Hexagon: Add V4 combine instructions and some more Def Pats for V2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@174331 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/Hexagon/HexagonISelLowering.h

@@ -52,6 +52,8 @@ namespace llvm {
       WrapperCP,
       WrapperCombineII,
       WrapperCombineRR,
+      WrapperCombineRI_V4,
+      WrapperCombineIR_V4,
       WrapperPackhl,
       WrapperSplatB,
       WrapperSplatH,

lib/Target/Hexagon/HexagonInstrInfo.td

@@ -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)),

lib/Target/Hexagon/HexagonInstrInfoV4.td

@@ -280,6 +280,19 @@ def COMBINE_Ir_V4 : ALU32_ir<(outs DoubleRegs:$dst),
             []>,
             Requires<[HasV4T]>;
 
+def HexagonWrapperCombineRI_V4 :
+  SDNode<"HexagonISD::WrapperCombineRI_V4", SDTHexagonI64I32I32>;
+def HexagonWrapperCombineIR_V4 :
+  SDNode<"HexagonISD::WrapperCombineIR_V4", SDTHexagonI64I32I32>;
+
+def : Pat <(HexagonWrapperCombineRI_V4 IntRegs:$r, s8ExtPred:$i),
+           (COMBINE_rI_V4 IntRegs:$r, s8ExtPred:$i)>,
+          Requires<[HasV4T]>;
+
+def : Pat <(HexagonWrapperCombineIR_V4 s8ExtPred:$i, IntRegs:$r),
+           (COMBINE_Ir_V4 s8ExtPred:$i, IntRegs:$r)>,
+          Requires<[HasV4T]>;
+
 let isExtendable = 1, opExtendable = 2, isExtentSigned = 0, opExtentBits = 6,
     neverHasSideEffects = 1, validSubTargets = HasV4SubT in
 def COMBINE_iI_V4 : ALU32_ii<(outs DoubleRegs:$dst),
@@ -1143,6 +1156,73 @@ def : Pat <(i32 (load (add (HexagonCONST32_GP tglobaladdr:$global),
                       u16ImmPred:$offset))),
            (i32 (LDriw_GP_V4 tglobaladdr:$global, u16ImmPred:$offset))>,
             Requires<[HasV4T]>;
+// zext i1->i64
+def : Pat <(i64 (zext (i1 PredRegs:$src1))),
+      (i64 (COMBINE_Ir_V4 0, (MUX_ii (i1 PredRegs:$src1), 1, 0)))>,
+      Requires<[HasV4T]>;
+
+// zext i32->i64
+def : Pat <(i64 (zext (i32 IntRegs:$src1))),
+      (i64 (COMBINE_Ir_V4 0, (i32 IntRegs:$src1)))>,
+      Requires<[HasV4T]>;
+// zext i8->i64
+def:  Pat <(i64 (zextloadi8 ADDRriS11_0:$src1)),
+      (i64 (COMBINE_Ir_V4 0, (LDriub ADDRriS11_0:$src1)))>,
+      Requires<[HasV4T]>;
+
+let AddedComplexity = 20 in
+def:  Pat <(i64 (zextloadi8 (add (i32 IntRegs:$src1),
+                                s11_0ExtPred:$offset))),
+      (i64 (COMBINE_Ir_V4 0, (LDriub_indexed IntRegs:$src1,
+                                  s11_0ExtPred:$offset)))>,
+      Requires<[HasV4T]>;
+
+// zext i16->i64
+def:  Pat <(i64 (zextloadi16 ADDRriS11_1:$src1)),
+      (i64 (COMBINE_Ir_V4 0, (LDriuh ADDRriS11_1:$src1)))>,
+      Requires<[HasV4T]>;
+
+let AddedComplexity = 20 in
+def:  Pat <(i64 (zextloadi16 (add (i32 IntRegs:$src1),
+                                  s11_1ExtPred:$offset))),
+      (i64 (COMBINE_Ir_V4 0, (LDriuh_indexed IntRegs:$src1,
+                                  s11_1ExtPred:$offset)))>,
+      Requires<[HasV4T]>;
+
+// anyext i16->i64
+def:  Pat <(i64 (extloadi16 ADDRriS11_2:$src1)),
+      (i64 (COMBINE_Ir_V4 0, (LDrih ADDRriS11_2:$src1)))>,
+      Requires<[HasV4T]>;
+
+let AddedComplexity = 20 in
+def:  Pat <(i64 (extloadi16 (add (i32 IntRegs:$src1),
+                                  s11_1ExtPred:$offset))),
+      (i64 (COMBINE_Ir_V4 0, (LDrih_indexed IntRegs:$src1,
+                                  s11_1ExtPred:$offset)))>,
+      Requires<[HasV4T]>;
+
+// zext i32->i64
+def:  Pat <(i64 (zextloadi32 ADDRriS11_2:$src1)),
+      (i64 (COMBINE_Ir_V4 0, (LDriw ADDRriS11_2:$src1)))>,
+      Requires<[HasV4T]>;
+
+let AddedComplexity = 100 in
+def:  Pat <(i64 (zextloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
+      (i64 (COMBINE_Ir_V4 0, (LDriw_indexed IntRegs:$src1,
+                                  s11_2ExtPred:$offset)))>,
+      Requires<[HasV4T]>;
+
+// anyext i32->i64
+def:  Pat <(i64 (extloadi32 ADDRriS11_2:$src1)),
+      (i64 (COMBINE_Ir_V4 0, (LDriw ADDRriS11_2:$src1)))>,
+      Requires<[HasV4T]>;
+
+let AddedComplexity = 100 in
+def:  Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
+      (i64 (COMBINE_Ir_V4 0, (LDriw_indexed IntRegs:$src1,
+                                  s11_2ExtPred:$offset)))>,
+      Requires<[HasV4T]>;
+
 
 
 //===----------------------------------------------------------------------===//

test/CodeGen/Hexagon/combine_ir.ll

@@ -0,0 +1,55 @@
+; RUN: llc -march=hexagon -mcpu=hexagonv4 < %s | FileCheck %s
+; CHECK: word
+; CHECK: combine(#0
+
+define void @word(i32* nocapture %a) nounwind {
+entry:
+  %0 = load i32* %a, align 4, !tbaa !0
+  %1 = zext i32 %0 to i64
+  %add.ptr = getelementptr inbounds i32* %a, i32 1
+  %2 = load i32* %add.ptr, align 4, !tbaa !0
+  %3 = zext i32 %2 to i64
+  %4 = shl nuw i64 %3, 32
+  %ins = or i64 %4, %1
+  tail call void @bar(i64 %ins) nounwind
+  ret void
+}
+
+declare void @bar(i64)
+
+; CHECK: halfword
+; CHECK: combine(#0
+
+define void @halfword(i16* nocapture %a) nounwind {
+entry:
+  %0 = load i16* %a, align 2, !tbaa !3
+  %1 = zext i16 %0 to i64
+  %add.ptr = getelementptr inbounds i16* %a, i32 1
+  %2 = load i16* %add.ptr, align 2, !tbaa !3
+  %3 = zext i16 %2 to i64
+  %4 = shl nuw nsw i64 %3, 16
+  %ins = or i64 %4, %1
+  tail call void @bar(i64 %ins) nounwind
+  ret void
+}
+
+; CHECK: byte
+; CHECK: combine(#0
+
+define void @byte(i8* nocapture %a) nounwind {
+entry:
+  %0 = load i8* %a, align 1, !tbaa !1
+  %1 = zext i8 %0 to i64
+  %add.ptr = getelementptr inbounds i8* %a, i32 1
+  %2 = load i8* %add.ptr, align 1, !tbaa !1
+  %3 = zext i8 %2 to i64
+  %4 = shl nuw nsw i64 %3, 8
+  %ins = or i64 %4, %1
+  tail call void @bar(i64 %ins) nounwind
+  ret void
+}
+
+!0 = metadata !{metadata !"int", metadata !1}
+!1 = metadata !{metadata !"omnipotent char", metadata !2}
+!2 = metadata !{metadata !"Simple C/C++ TBAA"}
+!3 = metadata !{metadata !"short", metadata !1}

test/CodeGen/Hexagon/struct_args.ll

@@ -1,5 +1,5 @@
 ; RUN: llc -march=hexagon -mcpu=hexagonv4 < %s | FileCheck %s
-; CHECK: r{{[0-9]}}:{{[0-9]}} = combine(r{{[0-9]}}, r{{[0-9]}})
+; CHECK: r{{[0-9]}}:{{[0-9]}} = combine({{r[0-9]|#0}}, r{{[0-9]}})
 ; CHECK: r{{[0-9]}}:{{[0-9]}} |= asl(r{{[0-9]}}:{{[0-9]}}, #32)
 
 %struct.small = type { i32, i32 }