fix(lift): SDIV of the most negative number by -1 gives the dividend, not a trap #1423
No reviewers
Labels
No labels
area/accessibility
area/audio
area/input
area/install
area/menu
area/renderer
area/saves
area/tooling
area/upscaling
beanstalk: approved
beanstalk: failed
beanstalk: needs approval
beanstalk: running
beanstalk: tested
ci: approved
good first issue
gpu/amd
gpu/igpu
gpu/intel
gpu/nvidia
help wanted
kind
bug
kind
chore
kind
docs
kind
feature
kind
performance
kind
question
needs measurements
platform/linux
platform/macos
platform/steam-deck
platform/windows
preview
priority
critical
priority
high
priority
low
regression
risk
high
risk
low
risk
medium
status
blocked
status
confirmed
status
duplicate
status
invalid
status
needs-info
status
wontfix
ci: approved
No milestone
No project
No assignees
2 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
MagicBeanDev/Longshot!1423
Loading…
Reference in a new issue
No description provided.
Delete branch "Victor/Longshot:fix/sdiv-int-min"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Type: fix · Risk: low · Tested on: Linux, RTX 3050 Ti Laptop 4 GB · Needs: none
Why
Lifted code that divides the most negative number by -1 kills the game with no message. AArch64's SDIV gives the dividend back there and goes on; the C the translator writes for it divides as C does, where that division is undefined, and x86's
idivtraps on it (SIGFPE on Linux,0xC0000095on Windows). Whether the game ever divides like that is not known: if it does, it is a crash nobody can explain from the log. Fixes #361.How to reproduce
Seen on: Linux (Arch), x86-64, gcc and clang; found by the equivalence run described in #361, which had to leave SDIV out because the lifted C stopped on these inputs
Settings and switches: none: no game is needed
Steps:
python -I -B tests/test_sdiv_overflow.pyon main (the check is this PR's first commit). It translatessdiv x0, x1, x2andsdiv w0, w1, w2as the lift does, compiles the C and runs it withINT64_MIN / -1andINT32_MIN / -1Expected: the dividend, as on the console
Actual: the program stops with signal 8 on those inputs, at
-O0and at-O2What changed
nx2/insn.py, which already writes the C for the instruction forms suyu's translator lacks, now writes SDIV's too: suyu's own text with a divisor of -1 taken as a negation in unsigned arithmetic, which wraps to AArch64's value for every dividend. Every other input gives what it gave, andthird_party/a64-kitis not edited (base code plus a patch outside it, as agreed on Discord for this issue).sdiv()innx2/insn.py, reached fromemit()for(w & 0x7FE0FC00) == 0x1AC00C00. A destination of XZR still goes to suyu, which writes no division for it. The text keeps suyu's shape (c->x[d]=(uint64_t)(...), the same register reads), so the lifter's later passes see what they saw.tests/test_sdiv_overflow.py(new): SDIV's X and W forms, a register read as XZR, one register for both operands, and UDIV beside them, each over 20 pairs of values at-O0and-O2, every run a process of its own so a trap names its case.What it costs: one compare more at each SDIV. The game's lifted code has 1,926 of them (384 on X registers, 1,542 on W), counted in a build of main @
dbad7d12. It changes the lifted C, so a game folder is translated and compiled again by its next build.Measurements
Verdict: it passes. The check stops on main at the 8 overflow runs of 280 and passes here, and the game plays as before with it. No speed was measured: one compare at each of 1,926 divisions.
Hardware: Arch Linux (kernel 7.2.11), NVIDIA GeForce RTX 3050 Ti Laptop GPU 4 GB, driver 615.78.08, Intel Core i7-11800H, 31 GB RAM
Scene: none for the check (no game). In the game: normal play from a save, about 8,400 frames presented, closed from the window
Method: the check's output at each commit; the translated C against Unicorn 2.1.4 on the same inputs; in the game, a few minutes of play on a build of this change
How to reproduce:
python -I -B tests/test_sdiv_overflow.py [--cc clang] [--cxx g++]544229a2ed90712ctests/test_sdiv_overflow.pysdiv x0, x1, x2,8000000000000000 / ffffffffffffffff8000000000000000sdiv w0, w1, w2,80000000 / ffffffff0000000080000000The Before column is the check run at this PR's first commit (
277effaa: main's sources with the new check).Flags and settings
None
Testing
Tested:
tests/test_sdiv_overflow.py: fails on main's sources, passes here, with gcc and clang++ and with clang and g++. It fails again with the fix broken three ways: SDIV left to suyu (8 runs), a divisor of -1 giving the dividend back for every dividend (10), and the W result sign-extended (18).The values the check expects are Unicorn's: all 140 of its form and value pairs give the same in Unicorn 2.1.4. Beyond the check, the translated C for the X and W forms gives Unicorn's result on 20,000 pairs of register values, 324 of them edge pairs (a local run, not part of the PR).
python -I -B tests/run_all.pyon Linux: 665 pass, 40 fail, 17 skipped. The 40 are the same 40 that fail on main @544229a2on this PC (664 pass there), and none of them reads a file this PR changes: the 9 checks that do pass or are skipped (nx2/test_insn.pyandnx2/test_leaf.pyneed the game's modules).The 40, failing on main as well
test_bug_report_defaults.py,test_cpu_report.py,test_crt_options.py,test_damage_debug_menu.py,test_display_startup_failure.py,test_dlss_feature_size.py,test_failed_compile_keeps_objects.py,test_framegen_preset.py,test_fsr_uniform_ownership.py,test_guest_imported.py,test_hud_layout_row.py,test_import_reach.py,test_invalidate_vcore.py,test_linux_preflight.py,test_linux_pressure.py,test_linux_update.py,test_mirror_budget.py,test_missing_nup.py,test_mods_size_table.py,test_nonuniform_image_arrays.py,test_parallel_init_order.py,test_pending_guest_import.py,test_pipeline_play_cache.py,test_present_thread.py,test_profiling_mode.py,test_readback_timeout_source.py,test_replay_write_order.py,test_rumble_menu.py,test_run_all_changed.py,test_slow_frames.py,test_staging_burst.py,test_switch_table.py,test_target_image_pool.py,test_tested_configurations.py,test_upscaler_round_trip.py,test_upscaler_rows.py,test_vcore_escape.py,test_vk_occlusion_queries.py,test_vk_swapchain_compat.py,test_walk_key.pyIn the game, on a build of main @
dbad7d12with this change (the lifted code translated and compiled again; all 1,926 signed divisions carry the new text): several minutes of normal play from a save, about 8,400 frames. Everything played as before, nothing stopped, and the game closed normally.Not tested:
Must be tested: None
Risks
What could break: every signed division in the game's lifted code goes through the new text. A mistake in it would show as wrong arithmetic anywhere; the check and the comparison with Unicorn are what stand against that.
Turning it off: revert the commit; there is no switch.
Compatibility: the lifted C changes, so the next build of a game folder translates and compiles the game again. Saves, settings, the shader and pipeline caches and the vk ABI are untouched.
Issues
Fixes #361
Checklist
Everyone
type(scope): description) of the status line's type.can't be run by hand is in Testing's Not testable by hand, with why.
in Risks.
main.python -I -B tests/run_all.pypasses (any failure that also fails onmainis named).third_party/untouched (a new folder with aPROVENANCE.mdinstead).When it applies
Assisted-by:/Co-Authored-By:trailers.mainand passing here.Tested in game: Linux (Arch, kernel 7.2.11), NVIDIA GeForce RTX 3050 Ti Laptop 4 GB, driver 615.78.08. A build of main @
dbad7d12with this change, the lifted code translated and compiled again (all 1,926 signed divisions carry the new text). Several minutes of normal play from a save, about 8,400 frames: everything played as before, nothing stopped, and the game closed normally.WIP: fix(lift): SDIV of the most negative number by -1 gives the dividend, not a trapto fix(lift): SDIV of the most negative number by -1 gives the dividend, not a trapBrought up to the ready-for-review standard (#1454): the checklist's Everyone list has two new boxes, added here from what this PR's description and comments show (please check they match what you know).