[llvm] [SelectionDAG] Add space-optimized forms of OPC_EmitNode/OPC_MorphNodeTo (PR #73502)

Wang Pengcheng via llvm-commits llvm-commits at lists.llvm.org
Mon Nov 27 03:18:12 PST 2023


https://github.com/wangpc-pp updated https://github.com/llvm/llvm-project/pull/73502

>From 86b6964ddf3b19da61b6a8563bc3d5d9b52e89df Mon Sep 17 00:00:00 2001
From: wangpc <wangpengcheng.pp at bytedance.com>
Date: Mon, 27 Nov 2023 19:14:10 +0800
Subject: [PATCH] [SelectionDAG] Add space-optimized forms of
 OPC_EmitNode/OPC_MorphNodeTo

If there is only one bit set in EmitNodeInfo, then we can encode it
implicitly to save one byte.

Overall this reduces the llc binary size with all in-tree targets by
about 168K.
---
 llvm/include/llvm/CodeGen/SelectionDAGISel.h  | 31 ++++++
 .../CodeGen/SelectionDAG/SelectionDAGISel.cpp | 99 +++++++++++++++++--
 llvm/test/TableGen/DAGDefaultOps.td           | 10 +-
 .../TableGen/dag-isel-regclass-emit-enum.td   |  4 +-
 llvm/test/TableGen/dag-isel-res-order.td      |  2 +-
 llvm/utils/TableGen/DAGISelMatcherEmitter.cpp | 52 +++++++---
 6 files changed, 173 insertions(+), 25 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/SelectionDAGISel.h b/llvm/include/llvm/CodeGen/SelectionDAGISel.h
index aa71be5d1960ff5..4876d04f71203b5 100644
--- a/llvm/include/llvm/CodeGen/SelectionDAGISel.h
+++ b/llvm/include/llvm/CodeGen/SelectionDAGISel.h
@@ -211,6 +211,37 @@ class SelectionDAGISel : public MachineFunctionPass {
     OPC_MorphNodeTo0,
     OPC_MorphNodeTo1,
     OPC_MorphNodeTo2,
+    // Space-optimized forms that implicitly encode EmitNodeInfo.
+    OPC_EmitNode0None,
+    OPC_EmitNode1None,
+    OPC_EmitNode2None,
+    OPC_MorphNodeTo0None,
+    OPC_MorphNodeTo1None,
+    OPC_MorphNodeTo2None,
+    OPC_EmitNode0Chain,
+    OPC_EmitNode1Chain,
+    OPC_EmitNode2Chain,
+    OPC_MorphNodeTo0Chain,
+    OPC_MorphNodeTo1Chain,
+    OPC_MorphNodeTo2Chain,
+    OPC_EmitNode0GlueInput,
+    OPC_EmitNode1GlueInput,
+    OPC_EmitNode2GlueInput,
+    OPC_MorphNodeTo0GlueInput,
+    OPC_MorphNodeTo1GlueInput,
+    OPC_MorphNodeTo2GlueInput,
+    OPC_EmitNode0GlueOutput,
+    OPC_EmitNode1GlueOutput,
+    OPC_EmitNode2GlueOutput,
+    OPC_MorphNodeTo0GlueOutput,
+    OPC_MorphNodeTo1GlueOutput,
+    OPC_MorphNodeTo2GlueOutput,
+    OPC_EmitNode0MemRefs,
+    OPC_EmitNode1MemRefs,
+    OPC_EmitNode2MemRefs,
+    OPC_MorphNodeTo0MemRefs,
+    OPC_MorphNodeTo1MemRefs,
+    OPC_MorphNodeTo2MemRefs,
     OPC_CompleteMatch,
     // Contains offset in table for pattern being selected
     OPC_Coverage
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
index a0af6faa7fbcefb..deef5c25b1238b3 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
@@ -3641,20 +3641,97 @@ void SelectionDAGISel::SelectCodeCommon(SDNode *NodeToMatch,
       continue;
     }
 
-    case OPC_EmitNode:     case OPC_MorphNodeTo:
-    case OPC_EmitNode0:    case OPC_EmitNode1:    case OPC_EmitNode2:
-    case OPC_MorphNodeTo0: case OPC_MorphNodeTo1: case OPC_MorphNodeTo2: {
+    case OPC_EmitNode:
+    case OPC_MorphNodeTo:
+    case OPC_EmitNode0:
+    case OPC_EmitNode1:
+    case OPC_EmitNode2:
+    case OPC_MorphNodeTo0:
+    case OPC_MorphNodeTo1:
+    case OPC_MorphNodeTo2:
+    case OPC_EmitNode0None:
+    case OPC_EmitNode1None:
+    case OPC_EmitNode2None:
+    case OPC_MorphNodeTo0None:
+    case OPC_MorphNodeTo1None:
+    case OPC_MorphNodeTo2None:
+    case OPC_EmitNode0Chain:
+    case OPC_EmitNode1Chain:
+    case OPC_EmitNode2Chain:
+    case OPC_MorphNodeTo0Chain:
+    case OPC_MorphNodeTo1Chain:
+    case OPC_MorphNodeTo2Chain:
+    case OPC_EmitNode0GlueInput:
+    case OPC_EmitNode1GlueInput:
+    case OPC_EmitNode2GlueInput:
+    case OPC_MorphNodeTo0GlueInput:
+    case OPC_MorphNodeTo1GlueInput:
+    case OPC_MorphNodeTo2GlueInput:
+    case OPC_EmitNode0GlueOutput:
+    case OPC_EmitNode1GlueOutput:
+    case OPC_EmitNode2GlueOutput:
+    case OPC_MorphNodeTo0GlueOutput:
+    case OPC_MorphNodeTo1GlueOutput:
+    case OPC_MorphNodeTo2GlueOutput:
+    case OPC_EmitNode0MemRefs:
+    case OPC_EmitNode1MemRefs:
+    case OPC_EmitNode2MemRefs:
+    case OPC_MorphNodeTo0MemRefs:
+    case OPC_MorphNodeTo1MemRefs:
+    case OPC_MorphNodeTo2MemRefs: {
       uint16_t TargetOpc = MatcherTable[MatcherIndex++];
       TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8;
-      unsigned EmitNodeInfo = MatcherTable[MatcherIndex++];
+      unsigned EmitNodeInfo;
+      if (Opcode >= OPC_EmitNode0None && Opcode <= OPC_MorphNodeTo2MemRefs) {
+        if (Opcode >= OPC_EmitNode0Chain && Opcode <= OPC_MorphNodeTo2Chain)
+          EmitNodeInfo = OPFL_Chain;
+        else if (Opcode >= OPC_EmitNode0GlueInput &&
+                 Opcode <= OPC_MorphNodeTo2GlueInput)
+          EmitNodeInfo = OPFL_GlueInput;
+        else if (Opcode >= OPC_EmitNode0GlueOutput &&
+                 Opcode <= OPC_MorphNodeTo2GlueOutput)
+          EmitNodeInfo = OPFL_GlueOutput;
+        else if (Opcode >= OPC_EmitNode0MemRefs &&
+                 Opcode <= OPC_MorphNodeTo2MemRefs)
+          EmitNodeInfo = OPFL_MemRefs;
+        else
+          EmitNodeInfo = OPFL_None;
+      } else
+        EmitNodeInfo = MatcherTable[MatcherIndex++];
       // Get the result VT list.
       unsigned NumVTs;
       // If this is one of the compressed forms, get the number of VTs based
       // on the Opcode. Otherwise read the next byte from the table.
       if (Opcode >= OPC_MorphNodeTo0 && Opcode <= OPC_MorphNodeTo2)
         NumVTs = Opcode - OPC_MorphNodeTo0;
+      else if (Opcode >= OPC_MorphNodeTo0None && Opcode <= OPC_MorphNodeTo2None)
+        NumVTs = Opcode - OPC_MorphNodeTo0None;
+      else if (Opcode >= OPC_MorphNodeTo0Chain &&
+               Opcode <= OPC_MorphNodeTo2Chain)
+        NumVTs = Opcode - OPC_MorphNodeTo0Chain;
+      else if (Opcode >= OPC_MorphNodeTo0GlueInput &&
+               Opcode <= OPC_MorphNodeTo2GlueInput)
+        NumVTs = Opcode - OPC_MorphNodeTo0GlueInput;
+      else if (Opcode >= OPC_MorphNodeTo0GlueOutput &&
+               Opcode <= OPC_MorphNodeTo2GlueOutput)
+        NumVTs = Opcode - OPC_MorphNodeTo0GlueOutput;
+      else if (Opcode >= OPC_MorphNodeTo0MemRefs &&
+               Opcode <= OPC_MorphNodeTo2MemRefs)
+        NumVTs = Opcode - OPC_MorphNodeTo0MemRefs;
       else if (Opcode >= OPC_EmitNode0 && Opcode <= OPC_EmitNode2)
         NumVTs = Opcode - OPC_EmitNode0;
+      else if (Opcode >= OPC_EmitNode0None && Opcode <= OPC_EmitNode2None)
+        NumVTs = Opcode - OPC_EmitNode0None;
+      else if (Opcode >= OPC_EmitNode0Chain && Opcode <= OPC_EmitNode2Chain)
+        NumVTs = Opcode - OPC_EmitNode0Chain;
+      else if (Opcode >= OPC_EmitNode0GlueInput &&
+               Opcode <= OPC_EmitNode2GlueInput)
+        NumVTs = Opcode - OPC_EmitNode0GlueInput;
+      else if (Opcode >= OPC_EmitNode0GlueOutput &&
+               Opcode <= OPC_EmitNode2GlueOutput)
+        NumVTs = Opcode - OPC_EmitNode0GlueOutput;
+      else if (Opcode >= OPC_EmitNode0MemRefs && Opcode <= OPC_EmitNode2MemRefs)
+        NumVTs = Opcode - OPC_EmitNode0MemRefs;
       else
         NumVTs = MatcherTable[MatcherIndex++];
       SmallVector<EVT, 4> VTs;
@@ -3727,8 +3804,18 @@ void SelectionDAGISel::SelectCodeCommon(SDNode *NodeToMatch,
 
       // Create the node.
       MachineSDNode *Res = nullptr;
-      bool IsMorphNodeTo = Opcode == OPC_MorphNodeTo ||
-                     (Opcode >= OPC_MorphNodeTo0 && Opcode <= OPC_MorphNodeTo2);
+      bool IsMorphNodeTo =
+          Opcode == OPC_MorphNodeTo ||
+          (Opcode >= OPC_MorphNodeTo0 && Opcode <= OPC_MorphNodeTo2) ||
+          (Opcode >= OPC_MorphNodeTo0None && Opcode <= OPC_MorphNodeTo2None) ||
+          (Opcode >= OPC_MorphNodeTo0Chain &&
+           Opcode <= OPC_MorphNodeTo2Chain) ||
+          (Opcode >= OPC_MorphNodeTo0GlueInput &&
+           Opcode <= OPC_MorphNodeTo2GlueInput) ||
+          (Opcode >= OPC_MorphNodeTo0GlueOutput &&
+           Opcode <= OPC_MorphNodeTo2GlueOutput) ||
+          (Opcode >= OPC_MorphNodeTo0MemRefs &&
+           Opcode <= OPC_MorphNodeTo2MemRefs);
       if (!IsMorphNodeTo) {
         // If this is a normal EmitNode command, just create the new node and
         // add the results to the RecordedNodes list.
diff --git a/llvm/test/TableGen/DAGDefaultOps.td b/llvm/test/TableGen/DAGDefaultOps.td
index 010ee205e6ff2f8..70c8413f2c053a4 100644
--- a/llvm/test/TableGen/DAGDefaultOps.td
+++ b/llvm/test/TableGen/DAGDefaultOps.td
@@ -77,20 +77,20 @@ def MulIRRPat : Pat<(mul i32:$x, i32:$y), (MulIRR Reg:$x, Reg:$y)>;
 // ADD-NEXT: OPC_RecordChild0
 // ADD-NEXT: OPC_RecordChild1
 // ADD-NEXT: OPC_EmitInteger32, 0
-// ADD-NEXT: OPC_MorphNodeTo1, TARGET_VAL(::AddRRI)
+// ADD-NEXT: OPC_MorphNodeTo1None, TARGET_VAL(::AddRRI)
 
 // ADDINT: SwitchOpcode{{.*}}TARGET_VAL(ISD::INTRINSIC_WO_CHAIN)
 // ADDINT-NEXT: OPC_CheckChild0Integer
 // ADDINT-NEXT: OPC_RecordChild1
 // ADDINT-NEXT: OPC_RecordChild2
 // ADDINT-NEXT: OPC_EmitInteger32, 2
-// ADDINT-NEXT: OPC_MorphNodeTo1, TARGET_VAL(::AddRRI)
+// ADDINT-NEXT: OPC_MorphNodeTo1None, TARGET_VAL(::AddRRI)
 
 // SUB: SwitchOpcode{{.*}}TARGET_VAL(ISD::SUB)
 // SUB-NEXT: OPC_RecordChild0
 // SUB-NEXT: OPC_RecordChild1
 // SUB-NEXT: OPC_EmitInteger32, 0
-// SUB-NEXT: OPC_MorphNodeTo1, TARGET_VAL(::SubRRI)
+// SUB-NEXT: OPC_MorphNodeTo1None, TARGET_VAL(::SubRRI)
 
 // MULINT: SwitchOpcode{{.*}}TARGET_VAL(ISD::INTRINSIC_W_CHAIN)
 // MULINT-NEXT: OPC_RecordNode
@@ -99,10 +99,10 @@ def MulIRRPat : Pat<(mul i32:$x, i32:$y), (MulIRR Reg:$x, Reg:$y)>;
 // MULINT-NEXT: OPC_RecordChild3
 // MULINT-NEXT: OPC_RecordChild4
 // MULINT-NEXT: OPC_EmitMergeInputChains
-// MULINT-NEXT: OPC_MorphNodeTo1, TARGET_VAL(::MulRRI)
+// MULINT-NEXT: OPC_MorphNodeTo1Chain, TARGET_VAL(::MulRRI)
 
 // MUL: SwitchOpcode{{.*}}TARGET_VAL(ISD::MUL)
 // MUL-NEXT: OPC_EmitInteger32, 0
 // MUL-NEXT: OPC_RecordChild0
 // MUL-NEXT: OPC_RecordChild1
-// MUL-NEXT: OPC_MorphNodeTo1, TARGET_VAL(::MulRRI)
+// MUL-NEXT: OPC_MorphNodeTo1Chain, TARGET_VAL(::MulRRI)
diff --git a/llvm/test/TableGen/dag-isel-regclass-emit-enum.td b/llvm/test/TableGen/dag-isel-regclass-emit-enum.td
index 27e214f1641d501..705bbf2c6b282c4 100644
--- a/llvm/test/TableGen/dag-isel-regclass-emit-enum.td
+++ b/llvm/test/TableGen/dag-isel-regclass-emit-enum.td
@@ -26,14 +26,14 @@ def GPRAbove127 : RegisterClass<"TestTarget", [i32], 32,
 // CHECK-NEXT: OPC_Scope, 13, /*->19*/ // 2 children in Scope
 // CHECK-NEXT: OPC_CheckChild1Integer, 0,
 // CHECK-NEXT: OPC_EmitInteger32, 0|128,2/*256*/,
-// CHECK-NEXT: OPC_MorphNodeTo1, TARGET_VAL(TargetOpcode::COPY_TO_REGCLASS), 0,
+// CHECK-NEXT: OPC_MorphNodeTo1None, TARGET_VAL(TargetOpcode::COPY_TO_REGCLASS),
 // CHECK-NEXT:     MVT::i32, 2/*#Ops*/, 1, 0,
 def : Pat<(i32 (add i32:$src, (i32 0))),
           (COPY_TO_REGCLASS GPRAbove127, GPR0:$src)>;
 
 // CHECK:      OPC_CheckChild1Integer, 2,
 // CHECK-NEXT: OPC_EmitStringInteger32, TestNamespace::GPR127RegClassID,
-// CHECK-NEXT: OPC_MorphNodeTo1, TARGET_VAL(TargetOpcode::COPY_TO_REGCLASS), 0,
+// CHECK-NEXT: OPC_MorphNodeTo1None, TARGET_VAL(TargetOpcode::COPY_TO_REGCLASS),
 // CHECK-NEXT:     MVT::i32, 2/*#Ops*/, 1, 0,
 def : Pat<(i32 (add i32:$src, (i32 1))),
           (COPY_TO_REGCLASS GPR127, GPR0:$src)>;
diff --git a/llvm/test/TableGen/dag-isel-res-order.td b/llvm/test/TableGen/dag-isel-res-order.td
index 4da6b6d6867c427..6937dd8c6b807c9 100644
--- a/llvm/test/TableGen/dag-isel-res-order.td
+++ b/llvm/test/TableGen/dag-isel-res-order.td
@@ -12,7 +12,7 @@ def REG : Register<"REG">;
 def GPR : RegisterClass<"TestTarget", [i32], 32, (add REG)>;
 
 // CHECK-LABEL: OPC_CheckOpcode, TARGET_VAL(ISD::UDIVREM)
-// CHECK: OPC_EmitNode2, TARGET_VAL(::INSTR)
+// CHECK: OPC_EmitNode2None, TARGET_VAL(::INSTR)
 // CHECK: Results = #2 #3
 // CHECK: OPC_CompleteMatch, 2, 3, 2
 def INSTR : Instruction {
diff --git a/llvm/utils/TableGen/DAGISelMatcherEmitter.cpp b/llvm/utils/TableGen/DAGISelMatcherEmitter.cpp
index 4da06197658cd0a..1c42e381ad70434 100644
--- a/llvm/utils/TableGen/DAGISelMatcherEmitter.cpp
+++ b/llvm/utils/TableGen/DAGISelMatcherEmitter.cpp
@@ -799,19 +799,49 @@ EmitMatcher(const Matcher *N, const unsigned Indent, unsigned CurrentIdx,
     const EmitNodeMatcherCommon *EN = cast<EmitNodeMatcherCommon>(N);
     OS << (isa<EmitNodeMatcher>(EN) ? "OPC_EmitNode" : "OPC_MorphNodeTo");
     bool CompressVTs = EN->getNumVTs() < 3;
-    if (CompressVTs)
+    bool CompressNodeInfo = false;
+    if (CompressVTs) {
       OS << EN->getNumVTs();
+      if (!EN->hasChain() && !EN->hasInGlue() && !EN->hasOutGlue() &&
+          !EN->hasMemRefs() && EN->getNumFixedArityOperands() == -1) {
+        CompressNodeInfo = true;
+        OS << "None";
+      } else if (EN->hasChain() && !EN->hasInGlue() && !EN->hasOutGlue() &&
+                 !EN->hasMemRefs() && EN->getNumFixedArityOperands() == -1) {
+        CompressNodeInfo = true;
+        OS << "Chain";
+      } else if (!EN->hasChain() && EN->hasInGlue() && !EN->hasOutGlue() &&
+                 !EN->hasMemRefs() && EN->getNumFixedArityOperands() == -1) {
+        CompressNodeInfo = true;
+        OS << "GlueInput";
+      } else if (!EN->hasChain() && !EN->hasInGlue() && EN->hasOutGlue() &&
+                 !EN->hasMemRefs() && EN->getNumFixedArityOperands() == -1) {
+        CompressNodeInfo = true;
+        OS << "GlueOutput";
+      } else if (!EN->hasChain() && !EN->hasInGlue() && !EN->hasOutGlue() &&
+                 EN->hasMemRefs() && EN->getNumFixedArityOperands() == -1) {
+        CompressNodeInfo = true;
+        OS << "MemRefs";
+      }
+    }
 
     const CodeGenInstruction &CGI = EN->getInstruction();
     OS << ", TARGET_VAL(" << CGI.Namespace << "::" << CGI.TheDef->getName()
-       << "), 0";
-
-    if (EN->hasChain())   OS << "|OPFL_Chain";
-    if (EN->hasInGlue())  OS << "|OPFL_GlueInput";
-    if (EN->hasOutGlue()) OS << "|OPFL_GlueOutput";
-    if (EN->hasMemRefs()) OS << "|OPFL_MemRefs";
-    if (EN->getNumFixedArityOperands() != -1)
-      OS << "|OPFL_Variadic" << EN->getNumFixedArityOperands();
+       << ")";
+
+    if (!CompressNodeInfo) {
+      OS << ", 0";
+      if (EN->hasChain())
+        OS << "|OPFL_Chain";
+      if (EN->hasInGlue())
+        OS << "|OPFL_GlueInput";
+      if (EN->hasOutGlue())
+        OS << "|OPFL_GlueOutput";
+      if (EN->hasMemRefs())
+        OS << "|OPFL_MemRefs";
+      if (EN->getNumFixedArityOperands() != -1)
+        OS << "|OPFL_Variadic" << EN->getNumFixedArityOperands();
+    }
     OS << ",\n";
 
     OS.indent(FullIndexWidth + Indent+4);
@@ -854,8 +884,8 @@ EmitMatcher(const Matcher *N, const unsigned Indent, unsigned CurrentIdx,
     } else
       OS << '\n';
 
-    return 5 + !CompressVTs + EN->getNumVTs() + NumOperandBytes +
-           NumCoveredBytes;
+    return 4 + !CompressVTs + !CompressNodeInfo + EN->getNumVTs() +
+           NumOperandBytes + NumCoveredBytes;
   }
   case Matcher::CompleteMatch: {
     const CompleteMatchMatcher *CM = cast<CompleteMatchMatcher>(N);



More information about the llvm-commits mailing list