[RISCV] Macro-fusion support for veyron-v1 CPU. (#70012)

mgudim · web-flow · commit 29ee66f4a096 · 2023-12-11T16:34:13.000-05:00
Support was added for the following fusions:
  auipc-addi, slli-srli, ld-add
Some parts of the code became repetative, so small refactoring of
existing lui-addi fusion was done.
diff --git a/llvm/lib/Target/RISCV/RISCVFeatures.td b/llvm/lib/Target/RISCV/RISCVFeatures.td
@@ -970,6 +970,16 @@ def TuneLUIADDIFusion
     : SubtargetFeature<"lui-addi-fusion", "HasLUIADDIFusion",
                        "true", "Enable LUI+ADDI macrofusion">;
 
+def TuneAUIPCADDIFusion
+    : SubtargetFeature<"auipc-addi-fusion", "HasAUIPCADDIFusion",
+                       "true", "Enable AUIPC+ADDI macrofusion">;
+def TuneShiftedZExtFusion
+    : SubtargetFeature<"shifted-zext-fusion", "HasShiftedZExtFusion",
+                       "true", "Enable SLLI+SRLI to be fused when computing (shifted) zero extension">;
+def TuneLDADDFusion
+    : SubtargetFeature<"ld-add-fusion", "HasLDADDFusion",
+                       "true", "Enable LD+ADD macrofusion.">;
+
 def TuneNoDefaultUnroll
     : SubtargetFeature<"no-default-unroll", "EnableDefaultUnroll", "false",
                        "Disable default unroll preference.">;
@@ -987,9 +997,12 @@ def TuneSiFive7 : SubtargetFeature<"sifive7", "RISCVProcFamily", "SiFive7",
                                    [TuneNoDefaultUnroll,
                                     TuneShortForwardBranchOpt]>;
 
-def TuneVentanaVeyron : SubtargetFeature<"ventana-veyron", "RISCVProcFamily", "VentanaVeyron",
-                                         "Ventana-Veyron Series processors",
-                                         [TuneLUIADDIFusion]>;
+def TuneVeyronFusions : SubtargetFeature<"ventana-veyron", "RISCVProcFamily", "VentanaVeyron",
+                                         "Ventana Veyron-Series processors",
+                                         [TuneLUIADDIFusion,
+                                          TuneAUIPCADDIFusion,
+                                          TuneShiftedZExtFusion,
+                                          TuneLDADDFusion]>;
 
 // Assume that lock-free native-width atomics are available, even if the target
 // and operating system combination would not usually provide them. The user
diff --git a/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp b/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
@@ -18,6 +18,101 @@
 
 using namespace llvm;
 
+static bool checkRegisters(Register FirstDest, const MachineInstr &SecondMI) {
+  if (!SecondMI.getOperand(1).isReg())
+    return false;
+
+  if (SecondMI.getOperand(1).getReg() != FirstDest)
+    return false;
+
+  // If the input is virtual make sure this is the only user.
+  if (FirstDest.isVirtual()) {
+    auto &MRI = SecondMI.getMF()->getRegInfo();
+    return MRI.hasOneNonDBGUse(FirstDest);
+  }
+
+  return SecondMI.getOperand(0).getReg() == FirstDest;
+}
+
+// Fuse load with add:
+// add rd, rs1, rs2
+// ld rd, 0(rd)
+static bool isLDADD(const MachineInstr *FirstMI, const MachineInstr &SecondMI) {
+  if (SecondMI.getOpcode() != RISCV::LD)
+    return false;
+
+  if (!SecondMI.getOperand(2).isImm())
+    return false;
+
+  if (SecondMI.getOperand(2).getImm() != 0)
+    return false;
+
+  // Given SecondMI, when FirstMI is unspecified, we must return
+  // if SecondMI may be part of a fused pair at all.
+  if (!FirstMI)
+    return true;
+
+  if (FirstMI->getOpcode() != RISCV::ADD)
+    return true;
+
+  return checkRegisters(FirstMI->getOperand(0).getReg(), SecondMI);
+}
+
+// Fuse these patterns:
+//
+// slli rd, rs1, 32
+// srli rd, rd, x
+// where 0 <= x <= 32
+//
+// and
+//
+// slli rd, rs1, 48
+// srli rd, rd, x
+static bool isShiftedZExt(const MachineInstr *FirstMI,
+                          const MachineInstr &SecondMI) {
+  if (SecondMI.getOpcode() != RISCV::SRLI)
+    return false;
+
+  if (!SecondMI.getOperand(2).isImm())
+    return false;
+
+  unsigned SRLIImm = SecondMI.getOperand(2).getImm();
+  bool IsShiftBy48 = SRLIImm == 48;
+  if (SRLIImm > 32 && !IsShiftBy48)
+    return false;
+
+  // Given SecondMI, when FirstMI is unspecified, we must return
+  // if SecondMI may be part of a fused pair at all.
+  if (!FirstMI)
+    return true;
+
+  if (FirstMI->getOpcode() != RISCV::SLLI)
+    return false;
+
+  unsigned SLLIImm = FirstMI->getOperand(2).getImm();
+  if (IsShiftBy48 ? (SLLIImm != 48) : (SLLIImm != 32))
+    return false;
+
+  return checkRegisters(FirstMI->getOperand(0).getReg(), SecondMI);
+}
+
+// Fuse AUIPC followed by ADDI
+// auipc rd, imm20
+// addi rd, rd, imm12
+static bool isAUIPCADDI(const MachineInstr *FirstMI,
+                        const MachineInstr &SecondMI) {
+  if (SecondMI.getOpcode() != RISCV::ADDI)
+    return false;
+  // Assume the 1st instr to be a wildcard if it is unspecified.
+  if (!FirstMI)
+    return true;
+
+  if (FirstMI->getOpcode() != RISCV::AUIPC)
+    return false;
+
+  return checkRegisters(FirstMI->getOperand(0).getReg(), SecondMI);
+}
+
 // Fuse LUI followed by ADDI or ADDIW.
 // rd = imm[31:0] which decomposes to
 // lui rd, imm[31:12]
@@ -27,29 +122,14 @@ static bool isLUIADDI(const MachineInstr *FirstMI,
   if (SecondMI.getOpcode() != RISCV::ADDI &&
       SecondMI.getOpcode() != RISCV::ADDIW)
     return false;
-
   // Assume the 1st instr to be a wildcard if it is unspecified.
   if (!FirstMI)
     return true;
 
   if (FirstMI->getOpcode() != RISCV::LUI)
     return false;
 
-  Register FirstDest = FirstMI->getOperand(0).getReg();
-
-  // Destination of LUI should be the ADDI(W) source register.
-  if (SecondMI.getOperand(1).getReg() != FirstDest)
-    return false;
-
-  // If the input is virtual make sure this is the only user.
-  if (FirstDest.isVirtual()) {
-    auto &MRI = SecondMI.getMF()->getRegInfo();
-    return MRI.hasOneNonDBGUse(FirstDest);
-  }
-
-  // If the FirstMI destination is non-virtual, it should match the SecondMI
-  // destination.
-  return SecondMI.getOperand(0).getReg() == FirstDest;
+  return checkRegisters(FirstMI->getOperand(0).getReg(), SecondMI);
 }
 
 static bool shouldScheduleAdjacent(const TargetInstrInfo &TII,
@@ -61,6 +141,15 @@ static bool shouldScheduleAdjacent(const TargetInstrInfo &TII,
   if (ST.hasLUIADDIFusion() && isLUIADDI(FirstMI, SecondMI))
     return true;
 
+  if (ST.hasAUIPCADDIFusion() && isAUIPCADDI(FirstMI, SecondMI))
+    return true;
+
+  if (ST.hasShiftedZExtFusion() && isShiftedZExt(FirstMI, SecondMI))
+    return true;
+
+  if (ST.hasLDADDFusion() && isLDADD(FirstMI, SecondMI))
+    return true;
+
   return false;
 }
 
diff --git a/llvm/lib/Target/RISCV/RISCVProcessors.td b/llvm/lib/Target/RISCV/RISCVProcessors.td
@@ -254,7 +254,7 @@ def VENTANA_VEYRON_V1 : RISCVProcessorModel<"veyron-v1",
                                              FeatureStdExtZicbop,
                                              FeatureStdExtZicboz,
                                              FeatureVendorXVentanaCondOps],
-                                             [TuneVentanaVeyron]>;
+                                             [TuneVeyronFusions]>;
 
 def XIANGSHAN_NANHU : RISCVProcessorModel<"xiangshan-nanhu",
                                           NoSchedModel,
diff --git a/llvm/lib/Target/RISCV/RISCVSubtarget.h b/llvm/lib/Target/RISCV/RISCVSubtarget.h
@@ -193,7 +193,10 @@ class RISCVSubtarget : public RISCVGenSubtargetInfo {
     return UserReservedRegister[i];
   }
 
-  bool hasMacroFusion() const { return hasLUIADDIFusion(); }
+  bool hasMacroFusion() const {
+    return hasLUIADDIFusion() || hasAUIPCADDIFusion() ||
+           hasShiftedZExtFusion() || hasLDADDFusion();
+  }
 
   // Vector codegen related methods.
   bool hasVInstructions() const { return HasStdExtZve32x; }
diff --git a/llvm/test/CodeGen/RISCV/macro-fusions-veyron-v1.mir b/llvm/test/CodeGen/RISCV/macro-fusions-veyron-v1.mir
@@ -0,0 +1,159 @@
+# REQUIRES: asserts
+# RUN: llc -mtriple=riscv64-linux-gnu  -mcpu=veyron-v1 -x=mir < %s \
+# RUN:   -debug-only=machine-scheduler -start-before=machine-scheduler 2>&1 \
+# RUN:   -mattr=+lui-addi-fusion,+auipc-addi-fusion,+shifted-zext-fusion,+ld-add-fusion \
+# RUN:   | FileCheck %s
+
+# CHECK: lui_addi:%bb.0
+# CHECK: Macro fuse: {{.*}}LUI - ADDI
+---
+name: lui_addi
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10
+    %1:gpr = COPY $x10
+    %2:gpr = LUI 1
+    %3:gpr = XORI %1, 2
+    %4:gpr = ADDI %2, 3
+    $x10 = COPY %3
+    $x11 = COPY %4
+    PseudoRET
+...
+
+# CHECK: auipc_addi
+# CHECK: Macro fuse: {{.*}}AUIPC - ADDI
+---
+name: auipc_addi
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10
+    %1:gpr = COPY $x10
+    %2:gpr = AUIPC 1
+    %3:gpr = XORI %1, 2
+    %4:gpr = ADDI %2, 3
+    $x10 = COPY %3
+    $x11 = COPY %4
+    PseudoRET
+...
+
+# CHECK: slli_srli
+# CHECK: Macro fuse: {{.*}}SLLI - SRLI
+---
+name: slli_srli
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10
+    %1:gpr = COPY $x10
+    %2:gpr = SLLI %1, 32
+    %3:gpr = XORI %1, 3
+    %4:gpr = SRLI %2, 4
+    $x10 = COPY %3
+    $x11 = COPY %4
+    PseudoRET
+...
+
+# CHECK: slli_srli_48
+# CHECK: Macro fuse: {{.*}}SLLI - SRLI
+---
+name: slli_srli_48
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10
+    %1:gpr = COPY $x10
+    %2:gpr = SLLI %1, 48
+    %3:gpr = XORI %1, 3
+    %4:gpr = SRLI %2, 48
+    $x10 = COPY %3
+    $x11 = COPY %4
+    PseudoRET
+...
+
+# CHECK: slli_srli_no_fusion_0
+# CHECK-NOT: Macro fuse: {{.*}}SLLI - SRLI
+---
+name: slli_srli_no_fusion_0
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10
+    %1:gpr = COPY $x10
+    %2:gpr = SLLI %1, 32
+    %3:gpr = XORI %1, 3
+    %4:gpr = SRLI %2, 33
+    $x10 = COPY %3
+    $x11 = COPY %4
+    PseudoRET
+...
+
+# CHECK: slli_srli_no_fusion_1
+# CHECK-NOT: Macro fuse: {{.*}}SLLI - SRLI
+---
+name: slli_srli_no_fusion_1
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10
+    %1:gpr = COPY $x10
+    %2:gpr = SLLI %1, 48
+    %3:gpr = XORI %1, 3
+    %4:gpr = SRLI %2, 4
+    $x10 = COPY %3
+    $x11 = COPY %4
+    PseudoRET
+...
+
+# CHECK: slli_srli_no_fusion_2
+# CHECK-NOT: Macro fuse: {{.*}}SLLI - SRLI
+---
+name: slli_srli_no_fusion_2
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10
+    %1:gpr = COPY $x10
+    %2:gpr = SLLI %1, 31
+    %3:gpr = XORI %1, 3
+    %4:gpr = SRLI %2, 4
+    $x10 = COPY %3
+    $x11 = COPY %4
+    PseudoRET
+...
+
+# CHECK: slli_srli_no_fusion_3
+# CHECK-NOT: Macro fuse: {{.*}}SLLI - SRLI
+---
+name: slli_srli_no_fusion_3
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10
+    %1:gpr = COPY $x10
+    %2:gpr = SLLI %1, 31
+    %3:gpr = XORI %1, 3
+    %4:gpr = SRLI %2, 48
+    $x10 = COPY %3
+    $x11 = COPY %4
+    PseudoRET
+...
+
+# CHECK: ld_add
+# CHECK: Macro fuse: {{.*}}ADD - LD
+---
+name: ld_add
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10, $x11
+    %1:gpr = COPY $x10
+    %2:gpr = COPY $x11
+    %3:gpr = ADD %1, %2
+    %4:gpr = XORI %2, 3
+    %5:gpr = LD %3, 0
+    $x10 = COPY %4
+    $x11 = COPY %5
+    PseudoRET
+...
