[llvm] [RISCV] Macro-fusion support for veyron-v1 CPU. (PR #70012)
Mikhail Gudim via llvm-commits
llvm-commits at lists.llvm.org
Thu Oct 26 22:25:04 PDT 2023
https://github.com/mgudim updated https://github.com/llvm/llvm-project/pull/70012
>From bbf0196a7d4200b143c11a66137c86d5b1f137df Mon Sep 17 00:00:00 2001
From: Mikhail Gudim <mgudim at gmail.com>
Date: Wed, 11 Oct 2023 23:47:05 -0400
Subject: [PATCH 1/3] [RISCV] Macro-fusion support for veyron-v1 CPU.
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.
---
llvm/lib/Target/RISCV/RISCVFeatures.td | 18 ++++
llvm/lib/Target/RISCV/RISCVMacroFusion.cpp | 110 ++++++++++++++++++---
llvm/lib/Target/RISCV/RISCVProcessors.td | 3 +-
llvm/lib/Target/RISCV/RISCVSubtarget.h | 10 +-
4 files changed, 121 insertions(+), 20 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVFeatures.td b/llvm/lib/Target/RISCV/RISCVFeatures.td
index 979bc0ea8c7d065..13565ed361dbc2d 100644
--- a/llvm/lib/Target/RISCV/RISCVFeatures.td
+++ b/llvm/lib/Target/RISCV/RISCVFeatures.td
@@ -937,6 +937,16 @@ 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 TuneSLLISRLIFusion
+ : SubtargetFeature<"slli-srli-fusion", "HasSLLISRLIFusion",
+ "true", "Enable SLLI+SRLI macrofusion">;
+def TuneLDADDFusion
+ : SubtargetFeature<"ld-add-fusion", "HasLDADDFusion",
+ "true", "Enable fusion of load with the last instruction of the address calculation">;
+
def TuneNoDefaultUnroll
: SubtargetFeature<"no-default-unroll", "EnableDefaultUnroll", "false",
"Disable default unroll preference.">;
@@ -954,6 +964,14 @@ def TuneSiFive7 : SubtargetFeature<"sifive7", "RISCVProcFamily", "SiFive7",
[TuneNoDefaultUnroll,
TuneShortForwardBranchOpt]>;
+def TuneVeyronFusions : SubtargetFeature<"ventana-veyron", "RISCVProcFamily", "VentanaVeyron",
+ "Ventana Veyron-Series processors",
+ [TuneLUIADDIFusion,
+ TuneAUIPCADDIFusion,
+ TuneSLLISRLIFusion,
+ 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
// is responsible for providing any necessary __sync implementations. Code
diff --git a/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp b/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
index 02a8d5c18fe1a0e..c33b3503aed0f97 100644
--- a/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
+++ b/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
@@ -18,6 +18,90 @@
using namespace llvm;
+static bool checkRegisters(Register FirstDest, const MachineInstr &SecondMI) {
+ 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
+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 SLLI by 32 feeding into SRLI by 32 or less or
+// SLLI by exactly 48 feeding into SRLI by exactly 48.
+static bool isSLLISRLI(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
+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;
+
+ // The first operand of ADDI might be a frame index.
+ if (!SecondMI.getOperand(1).isReg())
+ 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]
@@ -27,7 +111,6 @@ 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;
@@ -35,21 +118,11 @@ static bool isLUIADDI(const MachineInstr *FirstMI,
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)
+ // The first operand of ADDI might be a frame index.
+ if (!SecondMI.getOperand(1).isReg())
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 +134,15 @@ static bool shouldScheduleAdjacent(const TargetInstrInfo &TII,
if (ST.hasLUIADDIFusion() && isLUIADDI(FirstMI, SecondMI))
return true;
+ if (ST.hasAUIPCADDIFusion() && isAUIPCADDI(FirstMI, SecondMI))
+ return true;
+
+ if (ST.hasSLLISRLIFusion() && isSLLISRLI(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
index e4008d145ffa572..3a242e20edb5aa4 100644
--- a/llvm/lib/Target/RISCV/RISCVProcessors.td
+++ b/llvm/lib/Target/RISCV/RISCVProcessors.td
@@ -242,4 +242,5 @@ def VENTANA_VEYRON_V1 : RISCVProcessorModel<"veyron-v1",
FeatureStdExtZicbom,
FeatureStdExtZicbop,
FeatureStdExtZicboz,
- FeatureVendorXVentanaCondOps]>;
+ FeatureVendorXVentanaCondOps,
+ TuneVeyronFusions]>;
diff --git a/llvm/lib/Target/RISCV/RISCVSubtarget.h b/llvm/lib/Target/RISCV/RISCVSubtarget.h
index 6b915e61c136086..cd408101428634b 100644
--- a/llvm/lib/Target/RISCV/RISCVSubtarget.h
+++ b/llvm/lib/Target/RISCV/RISCVSubtarget.h
@@ -46,10 +46,7 @@ struct RISCVTuneInfo {
class RISCVSubtarget : public RISCVGenSubtargetInfo {
public:
- enum RISCVProcFamilyEnum : uint8_t {
- Others,
- SiFive7,
- };
+ enum RISCVProcFamilyEnum : uint8_t { Others, SiFive7, VentanaVeyron };
private:
virtual void anchor();
@@ -182,7 +179,10 @@ class RISCVSubtarget : public RISCVGenSubtargetInfo {
return UserReservedRegister[i];
}
- bool hasMacroFusion() const { return hasLUIADDIFusion(); }
+ bool hasMacroFusion() const {
+ return hasLUIADDIFusion() || hasAUIPCADDIFusion() || hasSLLISRLIFusion() ||
+ hasLDADDFusion();
+ }
// Vector codegen related methods.
bool hasVInstructions() const { return HasStdExtZve32x; }
>From d98d16628b244b8abfb7f2a9ffb7aa84f1bf3e51 Mon Sep 17 00:00:00 2001
From: Mikhail Gudim <mgudim at gmail.com>
Date: Wed, 25 Oct 2023 01:09:03 -0400
Subject: [PATCH 2/3] addressed review comments.
---
llvm/lib/Target/RISCV/RISCVMacroFusion.cpp | 23 ++++++++++++----------
1 file changed, 13 insertions(+), 10 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp b/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
index c33b3503aed0f97..5e64c5799fd2512 100644
--- a/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
+++ b/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
@@ -19,6 +19,9 @@
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;
@@ -53,8 +56,16 @@ static bool isLDADD(const MachineInstr *FirstMI, const MachineInstr &SecondMI) {
return checkRegisters(FirstMI->getOperand(0).getReg(), SecondMI);
}
-// Fuse SLLI by 32 feeding into SRLI by 32 or less or
-// SLLI by exactly 48 feeding into SRLI by exactly 48.
+// Fuse these patterns:
+//
+// $rd = slli $rs0, 32
+// $rd = srli $rs1, x
+// where 0 <= x <= 32
+//
+// and
+//
+// $rd = slli $rs0, 48
+// $rd = srli $rs1, 48
static bool isSLLISRLI(const MachineInstr *FirstMI,
const MachineInstr &SecondMI) {
if (SecondMI.getOpcode() != RISCV::SRLI)
@@ -95,10 +106,6 @@ static bool isAUIPCADDI(const MachineInstr *FirstMI,
if (FirstMI->getOpcode() != RISCV::AUIPC)
return false;
- // The first operand of ADDI might be a frame index.
- if (!SecondMI.getOperand(1).isReg())
- return false;
-
return checkRegisters(FirstMI->getOperand(0).getReg(), SecondMI);
}
@@ -118,10 +125,6 @@ static bool isLUIADDI(const MachineInstr *FirstMI,
if (FirstMI->getOpcode() != RISCV::LUI)
return false;
- // The first operand of ADDI might be a frame index.
- if (!SecondMI.getOperand(1).isReg())
- return false;
-
return checkRegisters(FirstMI->getOperand(0).getReg(), SecondMI);
}
>From a84f53e00a18f2ff77e266861df675c45de46605 Mon Sep 17 00:00:00 2001
From: Mikhail Gudim <mgudim at gmail.com>
Date: Thu, 26 Oct 2023 16:06:10 -0400
Subject: [PATCH 3/3] Corrected slli-srli fusion - the immediate shift ammounts
have to be the same.
---
llvm/lib/Target/RISCV/RISCVMacroFusion.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp b/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
index 5e64c5799fd2512..2d5094bd2122e54 100644
--- a/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
+++ b/llvm/lib/Target/RISCV/RISCVMacroFusion.cpp
@@ -88,7 +88,7 @@ static bool isSLLISRLI(const MachineInstr *FirstMI,
return false;
unsigned SLLIImm = FirstMI->getOperand(2).getImm();
- if (IsShiftBy48 ? (SLLIImm != 48) : (SLLIImm > 32))
+ if (IsShiftBy48 ? (SLLIImm != 48) : (SLLIImm != 32))
return false;
return checkRegisters(FirstMI->getOperand(0).getReg(), SecondMI);
More information about the llvm-commits
mailing list