[X86][SSE] Keep 32-bit target i64 vector shifts on SSE unit.

This patch improves the 32-bit target i64 constant matching to detect the shuffle vector splats that are introduced by i64 vector shift vectorization (D8416).

Differential Revision: http://reviews.llvm.org/D11327

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@243577 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index cea2aa73c23030468a414245a648925278e133bc..906f0464d84c92b3d2be22c6ce9e8d4e6dca89c4 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -17307,34 +17307,50 @@ static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
 
   // Special case in 32-bit mode, where i64 is expanded into high and low parts.
   if (!Subtarget->is64Bit() &&
-      (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
-      Amt.getOpcode() == ISD::BITCAST &&
-      Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
+      (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
+
+    // Peek through any splat that was introduced for i64 shift vectorization.
+    int SplatIndex = -1;
+    if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
+      if (SVN->isSplat()) {
+        SplatIndex = SVN->getSplatIndex();
+        Amt = Amt.getOperand(0);
+        assert(SplatIndex < (int)VT.getVectorNumElements() &&
+               "Splat shuffle referencing second operand");
+      }
+
+    if (Amt.getOpcode() != ISD::BITCAST ||
+        Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
+      return SDValue();
+
     Amt = Amt.getOperand(0);
     unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
                      VT.getVectorNumElements();
     unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
     uint64_t ShiftAmt = 0;
+    unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
     for (unsigned i = 0; i != Ratio; ++i) {
-      ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
+      ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
       if (!C)
         return SDValue();
       // 6 == Log2(64)
       ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
     }
-    // Check remaining shift amounts.
-    for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
-      uint64_t ShAmt = 0;
-      for (unsigned j = 0; j != Ratio; ++j) {
-        ConstantSDNode *C =
-          dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
-        if (!C)
+
+    // Check remaining shift amounts (if not a splat).
+    if (SplatIndex < 0) {
+      for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
+        uint64_t ShAmt = 0;
+        for (unsigned j = 0; j != Ratio; ++j) {
+          ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
+          if (!C)
+            return SDValue();
+          // 6 == Log2(64)
+          ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
+        }
+        if (ShAmt != ShiftAmt)
           return SDValue();
-        // 6 == Log2(64)
-        ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
       }
-      if (ShAmt != ShiftAmt)
-        return SDValue();
     }
 
     if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))

diff --git a/test/CodeGen/X86/vector-shift-ashr-128.ll b/test/CodeGen/X86/vector-shift-ashr-128.ll
index fff788a2298e3aa1a507404835337cec573c7636..64d4ea03762d7617b8d30caa5fec0136377bf544 100644 (file)
--- a/test/CodeGen/X86/vector-shift-ashr-128.ll
+++ b/test/CodeGen/X86/vector-shift-ashr-128.ll
@@ -877,21 +877,17 @@ define <2 x i64> @constant_shift_v2i64(<2 x i64> %a) nounwind {
 ;
 ; X32-SSE-LABEL: constant_shift_v2i64:
 ; X32-SSE:       # BB#0:
-; X32-SSE-NEXT:    movl $7, %eax
-; X32-SSE-NEXT:    movd %eax, %xmm2
 ; X32-SSE-NEXT:    movdqa {{.*#+}} xmm1 = [0,2147483648,0,2147483648]
-; X32-SSE-NEXT:    movdqa %xmm1, %xmm3
-; X32-SSE-NEXT:    psrlq %xmm2, %xmm3
-; X32-SSE-NEXT:    movl $1, %eax
-; X32-SSE-NEXT:    movd %eax, %xmm4
-; X32-SSE-NEXT:    psrlq %xmm4, %xmm1
-; X32-SSE-NEXT:    movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]
+; X32-SSE-NEXT:    movdqa %xmm1, %xmm2
+; X32-SSE-NEXT:    psrlq $7, %xmm2
+; X32-SSE-NEXT:    psrlq $1, %xmm1
+; X32-SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]
 ; X32-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X32-SSE-NEXT:    psrlq %xmm2, %xmm1
-; X32-SSE-NEXT:    psrlq %xmm4, %xmm0
+; X32-SSE-NEXT:    psrlq $7, %xmm1
+; X32-SSE-NEXT:    psrlq $1, %xmm0
 ; X32-SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
-; X32-SSE-NEXT:    xorpd %xmm3, %xmm1
-; X32-SSE-NEXT:    psubq %xmm3, %xmm1
+; X32-SSE-NEXT:    xorpd %xmm2, %xmm1
+; X32-SSE-NEXT:    psubq %xmm2, %xmm1
 ; X32-SSE-NEXT:    movdqa %xmm1, %xmm0
 ; X32-SSE-NEXT:    retl
   %shift = ashr <2 x i64> %a, <i64 1, i64 7>

diff --git a/test/CodeGen/X86/vector-shift-lshr-128.ll b/test/CodeGen/X86/vector-shift-lshr-128.ll
index 7ee0412b6adb9c68db9f772fb65408201b8f7a0c..1c988946a4672eb21de4c26f3ba19fe75de63440 100644 (file)
--- a/test/CodeGen/X86/vector-shift-lshr-128.ll
+++ b/test/CodeGen/X86/vector-shift-lshr-128.ll
@@ -655,13 +655,9 @@ define <2 x i64> @constant_shift_v2i64(<2 x i64> %a) nounwind {
 ;
 ; X32-SSE-LABEL: constant_shift_v2i64:
 ; X32-SSE:       # BB#0:
-; X32-SSE-NEXT:    movl $7, %eax
-; X32-SSE-NEXT:    movd %eax, %xmm2
 ; X32-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X32-SSE-NEXT:    psrlq %xmm2, %xmm1
-; X32-SSE-NEXT:    movl $1, %eax
-; X32-SSE-NEXT:    movd %eax, %xmm2
-; X32-SSE-NEXT:    psrlq %xmm2, %xmm0
+; X32-SSE-NEXT:    psrlq $7, %xmm1
+; X32-SSE-NEXT:    psrlq $1, %xmm0
 ; X32-SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
 ; X32-SSE-NEXT:    movapd %xmm1, %xmm0
 ; X32-SSE-NEXT:    retl

diff --git a/test/CodeGen/X86/vector-shift-shl-128.ll b/test/CodeGen/X86/vector-shift-shl-128.ll
index facbcc14c2c0d72678f48cfdc7e807b3cef4c776..fa6533d39cb75cfd8adfde2281a1f3693acea0bb 100644 (file)
--- a/test/CodeGen/X86/vector-shift-shl-128.ll
+++ b/test/CodeGen/X86/vector-shift-shl-128.ll
@@ -607,13 +607,9 @@ define <2 x i64> @constant_shift_v2i64(<2 x i64> %a) nounwind {
 ;
 ; X32-SSE-LABEL: constant_shift_v2i64:
 ; X32-SSE:       # BB#0:
-; X32-SSE-NEXT:    movl $7, %eax
-; X32-SSE-NEXT:    movd %eax, %xmm2
 ; X32-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X32-SSE-NEXT:    psllq %xmm2, %xmm1
-; X32-SSE-NEXT:    movl $1, %eax
-; X32-SSE-NEXT:    movd %eax, %xmm2
-; X32-SSE-NEXT:    psllq %xmm2, %xmm0
+; X32-SSE-NEXT:    psllq $7, %xmm1
+; X32-SSE-NEXT:    psllq $1, %xmm0
 ; X32-SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
 ; X32-SSE-NEXT:    movapd %xmm1, %xmm0
 ; X32-SSE-NEXT:    retl