[AMDGPU] v_dot4_i32_i8 op_sel not a valid operand #163647
Description
I'm trying to upgrade the version of LLVM compiler used in Debian, but a rocBLAS assembly file that built successfully on LLVM 19 and older is failing on LLVM 20 and newer. I've had to change the file name to *.txt in order to attach it to this issue: Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.txt
Success (LLVM 19):
# /usr/bin/clang++-19 -v -x assembler -target amdgcn-amd-amdhsa -mcode-object-version=4 -mcpu=gfx908 -mwavefrontsize64 -c -o Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.o Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.s
Debian clang version 19.1.7 (7)
Target: amdgcn-amd-amdhsa
Thread model: posix
InstalledDir: /usr/lib/llvm-19/bin
Found HIP installation: /usr, version 5.7.31921
(in-process)
"/usr/lib/llvm-19/bin/clang" -cc1as -mllvm --amdhsa-code-object-version=4 -triple amdgcn-amd-amdhsa -filetype obj -main-file-name Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.s -target-cpu gfx908 -target-feature +wavefrontsize64 -fdebug-compilation-dir=/root -dwarf-debug-producer "Debian clang version 19.1.7 (7)" -dwarf-version=5 -mrelocation-model pic -o Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.o Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.s
Failure (LLVM 21):
# /usr/bin/clang++-21 -v -x assembler -target amdgcn-amd-amdhsa -mcode-object-version=4 -mcpu=gfx908 -mwavefrontsize64 -c -o Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.o Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.s
Debian clang version 21.1.3 (1)
Target: amdgcn-amd-amdhsa
Thread model: posix
InstalledDir: /usr/lib/llvm-21/bin
Found HIP installation: /usr, version 5.7.31921
(in-process)
"/usr/lib/llvm-21/bin/clang" -cc1as -mllvm --amdhsa-code-object-version=4 -triple amdgcn-amd-amdhsa -filetype obj -main-file-name Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.s -target-cpu gfx908 -target-feature +wavefrontsize64 -fdebug-compilation-dir=/root -dwarf-debug-producer "Debian clang version 21.1.3 (1)" -dwarf-version=5 -mrelocation-model pic -o Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.o Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.s
<instantiation>:2:98: error: not a valid operand.
v_dot4_i32_i8 v[vgprValuC+0+0*8], v[vgprValuA_X0_I0+0], v[vgprValuB_X0_I0+0], v[vgprValuC+0+0*8] op_sel:[0,0] op_sel_hi:[1,1] //valuC[0]
^
Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.s:1472:1: note: while in macro instantiation
MAC_8x4_X0
^
<...>
<instantiation>:34:98: error: not a valid operand.
v_dot4_i32_i8 v[vgprValuC+7+3*8], v[vgprValuA_X0_I0+7], v[vgprValuB_X0_I0+3], v[vgprValuC+7+3*8] op_sel:[0,0] op_sel_hi:[1,1] //valuC[31]
^
Cijk_Ailk_Bjlk_4xi8II_BH_MT128x64x16_SN_1LDSB0_ASNjw6fz7pYwcB794PZbU_bbVKWbDcNIoVGGo9GD9Mvw=.s:2442:1: note: while in macro instantiation
MAC_8x4_X0
^