"""The Ampere -> Maxwell lowering takes IMNMX with an immediate wider than Maxwell's 20 bits (native op 0x817). Three of OoT's vertex programs stopped at it as "native integer min/max mode": IMNMX R, R, 0x7f0004, PT does not fit Maxwell IMNMX's signed 20-bit immediate. Such an immediate now goes into R251 (MOV32I) and the register form is used; one that fits keeps the immediate form. python -I -B tests/test_lower_imnmx_wide_immediate.py [--cxx clang++] """ import argparse import shutil import subprocess import tempfile from pathlib import Path RUN_ALL_ARGS = ["--cxx", "g++"] ROOT = Path(__file__).resolve().parents[1] FIXTURE = r'''#include #include #include #include #include #include "shader_recompiler/frontend/ampere/lower_to_maxwell.h" #define CHECK(x) do { if (!(x)) { std::fprintf(stderr, "FAIL:%d %s\n", __LINE__, #x); std::exit(1); } } while (0) using Shader::Ampere::Instruction; static Instruction insn(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { Instruction i{}; i.words[0] = a; i.words[1] = b; i.words[2] = c; i.words[3] = d; return i; } static const Instruction exit_insn = insn(0x0000794d, 0, 0x03800000, 0x000fea00); // EXIT static Shader::Ampere::LoweredProgram lower(Instruction i) { return Shader::Ampere::LowerToMaxwell(std::vector{i, exit_insn}); } static constexpr uint64_t PT = 7ULL << 16; static uint64_t mov32i(uint32_t d, uint32_t value) { return 0x0100000000000000ULL | PT | d | (15ULL << 12) | (uint64_t{value} << 20); } // Maxwell IMNMX (register form) Rd = min/max(Ra, Rb) by the predicate in 39-42, .U32 unless bit 48 static uint64_t imnmx_r(uint32_t d, uint32_t a, uint32_t b, bool sign = false) { return 0x5c20000000000000ULL | PT | d | (uint64_t{a} << 8) | (uint64_t{b} << 20) | (7ULL << 39) | (uint64_t{sign} << 48); } static uint64_t imnmx_i(uint32_t d, uint32_t a, uint32_t imm) { return 0x3820000000000000ULL | PT | d | (uint64_t{a} << 8) | (uint64_t{imm & 0x7ffff} << 20) | (uint64_t{imm >> 31} << 56) | (7ULL << 39); } int main() { { // IMNMX.U32 R26, R25, 0x7f0004, PT (from the log): the immediate in R251, then the register form const auto r = lower(insn(0x191a7817, 0x007f0004, 0x03800000, 0x000fc400)); CHECK(r); CHECK(r.instructions.size() >= 2); CHECK(r.instructions[0] == mov32i(251, 0x007f0004)); CHECK(r.instructions[1] == imnmx_r(0x1a, 0x19, 251)); } { // R38, R35 and R32, R28 (the other two) CHECK(lower(insn(0x23267817, 0x007f0004, 0x03800000, 0x000fe200)).instructions[1] == imnmx_r(0x26, 0x23, 251)); CHECK(lower(insn(0x1c207817, 0x007f0004, 0x03800000, 0x000fe200)).instructions[1] == imnmx_r(0x20, 0x1c, 251)); } { // an immediate that fits keeps the immediate form, positive and negative const auto r = lower(insn(0x191a7817, 0x0007ffff, 0x03800000, 0x000fc400)); CHECK(r && r.instructions[0] == imnmx_i(0x1a, 0x19, 0x0007ffff)); const auto n = lower(insn(0x191a7817, 0xfffffff0, 0x03800000, 0x000fc400)); CHECK(n && n.instructions[0] == imnmx_i(0x1a, 0x19, 0xfffffff0)); } { // the boundaries: 0x00080000 and 0xfff7ffff do not fit a signed 20-bit immediate, 0xfff80000 does const auto up = lower(insn(0x191a7817, 0x00080000, 0x03800000, 0x000fc400)); CHECK(up && up.instructions[0] == mov32i(251, 0x00080000) && up.instructions[1] == imnmx_r(0x1a, 0x19, 251)); const auto down = lower(insn(0x191a7817, 0xfff7ffff, 0x03800000, 0x000fc400)); CHECK(down && down.instructions[0] == mov32i(251, 0xfff7ffff) && down.instructions[1] == imnmx_r(0x1a, 0x19, 251)); const auto fits = lower(insn(0x191a7817, 0xfff80000, 0x03800000, 0x000fc400)); CHECK(fits && fits.instructions[0] == imnmx_i(0x1a, 0x19, 0xfff80000)); } { // .S32 (word 2 bit 9, native bit 73): bit 48 in the register form too const auto r = lower(insn(0x191a7817, 0x007f0004, 0x03800200, 0x000fc400)); CHECK(r && r.instructions[1] == imnmx_r(0x1a, 0x19, 251, true)); } std::printf("PASS: IMNMX with a wide immediate lowers through R251\n"); } ''' def main(): p = argparse.ArgumentParser(description=__doc__) p.add_argument("--cxx", default=shutil.which("clang++") and "clang++" or "g++") a = p.parse_args() with tempfile.TemporaryDirectory(prefix="lower-imnmx-") as tmp: cpp = Path(tmp) / "fixture.cpp" exe = Path(tmp) / "fixture" cpp.write_text(FIXTURE, encoding="utf-8") subprocess.run([a.cxx, "-std=c++20", "-O1", "-Wno-everything" if "clang" in a.cxx else "-w", "-I" + str(ROOT / "gpu" / "vk" / "eden"), "-I" + str(ROOT / "gpu" / "vk" / "compat"), str(cpp), "-o", str(exe)], check=True) subprocess.run([str(exe)], check=True) if __name__ == "__main__": main()