[X86][SSE] Disable shouldFoldConstantShiftPairToMask for btver1/btver2 targets (PR40758)

As detailed on PR40758, Bobcat/Jaguar can perform vector immediate shifts on the same pipes as vector ANDs with the same latency - so it doesn't make sense to replace a shl+lshr with a shift+and pair as it requires an additional mask (with the extra constant pool, loading and register pressure costs).

Differential Revision: https://reviews.llvm.org/D61068

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@359293 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -6882,6 +6882,8 @@ SDValue DAGCombiner::visitSHL(SDNode *N) {
   //                               (and (srl x, (sub c1, c2), MASK)
   // Only fold this if the inner shift has no other uses -- if it does, folding
   // this will increase the total number of instructions.
+  // TODO - drop hasOneUse requirement if c1 == c2?
+  // TODO - support non-uniform vector shift amounts.
   if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
       TLI.shouldFoldConstantShiftPairToMask(N, Level)) {
     if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
@@ -7188,6 +7190,7 @@ SDValue DAGCombiner::visitSRL(SDNode *N) {
   }
 
   // fold (srl (shl x, c), c) -> (and x, cst2)
+  // TODO - (srl (shl x, c1), c2).
   if (N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
       isConstantOrConstantVector(N1, /* NoOpaques */ true)) {
     SDLoc DL(N);

lib/Target/X86/X86.td

@@ -424,6 +424,11 @@ def FeatureFastHorizontalOps
         "Prefer horizontal vector math instructions (haddp, phsub, etc.) over "
         "normal vector instructions with shuffles", [FeatureSSE3]>;
 
+def FeatureFastVectorShiftMasks
+    : SubtargetFeature<
+        "fast-vector-shift-masks", "HasFastVectorShiftMasks", "true",
+        "Prefer a left/right vector logical shift pair over a shift+and pair">;
+
 // Merge branches using three-way conditional code.
 def FeatureMergeToThreeWayBranch : SubtargetFeature<"merge-to-threeway-branch",
                                         "ThreewayBranchProfitable", "true",
@@ -775,7 +780,8 @@ def ProcessorFeatures {
                                                       FeaturePOPCNT,
                                                       FeatureSlowSHLD,
                                                       FeatureLAHFSAHF,
-                                                      FeatureFast15ByteNOP];
+                                                      FeatureFast15ByteNOP,
+                                                      FeatureFastVectorShiftMasks];
   list<SubtargetFeature> BtVer1Features = BtVer1InheritableFeatures;
 
   // Jaguar

lib/Target/X86/X86ISelLowering.cpp

@@ -5013,7 +5013,18 @@ bool X86TargetLowering::hasAndNot(SDValue Y) const {
 
 bool X86TargetLowering::shouldFoldConstantShiftPairToMask(
     const SDNode *N, CombineLevel Level) const {
-  // TODO - some targets prefer immediate vector shifts to shift+mask.
+  assert((N->getOpcode() == ISD::SHL &&
+          N->getOperand(0).getOpcode() == ISD::SRL) ||
+         (N->getOpcode() == ISD::SRL &&
+          N->getOperand(0).getOpcode() == ISD::SHL) &&
+             "Expected shift-shift mask");
+
+  if (Subtarget.hasFastVectorShiftMasks() && N->getValueType(0).isVector()) {
+    // Only fold if the shift values are equal - so it folds to AND.
+    // TODO - we should fold if either is non-uniform but we don't do the
+    // fold for non-splats yet.
+    return N->getOperand(1) == N->getOperand(0).getOperand(1);
+  }
   return TargetLoweringBase::shouldFoldConstantShiftPairToMask(N, Level);
 }
 

lib/Target/X86/X86Subtarget.h

@@ -393,6 +393,9 @@ protected:
   /// Try harder to combine to horizontal vector ops if they are fast.
   bool HasFastHorizontalOps = false;
 
+  /// Prefer a left/right vector logical shifts pair over a shift+and pair.
+  bool HasFastVectorShiftMasks = false;
+
   /// Use a retpoline thunk rather than indirect calls to block speculative
   /// execution.
   bool UseRetpolineIndirectCalls = false;
@@ -644,6 +647,7 @@ public:
   bool hasFastSHLDRotate() const { return HasFastSHLDRotate; }
   bool hasFastBEXTR() const { return HasFastBEXTR; }
   bool hasFastHorizontalOps() const { return HasFastHorizontalOps; }
+  bool hasFastVectorShiftMasks() const { return HasFastVectorShiftMasks; }
   bool hasMacroFusion() const { return HasMacroFusion; }
   bool hasBranchFusion() const { return HasBranchFusion; }
   bool hasERMSB() const { return HasERMSB; }

test/CodeGen/X86/sse2-vector-shifts.ll

@@ -1,5 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+sse2 | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+sse2 | FileCheck %s --check-prefixes=CHECK,MASK
+; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+sse2,+fast-vector-shift-masks | FileCheck %s --check-prefixes=CHECK,SHIFT
+; RUN: llc < %s -mtriple=x86_64-pc-linux -mcpu=btver1 | FileCheck %s --check-prefixes=CHECK,SHIFT
 
 ; SSE2 Logical Shift Left
 
@@ -300,11 +302,17 @@ define <4 x i32> @shl_sra_v4i32(<4 x i32> %x) nounwind {
 }
 
 define <4 x i32> @shl_srl_v4i32(<4 x i32> %x) nounwind {
-; CHECK-LABEL: shl_srl_v4i32:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    pslld $3, %xmm0
-; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
-; CHECK-NEXT:    retq
+; MASK-LABEL: shl_srl_v4i32:
+; MASK:       # %bb.0:
+; MASK-NEXT:    pslld $3, %xmm0
+; MASK-NEXT:    pand {{.*}}(%rip), %xmm0
+; MASK-NEXT:    retq
+;
+; SHIFT-LABEL: shl_srl_v4i32:
+; SHIFT:       # %bb.0:
+; SHIFT-NEXT:    psrld $2, %xmm0
+; SHIFT-NEXT:    pslld $5, %xmm0
+; SHIFT-NEXT:    retq
   %shl0 = lshr <4 x i32> %x, <i32 2, i32 2, i32 2, i32 2>
   %shl1 = shl <4 x i32> %shl0, <i32 5, i32 5, i32 5, i32 5>
   ret <4 x i32> %shl1