Unbreak getOnesVector() / getZeroVector() to use valid ARM extended imm's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81262 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 5d8ee7e8542d01f7cd3760d2828b672537b16ca7..827e62bb8152e898857c4f759b1e7d6555465935 100644 (file)
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -2086,14 +2086,19 @@ static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
   // will be implemented with the NEON VNEG instruction.  However, VNEG does
   // not support i64 elements, so sometimes the zero vectors will need to be
   // explicitly constructed.  For those cases, and potentially other uses in
-  // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted
+  // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
   // to their dest type.  This ensures they get CSE'd.
   SDValue Vec;
-  SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
-  if (VT.getSizeInBits() == 64)
-    Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
-  else
-    Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
+  SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
+  SmallVector<SDValue, 8> Ops;
+  MVT TVT;
+
+  if (VT.getSizeInBits() == 64) {
+    Ops.assign(8, Cst); TVT = MVT::v8i8;
+  } else {
+    Ops.assign(16, Cst); TVT = MVT::v16i8;
+  }
+  Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
 
   return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
 }
@@ -2103,14 +2108,19 @@ static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
   assert(VT.isVector() && "Expected a vector type");
 
-  // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
-  // type.  This ensures they get CSE'd.
+  // Always build ones vectors as <16 x i32> or <8 x i32> bitcasted to their
+  // dest type. This ensures they get CSE'd.
   SDValue Vec;
-  SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
-  if (VT.getSizeInBits() == 64)
-    Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
-  else
-    Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
+  SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
+  SmallVector<SDValue, 8> Ops;
+  MVT TVT;
+
+  if (VT.getSizeInBits() == 64) {
+    Ops.assign(8, Cst); TVT = MVT::v8i8;
+  } else {
+    Ops.assign(16, Cst); TVT = MVT::v16i8;
+  }
+  Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
 
   return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
 }

diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 7241e633ac426226331ec79693e8aed42bf5d246..e9181231609b7eb05d01106ee6e2f6c792e80cc3 100644 (file)
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -1600,21 +1600,25 @@ def  VORRq    : N3VQ<0, 0, 0b10, 0b0001, 1, "vorr", v4i32, v4i32, or, 1>;
 def  VBICd    : N3V<0, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
                     (ins DPR:$src1, DPR:$src2), NoItinerary,
                     "vbic\t$dst, $src1, $src2", "",
-                    [(set DPR:$dst, (v2i32 (and DPR:$src1,(vnot DPR:$src2))))]>;
+                    [(set DPR:$dst, (v2i32 (and DPR:$src1,
+                                                (vnot_conv DPR:$src2))))]>;
 def  VBICq    : N3V<0, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
                     (ins QPR:$src1, QPR:$src2), NoItinerary,
                     "vbic\t$dst, $src1, $src2", "",
-                    [(set QPR:$dst, (v4i32 (and QPR:$src1,(vnot QPR:$src2))))]>;
+                    [(set QPR:$dst, (v4i32 (and QPR:$src1,
+                                                (vnot_conv QPR:$src2))))]>;
 
 //   VORN     : Vector Bitwise OR NOT
 def  VORNd    : N3V<0, 0, 0b11, 0b0001, 0, 1, (outs DPR:$dst),
                     (ins DPR:$src1, DPR:$src2), NoItinerary,
                     "vorn\t$dst, $src1, $src2", "",
-                    [(set DPR:$dst, (v2i32 (or DPR:$src1, (vnot DPR:$src2))))]>;
+                    [(set DPR:$dst, (v2i32 (or DPR:$src1,
+                                               (vnot_conv DPR:$src2))))]>;
 def  VORNq    : N3V<0, 0, 0b11, 0b0001, 1, 1, (outs QPR:$dst),
                     (ins QPR:$src1, QPR:$src2), NoItinerary,
                     "vorn\t$dst, $src1, $src2", "",
-                    [(set QPR:$dst, (v4i32 (or QPR:$src1, (vnot QPR:$src2))))]>;
+                    [(set QPR:$dst, (v4i32 (or QPR:$src1,
+                                               (vnot_conv QPR:$src2))))]>;
 
 //   VMVN     : Vector Bitwise NOT
 def  VMVNd    : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 0, 0,
@@ -1634,13 +1638,13 @@ def  VBSLd    : N3V<1, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
                     "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
                     [(set DPR:$dst,
                       (v2i32 (or (and DPR:$src2, DPR:$src1),
-                                 (and DPR:$src3, (vnot DPR:$src1)))))]>;
+                                 (and DPR:$src3, (vnot_conv DPR:$src1)))))]>;
 def  VBSLq    : N3V<1, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
                     (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
                     "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
                     [(set QPR:$dst,
                       (v4i32 (or (and QPR:$src2, QPR:$src1),
-                                 (and QPR:$src3, (vnot QPR:$src1)))))]>;
+                                 (and QPR:$src3, (vnot_conv QPR:$src1)))))]>;
 
 //   VBIF     : Vector Bitwise Insert if False
 //              like VBSL but with: "vbif\t$dst, $src3, $src1", "$src2 = $dst",

diff --git a/test/CodeGen/ARM/2009-09-09-AllOnes.ll b/test/CodeGen/ARM/2009-09-09-AllOnes.ll
new file mode 100644 (file)
index 0000000..f654a16
--- /dev/null
+++ b/test/CodeGen/ARM/2009-09-09-AllOnes.ll
@@ -0,0 +1,10 @@
+; RUN: llc -mattr=+neon < %s
+target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:32-f32:32:32-f64:32:32-v64:64:64-v128:128:128-a0:0:32"
+target triple = "thumbv7-elf"
+
+define arm_apcscc void @foo() {
+entry:
+  %0 = insertelement <4 x i32> undef, i32 -1, i32 3
+  store <4 x i32> %0, <4 x i32>* undef, align 16
+  unreachable
+}