[llvm] r344649 - [X86] Match (cmp (and (shr X, C), mask), 0) to BEXTR+TEST.
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 16 15:29:36 PDT 2018
Author: ctopper
Date: Tue Oct 16 15:29:36 2018
New Revision: 344649
URL: http://llvm.org/viewvc/llvm-project?rev=344649&view=rev
Log:
[X86] Match (cmp (and (shr X, C), mask), 0) to BEXTR+TEST.
Without this we match the CMP+AND to a TEST and then match the SHR separately. I'm trusting analyzeCompare to remove the TEST during the peephole pass. Otherwise we need to check the flag users to see if they only use the Z flag.
This recovers a case lost by r344270.
Differential Revision: https://reviews.llvm.org/D53310
Modified:
llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp
llvm/trunk/test/CodeGen/X86/tbm_patterns.ll
Modified: llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp?rev=344649&r1=344648&r2=344649&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp Tue Oct 16 15:29:36 2018
@@ -454,7 +454,7 @@ namespace {
}
bool foldLoadStoreIntoMemOperand(SDNode *Node);
- bool matchBEXTRFromAndImm(SDNode *Node);
+ MachineSDNode *matchBEXTRFromAndImm(SDNode *Node);
bool matchBEXTR(SDNode *Node);
bool shrinkAndImmediate(SDNode *N);
bool isMaskZeroExtended(SDNode *N) const;
@@ -2714,7 +2714,7 @@ bool X86DAGToDAGISel::matchBEXTR(SDNode
}
// See if this is an (X >> C1) & C2 that we can match to BEXTR/BEXTRI.
-bool X86DAGToDAGISel::matchBEXTRFromAndImm(SDNode *Node) {
+MachineSDNode *X86DAGToDAGISel::matchBEXTRFromAndImm(SDNode *Node) {
MVT NVT = Node->getSimpleValueType(0);
SDLoc dl(Node);
@@ -2729,30 +2729,30 @@ bool X86DAGToDAGISel::matchBEXTRFromAndI
// BEXTR?
if (!Subtarget->hasTBM() &&
!(Subtarget->hasBMI() && Subtarget->hasFastBEXTR()))
- return false;
+ return nullptr;
// Must have a shift right.
if (N0->getOpcode() != ISD::SRL && N0->getOpcode() != ISD::SRA)
- return false;
+ return nullptr;
// Shift can't have additional users.
if (!N0->hasOneUse())
- return false;
+ return nullptr;
// Only supported for 32 and 64 bits.
if (NVT != MVT::i32 && NVT != MVT::i64)
- return false;
+ return nullptr;
// Shift amount and RHS of and must be constant.
ConstantSDNode *MaskCst = dyn_cast<ConstantSDNode>(N1);
ConstantSDNode *ShiftCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
if (!MaskCst || !ShiftCst)
- return false;
+ return nullptr;
// And RHS must be a mask.
uint64_t Mask = MaskCst->getZExtValue();
if (!isMask_64(Mask))
- return false;
+ return nullptr;
uint64_t Shift = ShiftCst->getZExtValue();
uint64_t MaskSize = countPopulation(Mask);
@@ -2760,12 +2760,12 @@ bool X86DAGToDAGISel::matchBEXTRFromAndI
// Don't interfere with something that can be handled by extracting AH.
// TODO: If we are able to fold a load, BEXTR might still be better than AH.
if (Shift == 8 && MaskSize == 8)
- return false;
+ return nullptr;
// Make sure we are only using bits that were in the original value, not
// shifted in.
if (Shift + MaskSize > NVT.getSizeInBits())
- return false;
+ return nullptr;
SDValue New = CurDAG->getTargetConstant(Shift | (MaskSize << 8), dl, NVT);
unsigned ROpc = NVT == MVT::i64 ? X86::BEXTRI64ri : X86::BEXTRI32ri;
@@ -2794,9 +2794,7 @@ bool X86DAGToDAGISel::matchBEXTRFromAndI
NewNode = CurDAG->getMachineNode(ROpc, dl, NVT, Input, New);
}
- ReplaceUses(SDValue(Node, 0), SDValue(NewNode, 0));
- CurDAG->RemoveDeadNode(Node);
- return true;
+ return NewNode;
}
// Emit a PCMISTR(I/M) instruction.
@@ -3106,8 +3104,11 @@ void X86DAGToDAGISel::Select(SDNode *Nod
break;
case ISD::AND:
- if (matchBEXTRFromAndImm(Node))
+ if (MachineSDNode *NewNode = matchBEXTRFromAndImm(Node)) {
+ ReplaceUses(SDValue(Node, 0), SDValue(NewNode, 0));
+ CurDAG->RemoveDeadNode(Node);
return;
+ }
if (matchBEXTR(Node))
return;
if (AndImmShrink && shrinkAndImmediate(Node))
@@ -3551,6 +3552,22 @@ void X86DAGToDAGISel::Select(SDNode *Nod
// Save the original VT of the compare.
MVT CmpVT = N0.getSimpleValueType();
+ // If we are comparing (and (shr X, C, Mask) with 0, emit a BEXTR followed
+ // by a test instruction. The test should be removed later by
+ // analyzeCompare if we are using only the zero flag.
+ // TODO: Should we check the users and use the BEXTR flags directly?
+ if (isNullConstant(N1) && N0.getOpcode() == ISD::AND && N0.hasOneUse()) {
+ if (MachineSDNode *NewNode = matchBEXTRFromAndImm(N0.getNode())) {
+ unsigned TestOpc = CmpVT == MVT::i64 ? X86::TEST64rr
+ : X86::TEST32rr;
+ SDValue BEXTR = SDValue(NewNode, 0);
+ NewNode = CurDAG->getMachineNode(TestOpc, dl, MVT::i32, BEXTR, BEXTR);
+ ReplaceUses(SDValue(Node, 0), SDValue(NewNode, 0));
+ CurDAG->RemoveDeadNode(Node);
+ return;
+ }
+ }
+
// We can peek through truncates, but we need to be careful below.
if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse())
N0 = N0.getOperand(0);
@@ -3561,7 +3578,7 @@ void X86DAGToDAGISel::Select(SDNode *Nod
if (N0.getOpcode() == ISD::AND &&
N0.getNode()->hasOneUse() &&
N0.getValueType() != MVT::i8 &&
- X86::isZeroNode(N1)) {
+ isNullConstant(N1)) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
if (!C) break;
uint64_t Mask = C->getZExtValue();
Modified: llvm/trunk/test/CodeGen/X86/tbm_patterns.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/tbm_patterns.ll?rev=344649&r1=344648&r2=344649&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/tbm_patterns.ll (original)
+++ llvm/trunk/test/CodeGen/X86/tbm_patterns.ll Tue Oct 16 15:29:36 2018
@@ -53,8 +53,7 @@ define i32 @test_x86_tbm_bextri_u32_z2(i
; CHECK-LABEL: test_x86_tbm_bextri_u32_z2:
; CHECK: # %bb.0:
; CHECK-NEXT: movl %esi, %eax
-; CHECK-NEXT: shrl $4, %edi
-; CHECK-NEXT: testl $4095, %edi # imm = 0xFFF
+; CHECK-NEXT: bextrl $3076, %edi, %ecx # imm = 0xC04
; CHECK-NEXT: cmovnel %edx, %eax
; CHECK-NEXT: retq
%t0 = lshr i32 %a, 4
@@ -114,8 +113,7 @@ define i64 @test_x86_tbm_bextri_u64_z2(i
; CHECK-LABEL: test_x86_tbm_bextri_u64_z2:
; CHECK: # %bb.0:
; CHECK-NEXT: movq %rsi, %rax
-; CHECK-NEXT: shrl $4, %edi
-; CHECK-NEXT: testl $4095, %edi # imm = 0xFFF
+; CHECK-NEXT: bextrl $3076, %edi, %ecx # imm = 0xC04
; CHECK-NEXT: cmovneq %rdx, %rax
; CHECK-NEXT: retq
%t0 = lshr i64 %a, 4
More information about the llvm-commits
mailing list