aco: update SMEM offset workaround for LLVM 15

This isn't needed since LLVM 15's b0ccf38b018.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-By: Tatsuyuki Ishi <ishitatsuyuki@gmail.com>
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6663
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17036>
This commit is contained in:
Rhys Perry 2022-06-14 20:25:24 +01:00 committed by Marge Bot
parent d301883aba
commit 33e7ba2e3e
2 changed files with 14 additions and 5 deletions

View File

@ -308,22 +308,31 @@ disasm_instr(amd_gfx_level gfx_level, LLVMDisasmContextRef disasm, uint32_t* bin
size = l / 4;
}
#if LLVM_VERSION_MAJOR <= 14
/* See: https://github.com/GPUOpen-Tools/radeon_gpu_profiler/issues/65 and
* https://github.com/llvm/llvm-project/issues/38652
*/
if (gfx_level == GFX9 && (binary[pos] & 0xfc024000) == 0xc0024000) {
if (invalid) {
/* do nothing */
} else if (gfx_level == GFX9 && (binary[pos] & 0xfc024000) == 0xc0024000) {
/* SMEM with IMM=1 and SOE=1: LLVM ignores SOFFSET */
size_t len = strlen(outline);
snprintf(outline + len, outline_size - len, ", s%u", binary[pos + 1] >> 25);
char imm[16] = {0};
while (outline[--len] != ' ') ;
strncpy(imm, outline + len + 1, sizeof(imm) - 1);
snprintf(outline + len, outline_size - len, " s%u offset:%s", binary[pos + 1] >> 25, imm);
} else if (gfx_level >= GFX10 && (binary[pos] & 0xfc000000) == 0xf4000000 &&
(binary[pos + 1] & 0xfe000000) != 0xfa000000) {
/* SMEM non-NULL SOFFSET: LLVM ignores OFFSET */
uint32_t offset = binary[pos + 1] & 0x1fffff;
if (offset) {
size_t len = strlen(outline);
snprintf(outline + len, outline_size - len, ", 0x%x", offset);
snprintf(outline + len, outline_size - len, " offset:0x%x", offset);
}
}
#endif
return std::make_pair(invalid, size);
}

View File

@ -279,8 +279,8 @@ BEGIN_TEST(assembler.smem_offset)
//~gfx10! s_load_dword s7, s[6:7], 0x42 ; f40001c3 fa000042
bld.smem(aco_opcode::s_load_dword, dst, sbase, Operand::c32(0x42));
if (i >= GFX9) {
//~gfx9! s_load_dword s7, s[6:7], 0x42, s5 ; c00241c3 0a000042
//~gfx10! s_load_dword s7, s[6:7], s5, 0x42 ; f40001c3 0a000042
//~gfx9! s_load_dword s7, s[6:7], s5 offset:0x42 ; c00241c3 0a000042
//~gfx10! s_load_dword s7, s[6:7], s5 offset:0x42 ; f40001c3 0a000042
bld.smem(aco_opcode::s_load_dword, dst, sbase, Operand::c32(0x42), offset);
}