[InstCombine] Moved SSE vector shift constant folding into its own helper function...

This will make some upcoming bugfixes + improvements easier to manage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@243962 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineCalls.cpp b/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 8083e6a72e73b915b4c2f336dbf45b424722b253..e09c00253513f4508cd481c15c6041f98aa72eae 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -194,12 +194,44 @@ Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
     return MI;
   }
 
-  return nullptr;
-}
-
-static Value *SimplifyX86extend(const IntrinsicInst &II,
-                                InstCombiner::BuilderTy &Builder,
-                                bool SignExtend) {
+  return nullptr;\r
+}\r
+\r
+static Value *SimplifyX86immshift(const IntrinsicInst &II,\r
+                                  InstCombiner::BuilderTy &Builder,\r
+                                  bool ShiftLeft) {\r
+  // Simplify if count is constant. To 0 if >= BitWidth,\r
+  // otherwise to shl/lshr.\r
+  auto CDV = dyn_cast<ConstantDataVector>(II.getArgOperand(1));\r
+  auto CInt = dyn_cast<ConstantInt>(II.getArgOperand(1));\r
+  if (!CDV && !CInt)\r
+    return nullptr;\r
+  ConstantInt *Count;\r
+  if (CDV)\r
+    Count = cast<ConstantInt>(CDV->getElementAsConstant(0));\r
+  else\r
+    Count = CInt;\r
+\r
+  auto Vec = II.getArgOperand(0);\r
+  auto VT = cast<VectorType>(Vec->getType());\r
+  auto SVT = VT->getElementType();\r
+  if (Count->getZExtValue() > (SVT->getPrimitiveSizeInBits() - 1))\r
+    return ConstantAggregateZero::get(VT);\r
+\r
+  unsigned VWidth = VT->getNumElements();\r
+\r
+  // Get a constant vector of the same type as the first operand.\r
+  auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());\r
+\r
+  if (ShiftLeft)\r
+    return Builder.CreateShl(Vec, Builder.CreateVectorSplat(VWidth, VTCI));\r
+\r
+  return Builder.CreateLShr(Vec, Builder.CreateVectorSplat(VWidth, VTCI));\r
+}\r
+\r
+static Value *SimplifyX86extend(const IntrinsicInst &II,\r
+                                InstCombiner::BuilderTy &Builder,\r
+                                bool SignExtend) {\r
   VectorType *SrcTy = cast<VectorType>(II.getArgOperand(0)->getType());
   VectorType *DstTy = cast<VectorType>(II.getType());
   unsigned NumDstElts = DstTy->getNumElements();
@@ -718,86 +750,46 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
       II->setArgOperand(0, V);
       return II;
     }
-    break;
-  }
-
-  // Constant fold <A x Bi> << Ci.
-  // FIXME: We don't handle _dq because it's a shift of an i128, but is
-  // represented in the IR as <2 x i64>. A per element shift is wrong.
-  case Intrinsic::x86_sse2_psll_d:
-  case Intrinsic::x86_sse2_psll_q:
-  case Intrinsic::x86_sse2_psll_w:
+    break;\r
+  }\r
+\r
+  // Constant fold lshr( <A x Bi>, Ci ).\r
+  case Intrinsic::x86_sse2_psrl_d:\r
+  case Intrinsic::x86_sse2_psrl_q:\r
+  case Intrinsic::x86_sse2_psrl_w:\r
+  case Intrinsic::x86_sse2_psrli_d:\r
+  case Intrinsic::x86_sse2_psrli_q:\r
+  case Intrinsic::x86_sse2_psrli_w:\r
+  case Intrinsic::x86_avx2_psrl_d:\r
+  case Intrinsic::x86_avx2_psrl_q:\r
+  case Intrinsic::x86_avx2_psrl_w:\r
+  case Intrinsic::x86_avx2_psrli_d:\r
+  case Intrinsic::x86_avx2_psrli_q:\r
+  case Intrinsic::x86_avx2_psrli_w:\r
+    if (Value *V = SimplifyX86immshift(*II, *Builder, false))\r
+      return ReplaceInstUsesWith(*II, V);\r
+    break;\r
+\r
+  // Constant fold shl( <A x Bi>, Ci ).\r
+  case Intrinsic::x86_sse2_psll_d:\r
+  case Intrinsic::x86_sse2_psll_q:\r
+  case Intrinsic::x86_sse2_psll_w:\r
   case Intrinsic::x86_sse2_pslli_d:
   case Intrinsic::x86_sse2_pslli_q:
   case Intrinsic::x86_sse2_pslli_w:
   case Intrinsic::x86_avx2_psll_d:
   case Intrinsic::x86_avx2_psll_q:
   case Intrinsic::x86_avx2_psll_w:
-  case Intrinsic::x86_avx2_pslli_d:
-  case Intrinsic::x86_avx2_pslli_q:
-  case Intrinsic::x86_avx2_pslli_w:
-  case Intrinsic::x86_sse2_psrl_d:
-  case Intrinsic::x86_sse2_psrl_q:
-  case Intrinsic::x86_sse2_psrl_w:
-  case Intrinsic::x86_sse2_psrli_d:
-  case Intrinsic::x86_sse2_psrli_q:
-  case Intrinsic::x86_sse2_psrli_w:
-  case Intrinsic::x86_avx2_psrl_d:
-  case Intrinsic::x86_avx2_psrl_q:
-  case Intrinsic::x86_avx2_psrl_w:
-  case Intrinsic::x86_avx2_psrli_d:
-  case Intrinsic::x86_avx2_psrli_q:
-  case Intrinsic::x86_avx2_psrli_w: {
-    // Simplify if count is constant. To 0 if >= BitWidth,
-    // otherwise to shl/lshr.
-    auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
-    auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
-    if (!CDV && !CInt)
-      break;
-    ConstantInt *Count;
-    if (CDV)
-      Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
-    else
-      Count = CInt;
-
-    auto Vec = II->getArgOperand(0);
-    auto VT = cast<VectorType>(Vec->getType());
-    if (Count->getZExtValue() >
-        VT->getElementType()->getPrimitiveSizeInBits() - 1)
-      return ReplaceInstUsesWith(
-          CI, ConstantAggregateZero::get(Vec->getType()));
-
-    bool isPackedShiftLeft = true;
-    switch (II->getIntrinsicID()) {
-    default : break;
-    case Intrinsic::x86_sse2_psrl_d:
-    case Intrinsic::x86_sse2_psrl_q:
-    case Intrinsic::x86_sse2_psrl_w:
-    case Intrinsic::x86_sse2_psrli_d:
-    case Intrinsic::x86_sse2_psrli_q:
-    case Intrinsic::x86_sse2_psrli_w:
-    case Intrinsic::x86_avx2_psrl_d:
-    case Intrinsic::x86_avx2_psrl_q:
-    case Intrinsic::x86_avx2_psrl_w:
-    case Intrinsic::x86_avx2_psrli_d:
-    case Intrinsic::x86_avx2_psrli_q:
-    case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
-    }
-
-    unsigned VWidth = VT->getNumElements();
-    // Get a constant vector of the same type as the first operand.
-    auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
-    if (isPackedShiftLeft)
-      return BinaryOperator::CreateShl(Vec,
-          Builder->CreateVectorSplat(VWidth, VTCI));
-
-    return BinaryOperator::CreateLShr(Vec,
-        Builder->CreateVectorSplat(VWidth, VTCI));
-  }
-
-  case Intrinsic::x86_sse41_pmovsxbd:
-  case Intrinsic::x86_sse41_pmovsxbq:
-  case Intrinsic::x86_sse41_pmovsxbw:
+  case Intrinsic::x86_avx2_pslli_d:\r
+  case Intrinsic::x86_avx2_pslli_q:\r
+  case Intrinsic::x86_avx2_pslli_w:\r
+    if (Value *V = SimplifyX86immshift(*II, *Builder, true))\r
+      return ReplaceInstUsesWith(*II, V);\r
+    break;\r
+\r
+  case Intrinsic::x86_sse41_pmovsxbd:\r
+  case Intrinsic::x86_sse41_pmovsxbq:\r
+  case Intrinsic::x86_sse41_pmovsxbw:\r
   case Intrinsic::x86_sse41_pmovsxdq:
   case Intrinsic::x86_sse41_pmovsxwd:
   case Intrinsic::x86_sse41_pmovsxwq: