[Exegesis][RISCV] Add RISCV support for llvm-exegesis (#120467)

This patch also makes following amendments to core exegesis:
* Added distinction between regular registers aliasing check and
registers used as memory address in instruction.
* Added scratch memory space pointer register.
* General exegesis options were amended:
        * mattr - new option to pass a list of enabled target features

Llvm-exegesis RISCV port is a result of team effort. Below everyone
involved listed.
Co-authored-by: Konstantin Vladimirov
<[email protected]>
Co-authored-by: Dmitrii Petrov <[email protected]>
Co-authored-by: Dmitry Bushev <[email protected]>
Co-authored-by: Mark Goncharov <[email protected]>
Co-authored-by: Anastasiya Chernikova
<[email protected]>

Original pr: #89047

---------

Co-authored-by: Kazu Hirata <[email protected]>

Author: dybv-sc

llvm/test/tools/llvm-exegesis/RISCV/latency-by-extension-A.s

@@ -0,0 +1,59 @@
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=AMOAND_D -mattr="+a" | FileCheck --check-prefix=AMOAND_D %s
+
+AMOAND_D:      ---
+AMOAND_D-NEXT: mode: latency
+AMOAND_D-NEXT: key:
+AMOAND_D-NEXT:   instructions:
+AMOAND_D-NEXT:     - 'AMOAND_D [[RE01:X[0-9]+]] X10 [[RE01:X[0-9]+]]'
+AMOAND_D-NEXT: config: ''
+AMOAND_D-NEXT: register_initial_values:
+AMOAND_D-NEXT: - '[[RE01:X[0-9]+]]=0x0'
+AMOAND_D-DAG: ...
+
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=AMOADD_W -mattr="+a" | FileCheck --check-prefix=AMOADD_W %s
+
+AMOADD_W:      ---
+AMOADD_W-NEXT: mode: latency
+AMOADD_W-NEXT: key:
+AMOADD_W-NEXT:   instructions:
+AMOADD_W-NEXT:     - 'AMOADD_W [[RE02:X[0-9]+]] X10 [[RE02:X[0-9]+]]'
+AMOADD_W-NEXT: config: ''
+AMOADD_W-NEXT: register_initial_values:
+AMOADD_W-NEXT: - '[[RE02:X[0-9]+]]=0x0'
+AMOADD_W-DAG: ...
+
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=AMOMAXU_D -mattr="+a" | FileCheck --check-prefix=AMOMAXU_D %s
+
+AMOMAXU_D:      ---
+AMOMAXU_D-NEXT: mode: latency
+AMOMAXU_D-NEXT: key:
+AMOMAXU_D-NEXT:   instructions:
+AMOMAXU_D-NEXT:     - 'AMOMAXU_D [[RE03:X[0-9]+]] X10 [[RE03:X[0-9]+]]'
+AMOMAXU_D-NEXT: config: ''
+AMOMAXU_D-NEXT: register_initial_values:
+AMOMAXU_D-NEXT: - '[[RE03:X[0-9]+]]=0x0'
+AMOMAXU_D-DAG: ...
+
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=AMOMIN_W -mattr="+a" | FileCheck --check-prefix=AMOMIN_W %s
+
+AMOMIN_W:      ---
+AMOMIN_W-NEXT: mode: latency
+AMOMIN_W-NEXT: key:
+AMOMIN_W-NEXT:   instructions:
+AMOMIN_W-NEXT:     - 'AMOMIN_W [[RE04:X[0-9]+]] X10 [[RE04:X[0-9]+]]'
+AMOMIN_W-NEXT: config: ''
+AMOMIN_W-NEXT: register_initial_values:
+AMOMIN_W-NEXT: - '[[RE04:X[0-9]+]]=0x0'
+AMOMIN_W-DAG: ...
+
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=AMOXOR_D -mattr="+a" | FileCheck --check-prefix=AMOXOR_D %s
+
+AMOXOR_D:      ---
+AMOXOR_D-NEXT: mode: latency
+AMOXOR_D-NEXT: key:
+AMOXOR_D-NEXT:   instructions:
+AMOXOR_D-NEXT:     - 'AMOXOR_D [[RE05:X[0-9]+]] X10 [[RE05:X[0-9]+]]'
+AMOXOR_D-NEXT: config: ''
+AMOXOR_D-NEXT: register_initial_values:
+AMOXOR_D-NEXT: - '[[RE05:X[0-9]+]]=0x0'
+AMOXOR_D-DAG: ...

llvm/test/tools/llvm-exegesis/RISCV/latency-by-extension-C.s

@@ -0,0 +1,48 @@
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=C_ADDI -mattr=+c | FileCheck --check-prefix=C_ADDI %s
+
+C_ADDI:      ---
+C_ADDI-NEXT: mode: latency
+C_ADDI-NEXT: key:
+C_ADDI-NEXT:   instructions:
+C_ADDI-NEXT:     - 'C_ADDI [[REG01:X[0-9]+]] [[RE02:X[0-9]+]] [[IMM0:i_0x[0-9]+]]'
+
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=C_ADDIW -mattr=+c | FileCheck --check-prefix=C_ADDIW %s
+
+C_ADDIW:      ---
+C_ADDIW-NEXT: mode: latency
+C_ADDIW-NEXT: key:
+C_ADDIW-NEXT:   instructions:
+C_ADDIW-NEXT:     - 'C_ADDIW [[REG11:X[0-9]+]] [[RE12:X[0-9]+]] [[IMM1:i_0x[0-9]+]]'
+
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=C_ANDI -mattr=+c | FileCheck --check-prefix=C_ANDI %s
+
+C_ANDI:      ---
+C_ANDI-NEXT: mode: latency
+C_ANDI-NEXT: key:
+C_ANDI-NEXT:   instructions:
+C_ANDI-NEXT:     - 'C_ANDI [[REG31:X[0-9]+]] [[REG32:X[0-9]+]] [[IMM3:i_0x[0-9]+]]'
+
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=C_SLLI -mattr=+c | FileCheck --check-prefix=C_SLLI %s
+
+C_SLLI:      ---
+C_SLLI-NEXT: mode: latency
+C_SLLI-NEXT: key:
+C_SLLI-NEXT:   instructions:
+C_SLLI-NEXT:     - 'C_SLLI [[REG81:X[0-9]+]] [[REG82:X[0-9]+]] [[IMM8:i_0x[0-9]+]]'
+
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=C_SRAI -mattr=+c | FileCheck --check-prefix=C_SRAI %s
+
+C_SRAI:      ---
+C_SRAI-NEXT: mode: latency
+C_SRAI-NEXT: key:
+C_SRAI-NEXT:   instructions:
+C_SRAI-NEXT:     - 'C_SRAI [[REG91:X[0-9]+]] [[REG92:X[0-9]+]] [[IMM9:i_0x[0-9]+]]'
+
+# RUN: llvm-exegesis -mode=latency -mtriple=riscv64-unknown-linux-gnu --mcpu=generic --benchmark-phase=assemble-measured-code -opcode-name=C_SRLI -mattr=+c | FileCheck --check-prefix=C_SRLI %s
+
+C_SRLI:      ---
+C_SRLI-NEXT: mode: latency
+C_SRLI-NEXT: key:
+C_SRLI-NEXT:   instructions:
+C_SRLI-NEXT:     - 'C_SRLI [[REG101:X[0-9]+]] [[REG102:X[0-9]+]] [[IMM10:i_0x[0-9]+]]'
+C_SRLI-DAG: ...

llvm/test/tools/llvm-exegesis/RISCV/latency-by-opcode-name-FADD_D.s

@@ -0,0 +1,11 @@
+# RUN: llvm-exegesis -mtriple=riscv64-unknown-linux-gnu --mcpu=generic -mode=latency --benchmark-phase=assemble-measured-code -mattr=+d -opcode-name=FADD_D | FileCheck %s
+
+CHECK:      ---
+CHECK-NEXT: mode: latency
+CHECK-NEXT: key:
+CHECK-NEXT:   instructions:
+CHECK-NEXT:     - 'FADD_D [[REG1:F[0-9]+_D]] [[REG2:F[0-9]+_D]] [[REG3:F[0-9]+_D]] i_0x7'
+CHECK-NEXT: config: ''
+CHECK-NEXT: register_initial_values:
+CHECK-DAG: - '[[REG1]]=0x0'
+CHECK-DAG: ...

llvm/test/tools/llvm-exegesis/RISCV/lit.local.cfg

@@ -0,0 +1,3 @@
+if not ("RISCV" in config.root.targets):
+    # We need support for RISCV.
+    config.unsupported = True

llvm/tools/llvm-exegesis/lib/CMakeLists.txt

@@ -12,6 +12,9 @@ endif()
 if (LLVM_TARGETS_TO_BUILD MATCHES "Mips")
   list(APPEND LLVM_EXEGESIS_TARGETS "Mips")
 endif()
+if(LLVM_TARGETS_TO_BUILD MATCHES "RISCV")
+  list(APPEND LLVM_EXEGESIS_TARGETS "RISCV")
+endif()
 
 set(LLVM_EXEGESIS_TARGETS ${LLVM_EXEGESIS_TARGETS} PARENT_SCOPE)
 

llvm/tools/llvm-exegesis/lib/MCInstrDescView.cpp

@@ -95,11 +95,12 @@ Instruction::Instruction(const MCInstrDesc *Description, StringRef Name,
                          const BitVector *ImplDefRegs,
                          const BitVector *ImplUseRegs,
                          const BitVector *AllDefRegs,
-                         const BitVector *AllUseRegs)
+                         const BitVector *AllUseRegs,
+                         const BitVector *NonMemoryRegs)
     : Description(*Description), Name(Name), Operands(std::move(Operands)),
       Variables(std::move(Variables)), ImplDefRegs(*ImplDefRegs),
       ImplUseRegs(*ImplUseRegs), AllDefRegs(*AllDefRegs),
-      AllUseRegs(*AllUseRegs) {}
+      AllUseRegs(*AllUseRegs), NonMemoryRegs(*NonMemoryRegs) {}
 
 std::unique_ptr<Instruction>
 Instruction::create(const MCInstrInfo &InstrInfo,
@@ -166,6 +167,8 @@ Instruction::create(const MCInstrInfo &InstrInfo,
   BitVector ImplUseRegs = RATC.emptyRegisters();
   BitVector AllDefRegs = RATC.emptyRegisters();
   BitVector AllUseRegs = RATC.emptyRegisters();
+  BitVector NonMemoryRegs = RATC.emptyRegisters();
+
   for (const auto &Op : Operands) {
     if (Op.isReg()) {
       const auto &AliasingBits = Op.getRegisterAliasing().aliasedBits();
@@ -177,6 +180,8 @@ Instruction::create(const MCInstrInfo &InstrInfo,
         ImplDefRegs |= AliasingBits;
       if (Op.isUse() && Op.isImplicit())
         ImplUseRegs |= AliasingBits;
+      if (Op.isUse() && !Op.isMemory())
+        NonMemoryRegs |= AliasingBits;
     }
   }
   // Can't use make_unique because constructor is private.
@@ -185,7 +190,8 @@ Instruction::create(const MCInstrInfo &InstrInfo,
       std::move(Variables), BVC.getUnique(std::move(ImplDefRegs)),
       BVC.getUnique(std::move(ImplUseRegs)),
       BVC.getUnique(std::move(AllDefRegs)),
-      BVC.getUnique(std::move(AllUseRegs))));
+      BVC.getUnique(std::move(AllUseRegs)),
+      BVC.getUnique(std::move(NonMemoryRegs))));
 }
 
 const Operand &Instruction::getPrimaryOperand(const Variable &Var) const {
@@ -240,6 +246,12 @@ bool Instruction::hasAliasingRegisters(
                                      ForbiddenRegisters);
 }
 
+bool Instruction::hasAliasingNotMemoryRegisters(
+    const BitVector &ForbiddenRegisters) const {
+  return anyCommonExcludingForbidden(AllDefRegs, NonMemoryRegs,
+                                     ForbiddenRegisters);
+}
+
 bool Instruction::hasOneUseOrOneDef() const {
   return AllDefRegs.count() || AllUseRegs.count();
 }

llvm/tools/llvm-exegesis/lib/MCInstrDescView.h

@@ -133,6 +133,12 @@ struct Instruction {
   // aliasing Use and Def registers.
   bool hasAliasingRegisters(const BitVector &ForbiddenRegisters) const;
 
+  // Whether this instruction is self aliasing through some registers.
+  // Repeating this instruction may execute sequentially by picking aliasing
+  // Def and Not Memory Use registers. It may also execute in parallel by
+  // picking non aliasing Def and Not Memory Use registers.
+  bool hasAliasingNotMemoryRegisters(const BitVector &ForbiddenRegisters) const;
+
   // Whether this instruction's registers alias with OtherInstr's registers.
   bool hasAliasingRegistersThrough(const Instruction &OtherInstr,
                                    const BitVector &ForbiddenRegisters) const;
@@ -160,12 +166,15 @@ struct Instruction {
   const BitVector &ImplUseRegs; // The set of aliased implicit use registers.
   const BitVector &AllDefRegs;  // The set of all aliased def registers.
   const BitVector &AllUseRegs;  // The set of all aliased use registers.
+  // The set of all aliased not memory use registers.
+  const BitVector &NonMemoryRegs;
+
 private:
   Instruction(const MCInstrDesc *Description, StringRef Name,
               SmallVector<Operand, 8> Operands,
               SmallVector<Variable, 4> Variables, const BitVector *ImplDefRegs,
               const BitVector *ImplUseRegs, const BitVector *AllDefRegs,
-              const BitVector *AllUseRegs);
+              const BitVector *AllUseRegs, const BitVector *NonMemoryRegs);
 };
 
 // Instructions are expensive to instantiate. This class provides a cache of

llvm/tools/llvm-exegesis/lib/RISCV/CMakeLists.txt

@@ -0,0 +1,22 @@
+include_directories(
+  ${LLVM_MAIN_SRC_DIR}/lib/Target/RISCV
+  ${LLVM_BINARY_DIR}/lib/Target/RISCV
+)
+
+set(LLVM_LINK_COMPONENTS
+  CodeGen
+  RISCV
+  Exegesis
+  Core
+  Support
+  )
+
+add_llvm_library(LLVMExegesisRISCV
+  DISABLE_LLVM_LINK_LLVM_DYLIB
+  STATIC
+  Target.cpp
+
+  DEPENDS
+  intrinsics_gen
+  RISCVCommonTableGen
+  )