Smarter isel of ldrsb / ldrsh. Only make use of these when [r,r] address is feasible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75360 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/ARMISelDAGToDAG.cpp

@@ -454,16 +454,11 @@ bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue Op, SDValue N,
   // FIXME dl should come from the parent load or store, not the address
   DebugLoc dl = Op.getDebugLoc();
   if (N.getOpcode() != ISD::ADD) {
-    Base = N;
+    ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
-    // We must materialize a zero in a reg! Returning a constant here
+    if (!NC || NC->getZExtValue() != 0)
-    // wouldn't work without additional code to position the node within
+      return false;
-    // ISel's topological ordering in a place where ISel will process it
+
-    // normally.  Instead, just explicitly issue a tMOVri8 node!
+    Base = Offset = N;
-    SDValue CC = CurDAG->getRegister(ARM::CPSR, MVT::i32);
-    SDValue Pred = CurDAG->getTargetConstant(0xEULL, MVT::i32);
-    SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
-    SDValue Ops[] = { CC, CurDAG->getTargetConstant(0, MVT::i32), Pred, PredReg };
-    Offset = SDValue(CurDAG->getTargetNode(ARM::tMOVi8, dl, MVT::i32, Ops,4),0);
     return true;
   }
 

lib/Target/ARM/ARMInstrThumb.td

@@ -224,10 +224,12 @@ def tLDRH : T1pI2<(outs tGPR:$dst), (ins t_addrmode_s2:$addr),
                 "ldrh", " $dst, $addr",
                 [(set tGPR:$dst, (zextloadi16 t_addrmode_s2:$addr))]>;
 
+let AddedComplexity = 10 in
 def tLDRSB : T1pI1<(outs tGPR:$dst), (ins t_addrmode_rr:$addr),
                  "ldrsb", " $dst, $addr",
                  [(set tGPR:$dst, (sextloadi8 t_addrmode_rr:$addr))]>;
 
+let AddedComplexity = 10 in
 def tLDRSH : T1pI2<(outs tGPR:$dst), (ins t_addrmode_rr:$addr),
                  "ldrsh", " $dst, $addr",
                  [(set tGPR:$dst, (sextloadi16 t_addrmode_rr:$addr))]>;
@@ -620,6 +622,14 @@ def : T1Pat<(extloadi1  t_addrmode_s1:$addr),  (tLDRB t_addrmode_s1:$addr)>;
 def : T1Pat<(extloadi8  t_addrmode_s1:$addr),  (tLDRB t_addrmode_s1:$addr)>;
 def : T1Pat<(extloadi16 t_addrmode_s2:$addr),  (tLDRH t_addrmode_s2:$addr)>;
 
+// If it's possible to use [r,r] address mode for sextload, select to
+// ldr{b|h} + sxt{b|h} instead.
+def : TPat<(sextloadi8 t_addrmode_s1:$addr),
+           (tSXTB (tLDRB t_addrmode_s1:$addr))>;
+def : TPat<(sextloadi16 t_addrmode_s2:$addr),
+           (tSXTH (tLDRH t_addrmode_s2:$addr))>;
+
+
 // Large immediate handling.
 
 // Two piece imms.

lib/Target/ARM/README-Thumb.txt

@@ -244,12 +244,3 @@ to toggle the 's' bit since they do not set CPSR when they are inside IT blocks.
 Make use of hi register variants of cmp: tCMPhir / tCMPZhir.
 
 //===---------------------------------------------------------------------===//
-
-Rather than generating ldrsb, sometimes it's better to select to ldrb + sxtb.
-The problem is ldrsb addressing mode [r, r] means the zero offset requires an
-extra move. e.g. ldr_ext.ll test3:
-	movs r1, #0
-	ldrsb r0, [r0, r1]
-=>
-	ldrb	r0, [r0, #0]
-	sxtb	r0, r0

test/CodeGen/Thumb/ldr_ext.ll

@@ -1,27 +1,34 @@
-; RUN: llvm-as < %s | llc -march=thumb | grep ldrb | count 1
+; RUN: llvm-as < %s | llc -march=thumb | FileCheck %s
-; RUN: llvm-as < %s | llc -march=thumb | grep ldrh | count 1
-; RUN: llvm-as < %s | llc -march=thumb | grep ldrsb | count 1
-; RUN: llvm-as < %s | llc -march=thumb | grep ldrsh | count 1
 
 define i32 @test1(i8* %v.pntr.s0.u1) {
+; CHECK: test1:
+; CHECK: ldrb
     %tmp.u = load i8* %v.pntr.s0.u1
     %tmp1.s = zext i8 %tmp.u to i32
     ret i32 %tmp1.s
 }
 
 define i32 @test2(i16* %v.pntr.s0.u1) {
+; CHECK: test2:
+; CHECK: ldrh
     %tmp.u = load i16* %v.pntr.s0.u1
     %tmp1.s = zext i16 %tmp.u to i32
     ret i32 %tmp1.s
 }
 
 define i32 @test3(i8* %v.pntr.s1.u0) {
+; CHECK: test3:
+; CHECK: ldrb 
+; CHECK: sxtb
    %tmp.s = load i8* %v.pntr.s1.u0
     %tmp1.s = sext i8 %tmp.s to i32
     ret i32 %tmp1.s
 }
 
 define i32 @test4() {
+; CHECK: test4:
+; CHECK: movs
+; CHECK: ldrsh 
     %tmp.s = load i16* null
     %tmp1.s = sext i16 %tmp.s to i32
     ret i32 %tmp1.s