + ..., prefer to generate multiply-add,
// over add-add with individual multiplies as inputs.
let AddedComplexity = 10 in {
@@ -1510,7 +1550,7 @@ let AddedComplexity = 110 in { // greater than S2_asl_r_r_and/or/xor.
// S4_addaddi and S4_subaddi don't have tied operands, so give them
// a bit of preference.
-let AddedComplexity = 30 in {
+let AddedComplexity = 30, Predicates = [UseCompound] in {
def: Pat<(add I32:$Rs, (Su<Add> I32:$Ru, anyimm:$s6)),
(S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
def: Pat<(add anyimm:$s6, (Su<Add> I32:$Rs, I32:$Ru)),
@@ -1523,8 +1563,10 @@ let AddedComplexity = 30 in {
(S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
}
+let Predicates = [UseCompound] in
def: Pat<(or I32:$Ru, (Su<And> I32:$Rx, anyimm:$s10)),
(S4_or_andix IntRegs:$Ru, IntRegs:$Rx, imm:$s10)>;
+
def: Pat<(or I32:$Rx, (Su<And> I32:$Rs, anyimm:$s10)),
(S4_or_andi IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
def: Pat<(or I32:$Rx, (Su<Or> I32:$Rs, anyimm:$s10)),
@@ -1625,7 +1667,7 @@ def : Pat <(mulhs I64:$Rss, I64:$Rtt),
// will put the immediate addend into a register, while these instructions will
// use it directly. Such a construct does not appear in the middle of a gep,
// where M2_macsip would be preferable.
-let AddedComplexity = 20 in {
+let AddedComplexity = 20, Predicates = [UseCompound] in {
def: Pat<(add (Su<Mul> I32:$Rs, u6_0ImmPred:$U6), anyimm:$u6),
(M4_mpyri_addi imm:$u6, IntRegs:$Rs, imm:$U6)>;
def: Pat<(add (Su<Mul> I32:$Rs, I32:$Rt), anyimm:$u6),
@@ -1633,13 +1675,14 @@ let AddedComplexity = 20 in {
}
// Keep these instructions less preferable to M2_macsip/M2_macsin.
-def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, u6_2ImmPred:$u6_2)),
- (M4_mpyri_addr_u2 IntRegs:$Ru, imm:$u6_2, IntRegs:$Rs)>;
-def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, anyimm:$u6)),
- (M4_mpyri_addr IntRegs:$Ru, IntRegs:$Rs, imm:$u6)>;
-def: Pat<(add I32:$Ru, (Su<Mul> I32:$Ry, I32:$Rs)),
- (M4_mpyrr_addr IntRegs:$Ru, IntRegs:$Ry, IntRegs:$Rs)>;
-
+let Predicates = [UseCompound] in {
+ def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, u6_2ImmPred:$u6_2)),
+ (M4_mpyri_addr_u2 IntRegs:$Ru, imm:$u6_2, IntRegs:$Rs)>;
+ def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, anyimm:$u6)),
+ (M4_mpyri_addr IntRegs:$Ru, IntRegs:$Rs, imm:$u6)>;
+ def: Pat<(add I32:$Ru, (Su<Mul> I32:$Ry, I32:$Rs)),
+ (M4_mpyrr_addr IntRegs:$Ru, IntRegs:$Ry, IntRegs:$Rs)>;
+}
def: Pat<(fma F32:$Rs, F32:$Rt, F32:$Rx),
(F2_sffma F32:$Rx, F32:$Rs, F32:$Rt)>;
@@ -1648,7 +1691,6 @@ def: Pat<(fma (fneg F32:$Rs), F32:$Rt, F32:$Rx),
def: Pat<(fma F32:$Rs, (fneg F32:$Rt), F32:$Rx),
(F2_sffms F32:$Rx, F32:$Rs, F32:$Rt)>;
-
def: Pat<(mul V2I32:$Rs, V2I32:$Rt),
(PS_vmulw V2I32:$Rs, V2I32:$Rt)>;
def: Pat<(add V2I32:$Rx, (mul V2I32:$Rs, V2I32:$Rt)),
generated by cgit v1.2.3 (git 2.25.1) at 2025-07-04 20:10:40 +0000