llvmbot
commented
2 weeks ago @llvm/issue-subscribers-backend-risc-v
Author: None (ChiHungWei)
Ubuntu 24.04.1 LTS on x86_64, LLVM 20.0.0git with riscv-gnu-toolchain installed at /opt/riscv_2
The following problem occurs while cross compiling [ggml.c](https://github.com/ggerganov/llama.cpp/blob/master/ggml/src/ggml.c) ,
```
clang: /home/tomlord/workspace/jason/llvm-project/llvm/include/llvm/CodeGen/MachineOperand.h:557: int64_t llvm::MachineOperand::getImm() const: Assertion `isImm() && "Wrong MachineOperand accessor"' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace, preprocessed source, and associated run script.
Stack dump:
1. <eof> parser at end of file
2. Code generation
3. Running pass 'Function Pass Manager' on module 'ggml/src/ggml.c'.
4. Running pass 'Prologue/Epilogue Insertion & Frame Finalization' on function '@ggml_compute_forward_conv_transpose_2d'
```
Compiling flags(under the directory of `/llama.cpp`):
`/home/tomlord/workspace/jason/llvm-project/riscv-custom/bin/clang -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon -I/home/tomlord/workspace/jason/llvm-project/riscv-custom/include -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -D_GLIBCXX_ASSERTIONS -DGGML_USE_LLAMAFILE --gcc-toolchain=/opt/riscv_2 --sysroot=/opt/riscv_2/sysroot --target=riscv64-unknown-linux-gnu -fopenmp=libomp -v -std=c11 -fuse-ld=lld -fPIC -march=rv64gc -mabi=lp64d -O0 -g -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int -Werror=implicit-function-declaration -pthread -Wunreachable-code-break -Wunreachable-code-return -Wdouble-promotion -c ggml/src/ggml.c -o ggml/src/ggml.o`
Reduced .bc file that bugpoint emits:
[https://github.com/ChiHungWei/For-LLVM-Bug-Report](https://github.com/ChiHungWei/For-LLVM-Bug-Report)
I've added a custom instruction in [llvm-project/llvm/lib/Target/RISCV/RISCVInstrInfo.td](https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/RISCV/RISCVInstrInfo.td) as following,
```
//===----------------------------------------------------------------------===//
// Instruction class templates
//===----------------------------------------------------------------------===//
let hasSideEffects = 0, mayLoad = 0, mayStore = 0 in
class ALU_rrr<bits<2> funct2, bits<3> funct3, string opcodestr,
bit Commutable = 0>
: RVInstR4<funct2, funct3, OPC_OP, (outs GPR:$rd), (ins GPR:$rs1, GPR:$rs2, GPR:$rs3),
opcodestr, "$rd, $rs1, $rs2, $rs3"> {
let isCommutable = Commutable;
}
//===----------------------------------------------------------------------===//
//Instructions
//===----------------------------------------------------------------------===//
def MULADD : ALU_rrr<0b10, 0b100,"muladd">,
Sched<[WriteIMul, ReadIMul, ReadIMul]>;
//===----------------------------------------------------------------------===//
//codegen pattern
//===----------------------------------------------------------------------===//
def : Pat< (add (mul (XLenVT GPR:$rs1), (XLenVT GPR:$rs2)), GPR:$rs3),
(MULADD GPR:$rs1, GPR:$rs2, GPR:$rs3) >;
```
Also, in [llvm-project/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp](https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp). I add a case inside the `switch (Opcode)` of function `void RISCVDAGToDAGISel::Select(SDNode *Node)` to help select the new custom instruction "MULADD",
```
case ISD::ADD: {
SDValue opn_0 = Node->getOperand(0);
SDValue opn_1 = Node->getOperand(1);
//None of the operand comes from the result of mul
if(!(opn_0.getOpcode() == ISD::MUL || opn_1.getOpcode() == ISD::MUL)){
break;
}
//operand 0 is the mul node
else if(opn_0.getOpcode() == ISD::MUL){
SDValue rs1 = opn_0.getNode()->getOperand(0);
SDValue rs2 = opn_0.getNode()->getOperand(1);
SDValue rs3 = opn_1;
SDNode *mul_node = opn_0.getNode();
//create a new node with muladd target-specific instruction
SDNode *muladd = CurDAG->getMachineNode(RISCV::MULADD, DL, VT, rs1,rs2,rs3);
ReplaceNode(Node, muladd);
CurDAG->RemoveDeadNode(mul_node);
return;
}
//operand 1 is the mul node
else {
SDValue rs1 = opn_1.getNode()->getOperand(0);
SDValue rs2 = opn_1.getNode()->getOperand(1);
SDValue rs3 = opn_0;
SDNode *mul_node = opn_1.getNode();
//create a new node with muladd target-specific instruction
SDNode *muladd = CurDAG->getMachineNode(RISCV::MULADD, DL, VT, rs1,rs2,rs3);
ReplaceNode(Node, muladd);
CurDAG->RemoveDeadNode(mul_node);
return;
}
}
```
The newly defined instruction works well in a simple c file like this,
```
#include <stdlib.h>
#include <stdio.h>
int main() {
int a = 3;
int b = 103;
int c = 127;
a = a * b + c;
printf("The answer of 3*103 + 127 = %d", a);
return 0;
}
```
However, it fails when compiling` ggml.c ` as shown above.
I've surveyed vendor extension like [RISCVInstrInfoXTHead.td](https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/RISCV/RISCVInstrInfoXTHead.td) and [RISCVInstrInfoXVentana.td](https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/RISCV/RISCVInstrInfoXVentana.td), and they use roughly the same way defining custom instructions. So I am confused what causes this issue.
topperc
Ubuntu 24.04.1 LTS on x86_64, LLVM 20.0.0git with riscv-gnu-toolchain installed at /opt/riscv_2
The following problem occurs while cross compiling ggml.c ,
Compiling flags(under the directory of
/llama.cpp):/home/tomlord/workspace/jason/llvm-project/riscv-custom/bin/clang -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon -I/home/tomlord/workspace/jason/llvm-project/riscv-custom/include -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -D_GLIBCXX_ASSERTIONS -DGGML_USE_LLAMAFILE --gcc-toolchain=/opt/riscv_2 --sysroot=/opt/riscv_2/sysroot --target=riscv64-unknown-linux-gnu -fopenmp=libomp -v -std=c11 -fuse-ld=lld -fPIC -march=rv64gc -mabi=lp64d -O0 -g -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int -Werror=implicit-function-declaration -pthread -Wunreachable-code-break -Wunreachable-code-return -Wdouble-promotion -c ggml/src/ggml.c -o ggml/src/ggml.oReduced .bc file that bugpoint emits: https://github.com/ChiHungWei/For-LLVM-Bug-Report
I've added a custom instruction in llvm-project/llvm/lib/Target/RISCV/RISCVInstrInfo.td as following,
Also, in llvm-project/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp. I add a case inside the
switch (Opcode)of functionvoid RISCVDAGToDAGISel::Select(SDNode *Node)to help select the new custom instruction "MULADD",The newly defined instruction works well in a simple c file like this,
However, it fails when compiling
ggml.cas shown above. I've surveyed vendor extension like RISCVInstrInfoXTHead.td and RISCVInstrInfoXVentana.td, and they use roughly the same way defining custom instructions. So I am confused what causes this issue.