[X86][SelectionDAGBuilder] Move the hack for handling MMX shift by i32 intrinsics into the X86 backend.

This intrinsics should be shift by immediate, but gcc allows any
i32 scalar and clang needs to match that. So we try to detect the
non-constant case and move the data from an integer register to an
MMX register.

Previously this was done by creating a v2i32 build_vector and
bitcast in SelectionDAGBuilder. This had to be done early since
v2i32 isn't a legal type. The bitcast+build_vector would be DAG
combined to X86ISD::MMX_MOVW2D which isel will turn into a
GPR->MMX MOVD.

This commit just moves the whole thing to lowering and emits
the X86ISD::MMX_MOVW2D directly to avoid the illegal type. The
test changes just seem to be due to nodes being linearized in a
different order.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@372535 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -5969,65 +5969,6 @@ void SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I,
   case Intrinsic::masked_compressstore:
     visitMaskedStore(I, true /* IsCompressing */);
     return;
-  case Intrinsic::x86_mmx_pslli_w:
-  case Intrinsic::x86_mmx_pslli_d:
-  case Intrinsic::x86_mmx_pslli_q:
-  case Intrinsic::x86_mmx_psrli_w:
-  case Intrinsic::x86_mmx_psrli_d:
-  case Intrinsic::x86_mmx_psrli_q:
-  case Intrinsic::x86_mmx_psrai_w:
-  case Intrinsic::x86_mmx_psrai_d: {
-    SDValue ShAmt = getValue(I.getArgOperand(1));
-    if (isa<ConstantSDNode>(ShAmt)) {
-      visitTargetIntrinsic(I, Intrinsic);
-      return;
-    }
-    unsigned NewIntrinsic = 0;
-    EVT ShAmtVT = MVT::v2i32;
-    switch (Intrinsic) {
-    case Intrinsic::x86_mmx_pslli_w:
-      NewIntrinsic = Intrinsic::x86_mmx_psll_w;
-      break;
-    case Intrinsic::x86_mmx_pslli_d:
-      NewIntrinsic = Intrinsic::x86_mmx_psll_d;
-      break;
-    case Intrinsic::x86_mmx_pslli_q:
-      NewIntrinsic = Intrinsic::x86_mmx_psll_q;
-      break;
-    case Intrinsic::x86_mmx_psrli_w:
-      NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
-      break;
-    case Intrinsic::x86_mmx_psrli_d:
-      NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
-      break;
-    case Intrinsic::x86_mmx_psrli_q:
-      NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
-      break;
-    case Intrinsic::x86_mmx_psrai_w:
-      NewIntrinsic = Intrinsic::x86_mmx_psra_w;
-      break;
-    case Intrinsic::x86_mmx_psrai_d:
-      NewIntrinsic = Intrinsic::x86_mmx_psra_d;
-      break;
-    default: llvm_unreachable("Impossible intrinsic");  // Can't reach here.
-    }
-
-    // The vector shift intrinsics with scalars uses 32b shift amounts but
-    // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
-    // to be zero.
-    // We must do this early because v2i32 is not a legal type.
-    SDValue ShOps[2];
-    ShOps[0] = ShAmt;
-    ShOps[1] = DAG.getConstant(0, sdl, MVT::i32);
-    ShAmt =  DAG.getBuildVector(ShAmtVT, sdl, ShOps);
-    EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
-    ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt);
-    Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT,
-                       DAG.getConstant(NewIntrinsic, sdl, MVT::i32),
-                       getValue(I.getArgOperand(0)), ShAmt);
-    setValue(&I, Res);
-    return;
-  }
   case Intrinsic::powi:
     setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)),
                             getValue(I.getArgOperand(1)), DAG));

lib/Target/X86/X86ISelLowering.cpp

@@ -23738,6 +23738,58 @@ SDValue X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
                                                  MaskVT, Operation);
     return DAG.getMergeValues({Result0, Result1}, DL);
   }
+  case Intrinsic::x86_mmx_pslli_w:
+  case Intrinsic::x86_mmx_pslli_d:
+  case Intrinsic::x86_mmx_pslli_q:
+  case Intrinsic::x86_mmx_psrli_w:
+  case Intrinsic::x86_mmx_psrli_d:
+  case Intrinsic::x86_mmx_psrli_q:
+  case Intrinsic::x86_mmx_psrai_w:
+  case Intrinsic::x86_mmx_psrai_d: {
+    SDValue ShAmt = Op.getOperand(2);
+    // If the argument is a constant, this is fine.
+    if (isa<ConstantSDNode>(ShAmt))
+      return Op;
+
+    unsigned NewIntrinsic;
+    switch (IntNo) {
+    default: llvm_unreachable("Impossible intrinsic");  // Can't reach here.
+    case Intrinsic::x86_mmx_pslli_w:
+      NewIntrinsic = Intrinsic::x86_mmx_psll_w;
+      break;
+    case Intrinsic::x86_mmx_pslli_d:
+      NewIntrinsic = Intrinsic::x86_mmx_psll_d;
+      break;
+    case Intrinsic::x86_mmx_pslli_q:
+      NewIntrinsic = Intrinsic::x86_mmx_psll_q;
+      break;
+    case Intrinsic::x86_mmx_psrli_w:
+      NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
+      break;
+    case Intrinsic::x86_mmx_psrli_d:
+      NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
+      break;
+    case Intrinsic::x86_mmx_psrli_q:
+      NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
+      break;
+    case Intrinsic::x86_mmx_psrai_w:
+      NewIntrinsic = Intrinsic::x86_mmx_psra_w;
+      break;
+    case Intrinsic::x86_mmx_psrai_d:
+      NewIntrinsic = Intrinsic::x86_mmx_psra_d;
+      break;
+    }
+
+    // The vector shift intrinsics with scalars uses 32b shift amounts but
+    // the sse2/mmx shift instructions reads 64 bits. Copy the 32 bits to an
+    // MMX register.
+    SDLoc DL(Op);
+    ShAmt = DAG.getNode(X86ISD::MMX_MOVW2D, DL, MVT::x86mmx, ShAmt);
+    return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(),
+                       DAG.getConstant(NewIntrinsic, DL, MVT::i32),
+                       Op.getOperand(1), ShAmt);
+
+  }
   }
 }
 

test/CodeGen/X86/bitcast-mmx.ll

@@ -34,10 +34,10 @@ define i64 @t1(i64 %x, i32 %n) nounwind {
 ; X86-NEXT:    movl %esp, %ebp
 ; X86-NEXT:    andl $-8, %esp
 ; X86-NEXT:    subl $8, %esp
-; X86-NEXT:    movd 16(%ebp), %mm0
-; X86-NEXT:    movq 8(%ebp), %mm1
-; X86-NEXT:    psllq %mm0, %mm1
-; X86-NEXT:    movq %mm1, (%esp)
+; X86-NEXT:    movq 8(%ebp), %mm0
+; X86-NEXT:    movd 16(%ebp), %mm1
+; X86-NEXT:    psllq %mm1, %mm0
+; X86-NEXT:    movq %mm0, (%esp)
 ; X86-NEXT:    movl (%esp), %eax
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
 ; X86-NEXT:    movl %ebp, %esp
@@ -46,10 +46,10 @@ define i64 @t1(i64 %x, i32 %n) nounwind {
 ;
 ; X64-LABEL: t1:
 ; X64:       # %bb.0: # %entry
-; X64-NEXT:    movd %esi, %mm0
-; X64-NEXT:    movq %rdi, %mm1
-; X64-NEXT:    psllq %mm0, %mm1
-; X64-NEXT:    movq %mm1, %rax
+; X64-NEXT:    movq %rdi, %mm0
+; X64-NEXT:    movd %esi, %mm1
+; X64-NEXT:    psllq %mm1, %mm0
+; X64-NEXT:    movq %mm0, %rax
 ; X64-NEXT:    retq
 entry:
   %0 = bitcast i64 %x to x86_mmx
@@ -65,11 +65,11 @@ define i64 @t2(i64 %x, i32 %n, i32 %w) nounwind {
 ; X86-NEXT:    movl %esp, %ebp
 ; X86-NEXT:    andl $-8, %esp
 ; X86-NEXT:    subl $8, %esp
-; X86-NEXT:    movd 16(%ebp), %mm0
-; X86-NEXT:    movd 20(%ebp), %mm1
-; X86-NEXT:    psllq %mm0, %mm1
-; X86-NEXT:    por 8(%ebp), %mm1
-; X86-NEXT:    movq %mm1, (%esp)
+; X86-NEXT:    movd 20(%ebp), %mm0
+; X86-NEXT:    movd 16(%ebp), %mm1
+; X86-NEXT:    psllq %mm1, %mm0
+; X86-NEXT:    por 8(%ebp), %mm0
+; X86-NEXT:    movq %mm0, (%esp)
 ; X86-NEXT:    movl (%esp), %eax
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
 ; X86-NEXT:    movl %ebp, %esp
@@ -78,12 +78,12 @@ define i64 @t2(i64 %x, i32 %n, i32 %w) nounwind {
 ;
 ; X64-LABEL: t2:
 ; X64:       # %bb.0: # %entry
-; X64-NEXT:    movd %esi, %mm0
-; X64-NEXT:    movd %edx, %mm1
-; X64-NEXT:    psllq %mm0, %mm1
-; X64-NEXT:    movq %rdi, %mm0
-; X64-NEXT:    por %mm1, %mm0
-; X64-NEXT:    movq %mm0, %rax
+; X64-NEXT:    movd %edx, %mm0
+; X64-NEXT:    movd %esi, %mm1
+; X64-NEXT:    psllq %mm1, %mm0
+; X64-NEXT:    movq %rdi, %mm1
+; X64-NEXT:    por %mm0, %mm1
+; X64-NEXT:    movq %mm1, %rax
 ; X64-NEXT:    retq
 entry:
   %0 = insertelement <2 x i32> undef, i32 %w, i32 0

test/CodeGen/X86/mmx-fold-load.ll

@@ -585,22 +585,22 @@ define void @test_psrlq_by_volatile_shift_amount(x86_mmx* %t) nounwind {
 ; X86-NEXT:    pushl %eax
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    movl $1, (%esp)
-; X86-NEXT:    movd (%esp), %mm0
 ; X86-NEXT:    movl $255, %ecx
-; X86-NEXT:    movd %ecx, %mm1
-; X86-NEXT:    psrlq %mm0, %mm1
-; X86-NEXT:    movq %mm1, (%eax)
+; X86-NEXT:    movd %ecx, %mm0
+; X86-NEXT:    movd (%esp), %mm1
+; X86-NEXT:    psrlq %mm1, %mm0
+; X86-NEXT:    movq %mm0, (%eax)
 ; X86-NEXT:    popl %eax
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_psrlq_by_volatile_shift_amount:
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    movl $1, -{{[0-9]+}}(%rsp)
-; X64-NEXT:    movd -{{[0-9]+}}(%rsp), %mm0
 ; X64-NEXT:    movl $255, %eax
-; X64-NEXT:    movd %eax, %mm1
-; X64-NEXT:    psrlq %mm0, %mm1
-; X64-NEXT:    movq %mm1, (%rdi)
+; X64-NEXT:    movd %eax, %mm0
+; X64-NEXT:    movd -{{[0-9]+}}(%rsp), %mm1
+; X64-NEXT:    psrlq %mm1, %mm0
+; X64-NEXT:    movq %mm0, (%rdi)
 ; X64-NEXT:    retq
 entry:
   %0 = alloca i32, align 4