[llvm-dev] [RISC-V]How to change the behavior of register allocation for RISC-V inline assembly
Junning Wu via llvm-dev
llvm-dev at lists.llvm.org
Mon Apr 6 17:50:32 PDT 2020
*The problem is, I have an instruction named LQP, which load 4x32-bit from
memory and store them into 4 consecutive GPRs, such as, LQP a0, a4, a5, 4,*
* the loaded data will be stored in (a3,a2,a1,a0), the data address is
4(a4+a5).*
*The RISC-V inline assembly like this:*
asm volatile
(
"lqp %[z], %[x], %[y], 4\n\t"
: [z] "=r" (c)
: [x] "r" (a), [y] "r" (b)
) ;
*To get this done, I creat to GPR groups and change the LQP instruction's
definition, like this:*
def GPRA0 : RegisterClass<"RISCV", [XLenVT], 32, (add X10)> {
let RegInfos = RegInfoByHwMode<
[RV32, RV64, DefaultMode],
[RegInfo<32,32,32>, RegInfo<64,64,64>, RegInfo<32,32,32>]>;
}
def GPRNOA0A1A2A3 : RegisterClass<"RISCV", [XLenVT], 32, (add
(sequence "X%u", 14, 17),
(sequence "X%u", 5, 7),
(sequence "X%u", 28, 31),
(sequence "X%u", 8, 9),
(sequence "X%u", 18, 27),
(sequence "X%u", 0, 4)
)> {
let RegInfos = RegInfoByHwMode<
[RV32, RV64, DefaultMode],
[RegInfo<32,32,32>, RegInfo<64,64,64>, RegInfo<32,32,32>]>;
}
let hasSideEffects = 0, mayLoad = 1, mayStore = 0 in
def LQP : RVInstRI<0b11, 0b000, OPC_HX_CUS0,
(outs GPRA0:$rd), (ins GPRNOA0A1A2A3:$rs1,
GPRNOA0A1A2A3:$rs2, simm5:$shift),
"lqp", "$rd, $rs1, $rs2, $shift">, Sched<[]> {
bits<5> shift;
bits<5> rs1;
bits<5> rs2;
bits<5> rd;
let Inst{31-30} = 0b11;
let Inst{29-25} = shift;
let Inst{24-20} = rs2;
let Inst{19-15} = rs1;
let Inst{14-12} = 0b000;
let Inst{11-7} = rd;
let Opcode = OPC_HX_CUS0.Value;
}
*and the DecodeStatus as well, *
static DecodeStatus DecodeGPRA0RegisterClass(MCInst &Inst, uint64_t RegNo,
uint64_t Address,
const void *Decoder) {
if (RegNo != 10) {
return MCDisassembler::Fail;
}
return DecodeGPRRegisterClass(Inst, RegNo, Address, Decoder);
}
static DecodeStatus DecodeGPRNOA0A1A2A3RegisterClass(MCInst &Inst, uint64_t
RegNo,
uint64_t Address,
const void *Decoder) {
if (RegNo == 10 || RegNo == 11 || RegNo == 12 || RegNo == 13) {
return MCDisassembler::Fail;
}
return DecodeGPRRegisterClass(Inst, RegNo, Address, Decoder);
}
static DecodeStatus DecodeGPRNOA0A1RegisterClass(MCInst &Inst, uint64_t
RegNo,
uint64_t Address,
const void *Decoder) {
if (RegNo == 10 || RegNo == 11) {
return MCDisassembler::Fail;
}
return DecodeGPRRegisterClass(Inst, RegNo, Address, Decoder);
}
*Rebuild LLVM+clang, and compile the above c code which got such an error:*
clang -I./env -I./common -I./src/test_newinst
-I/home/llvm/workspace/llvm/llvm-project/llvm_install/riscv32-unknown-elf/include
-mcmodel=medany -static -std=gnu99 -fno-common -fno-builtin-printf
-march=rv32imac -mabi=ilp32 -DMB_ADDR=0x11ffC -O3
--target=riscv32-unknown-elf
--sysroot=/home/llvm/workspace/riscv/riscv-tc-20200316/bin/riscv32-unknown-elf
--gcc-toolchain=/home/llvm/workspace/riscv/riscv-tc-20200316 -o
./build/test_newinst/test_newinst ./src/test_newinst/main.c
./common/syscalls.c ./common/dev.c ./common/crt.S -static -nostartfiles
-lm -lgcc -T ./common/test.ld
./src/test_newinst/main.c:117:5: error: invalid operand for instruction
"lqp %[z], %[x], %[y], 4\n\t"
^
<inline asm>:1:8: note: instantiated into assembly here
lqp a2, a0, a1, 4
*This error is obvious, and due to the Match_InvalidOperand, I wonder why
the changes do not make any effect for LQP instruction.*
*And I have checked the OperandInfo111[] and RISCVInsts[] for the LQP
instruction, *
static const MCOperandInfo OperandInfo111[] = { { RISCV::GPRA0RegClassID,
0, MCOI::OPERAND_REGISTER, 0 }, { RISCV::GPRNOA0A1A2A3RegClassID, 0,
MCOI::OPERAND_REGISTER, 0 }, { RISCV::GPRNOA0A1A2A3RegClassID, 0,
MCOI::OPERAND_REGISTER, 0 }, { -1, 0, RISCVOp::OPERAND_SIMM5, 0 }, };
{ 457, 4, 1, 4, 0, 0|(1ULL<<MCID::MayLoad), 0x12ULL, nullptr, nullptr,
OperandInfo111, -1 ,nullptr }, // Inst #457 = LQP
*Do I miss something or what I can do in this situation?*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200407/0746669a/attachment.html>
More information about the llvm-dev
mailing list