Hexagon: Expand cttz, ctlz, and ctpop for now.

[oota-llvm.git] / lib / Target / Hexagon / HexagonInstrInfo.td

diff --git a/lib/Target/Hexagon/HexagonInstrInfo.td b/lib/Target/Hexagon/HexagonInstrInfo.td
index 11c01672f0c1874005f144e7bf329f93f8ca8cbd..082772a042eb79461a02a7e686d6a0dba2f60a3a 100644 (file)
--- a/lib/Target/Hexagon/HexagonInstrInfo.td
+++ b/lib/Target/Hexagon/HexagonInstrInfo.td
@@ -808,7 +808,7 @@ let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC],
 // JR +
 //===----------------------------------------------------------------------===//
 def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone,
-                               [SDNPHasChain, SDNPOptInGlue]>;
+                               [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
 
 // Jump to address from register.
 let isPredicable =1, isReturn = 1, isTerminator = 1, isBarrier = 1,
@@ -1885,7 +1885,7 @@ def SDHexagonBARRIER: SDTypeProfile<0, 0, []>;
 def HexagonBARRIER: SDNode<"HexagonISD::BARRIER", SDHexagonBARRIER,
                            [SDNPHasChain]>;
 
-let hasSideEffects = 1, isHexagonSolo = 1 in
+let hasSideEffects = 1, isSolo = 1 in
 def BARRIER : SYSInst<(outs), (ins),
                      "barrier",
                      [(HexagonBARRIER)]>;
@@ -1960,9 +1960,9 @@ def LOOP0_r : CRInst<(outs), (ins brtarget:$offset, IntRegs:$src2),
 
 let isBranch = 1, isTerminator = 1, neverHasSideEffects = 1,
     Defs = [PC, LC0], Uses = [SA0, LC0] in {
-def ENDLOOP0 : Marker<(outs), (ins brtarget:$offset),
-                      ":endloop0",
-                      []>;
+def ENDLOOP0 : Endloop<(outs), (ins brtarget:$offset),
+                       ":endloop0",
+                       []>;
 }
 
 // Support for generating global address.
@@ -2825,23 +2825,42 @@ def : Pat <(i32 (zext (i1 PredRegs:$src1))),
 
 // i1 -> i64
 def : Pat <(i64 (zext (i1 PredRegs:$src1))),
-      (i64 (COMBINE_rr (TFRI 0), (MUX_ii (i1 PredRegs:$src1), 1, 0)))>;
+      (i64 (COMBINE_rr (TFRI 0), (MUX_ii (i1 PredRegs:$src1), 1, 0)))>,
+      Requires<[NoV4T]>;
 
 // i32 -> i64
 def : Pat <(i64 (zext (i32 IntRegs:$src1))),
-      (i64 (COMBINE_rr (TFRI 0), (i32 IntRegs:$src1)))>;
+      (i64 (COMBINE_rr (TFRI 0), (i32 IntRegs:$src1)))>,
+      Requires<[NoV4T]>;
 
 // i8 -> i64
 def:  Pat <(i64 (zextloadi8 ADDRriS11_0:$src1)),
-      (i64 (COMBINE_rr (TFRI 0), (LDriub ADDRriS11_0:$src1)))>;
+      (i64 (COMBINE_rr (TFRI 0), (LDriub ADDRriS11_0:$src1)))>,
+      Requires<[NoV4T]>;
+
+let AddedComplexity = 20 in
+def:  Pat <(i64 (zextloadi8 (add (i32 IntRegs:$src1),
+                                s11_0ExtPred:$offset))),
+      (i64 (COMBINE_rr (TFRI 0), (LDriub_indexed IntRegs:$src1,
+                                  s11_0ExtPred:$offset)))>,
+      Requires<[NoV4T]>;
 
 // i16 -> i64
 def:  Pat <(i64 (zextloadi16 ADDRriS11_1:$src1)),
-      (i64 (COMBINE_rr (TFRI 0), (LDriuh ADDRriS11_1:$src1)))>;
+      (i64 (COMBINE_rr (TFRI 0), (LDriuh ADDRriS11_1:$src1)))>,
+      Requires<[NoV4T]>;
+
+let AddedComplexity = 20 in
+def:  Pat <(i64 (zextloadi16 (add (i32 IntRegs:$src1),
+                                  s11_1ExtPred:$offset))),
+      (i64 (COMBINE_rr (TFRI 0), (LDriuh_indexed IntRegs:$src1,
+                                  s11_1ExtPred:$offset)))>,
+      Requires<[NoV4T]>;
 
 // i32 -> i64
 def:  Pat <(i64 (zextloadi32 ADDRriS11_2:$src1)),
-      (i64 (COMBINE_rr (TFRI 0), (LDriw ADDRriS11_2:$src1)))>;
+      (i64 (COMBINE_rr (TFRI 0), (LDriw ADDRriS11_2:$src1)))>,
+      Requires<[NoV4T]>;
 
 def:  Pat <(i32 (zextloadi1 ADDRriS11_0:$src1)),
       (i32 (LDriw ADDRriS11_0:$src1))>;
@@ -2862,15 +2881,41 @@ def : Pat <(i64 (anyext (i1 PredRegs:$src1))),
 // Any extended 64-bit load.
 // anyext i32 -> i64
 def:  Pat <(i64 (extloadi32 ADDRriS11_2:$src1)),
-      (i64 (COMBINE_rr (TFRI 0), (LDriw ADDRriS11_2:$src1)))>;
+      (i64 (COMBINE_rr (TFRI 0), (LDriw ADDRriS11_2:$src1)))>,
+      Requires<[NoV4T]>;
+
+// When there is an offset we should prefer the pattern below over the pattern above.
+// The complexity of the above is 13 (gleaned from HexagonGenDAGIsel.inc)
+// So this complexity below is comfortably higher to allow for choosing the below.
+// If this is not done then we generate addresses such as
+// ********************************************
+//        r1 = add (r0, #4)
+//        r1 = memw(r1 + #0)
+//  instead of
+//        r1 = memw(r0 + #4)
+// ********************************************
+let AddedComplexity = 100 in
+def:  Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
+      (i64 (COMBINE_rr (TFRI 0), (LDriw_indexed IntRegs:$src1,
+                                  s11_2ExtPred:$offset)))>,
+      Requires<[NoV4T]>;
 
 // anyext i16 -> i64.
 def:  Pat <(i64 (extloadi16 ADDRriS11_2:$src1)),
-      (i64 (COMBINE_rr (TFRI 0), (LDrih ADDRriS11_2:$src1)))>;
+      (i64 (COMBINE_rr (TFRI 0), (LDrih ADDRriS11_2:$src1)))>,
+      Requires<[NoV4T]>;
+
+let AddedComplexity = 20 in
+def:  Pat <(i64 (extloadi16 (add (i32 IntRegs:$src1),
+                                  s11_1ExtPred:$offset))),
+      (i64 (COMBINE_rr (TFRI 0), (LDrih_indexed IntRegs:$src1,
+                                  s11_1ExtPred:$offset)))>,
+      Requires<[NoV4T]>;
 
 // Map from Rdd = zxtw(Rs) -> Rdd = combine(0, Rs).
 def : Pat<(i64 (zext (i32 IntRegs:$src1))),
-      (i64 (COMBINE_rr (TFRI 0), (i32 IntRegs:$src1)))>;
+      (i64 (COMBINE_rr (TFRI 0), (i32 IntRegs:$src1)))>,
+      Requires<[NoV4T]>;
 
 // Multiply 64-bit unsigned and use upper result.
 def : Pat <(mulhu (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2)),