ClassifyInstruction assumed a ModRM byte right after every opcode that
was not a near branch. A length-changing rewrite therefore adjusted
immediates such as MOVABS imm64 that looked like a RIP-relative ModRM,
and missed RIP-relative operands of VEX and 0F 38 / 0F 3A instructions
and the rel8 targets of LOOP and JRCXZ. Take the RIP-relative
displacement and branch target from DecodeInstruction instead.
Guest C built with -fexceptions binds __gcc_personality_v0 to
__gcc_personality_v0_nid_postfix, and the unwinder recognizes the frame
by that address. exception_c_cleanup links the host personality, so it
passed without the export.
The new test builds a call-site table by hand and drives the routine the
way the unwinder does: through the _nid_postfix symbol and through
CallPersonality. It covers the cleanup phase installing the landing pad,
the search phase continuing, ranges before and after the call site, a
call site with no landing pad, no LSDA, a signal frame, an unsupported
version, and an unknown personality. Without the export it fails to link.
On macOS the host keeps libc++abi: libc++ binds __cxa_throw two-level,
so moving only the personality into this runtime would leave the two
halves disagreeing on the exception header. The host personalities stay
libc++abi's and libunwind's, and libc.prx serves guest code:
- the guest's libstdc++ std::exception_ptr is modelled directly, since
libc++'s differs, and the Itanium type_info layouts that libc++abi
keeps private are declared for the runtime;
- guest frames are found through their PT_GNU_EH_FRAME, using a
dl_iterate_phdr emulation over the ELF program headers that relinked
images keep in __APS5DATA,__meta (written by relinker --macos, which
comes in a later pull request);
- host frames are found by scanning __TEXT,__eh_frame;
- sceKernelGetModuleInfoForUnwind uses the same emulation and falls back
to dladdr for host images.
The libc++ paths compile only with libc++/libc++abi and the rest only
on __APPLE__, so Linux and Windows builds are unchanged.
exception_c_cleanup throws from host code, which this runtime does not
serve on macOS, so it is not built there.
Guest C built with -fexceptions imports __gcc_personality_v0 for its
cleanup frames, and libc.prx did not export it. The routine becomes
__gcc_personality_v0_nid_postfix, the host __gcc_personality_v0
forwards to it as __gxx_personality_v0 already does, and
CallPersonality recognizes both names.
The GFX10 lowering of a wave64 reduction scans rows with DPP row_shr
1/2/4/8 and v_permlanex16_b32 under EXEC -1, then joins lanes 31 and 63
with a scalar op. The console fills lanes outside the entry EXEC with
the identity, but on the host they can be helpers that sit out subgroup
operations or invocations that were never launched: on 64 lanes (RADV)
the emulated chain reads 0 or stale values from them, and on 32 lanes
lane 63 throws since #997.
EmitReadLane now recognises the sequence on any host for UMin, UMax,
SMin, SMax, IAdd, AND and OR when the keys are the identity outside the
entry EXEC. On 64 lanes, lanes 31 and 63 become subgroup reductions of
the keys of invocations 0-31 and 32-63; on 32 lanes or fewer, lane 31
reduces all keys and lane 63 is the identity. It throws without
subgroup arithmetic or on a subgroup wider than the wave. VulkanDevice
reports GroupNonUniformArithmetic and ShaderValidation checks it.
guest_formatting_tests runs only on Windows and windows_exception_tests only
exists there, so the stale-entry check failed the Windows job.
Co-Authored-By: Claude <[email protected]>
sceKernelGetProcessTime and sceKernelGetProcessTimeCounter subtracted
GetStartNanos() from GetMonotonicNanos() in one expression. GCC reads the
clock first, so on the first call in a process the start was captured
after it and the result wrapped to about 2^64.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
UD1 (0F B9 /r) always raises #UD, like UD2, but DecodeInstruction
reported it as sequential. Reachability analysis then followed the
bytes after it, so an import reachable only through that fallthrough
was kept by the strict unused NID filter. UD1 keeps its ModRM, so a
RIP-relative operand is still reported.
#1398 and #1399 both rewrite the sceRtcParseDateTime entry that the s_sethalt entry followed. It now follows the libSceAudio3d entries, which no open pull request edits, so either can merge first without a conflict.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#1372 adds its test after the vop1 decode tests and #1314 its debt entry after s_setkill, the places where this branch added its own. The wave control decode test now precedes the vop1 decode tests and the s_sethalt entry follows sceRtcParseDateTime, so either pull request can merge first without a conflict.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The test is added to ctest in its own BUILD_TESTING block, as agc_vop1_sdwa_move_decode is, instead of in the shared test list that #1187 also extends.
s_sethalt with SIMM16[0] clear is a no-op; s_sethalt 1, s_sendmsghalt, a reached s_code_end, s_rfe_b64, the ds_gws_* instructions and ds_ordered_count throw with the reason at translation. Linear decoding stops at s_code_end like at s_endpgm, so code padded with s_code_end after a final backward branch decodes. Encodings from LLVM SOPInstructions.td and DSInstructions.td, checked against the gfx10 MC tests; semantics from the RDNA 2 ISA guide.
The Windows job failed the check with two names that are not test
executables. `libhost_thread_local_test_library` is a support library
whose RUNTIME_OUTPUT_DIRECTORY puts it in the test directory; on Linux
its output is a library artifact and never lands there, which is why the
Linux job stayed green. `windows_exception_tests` is a MINGW-only test
executable that no `add_test` registers.
Ninja prints the rule beside each target, and CMake links an executable
with `CXX_EXECUTABLE_LINKER__<target>_<config>` and a shared library with
`CXX_SHARED_LIBRARY_LINKER__...`, on every platform. Requiring EXECUTABLE
in the rule leaves the library out and keeps the Linux count at 362.
`windows_exception_tests` is a real instance of the class this check
looks for, so record it beside the other five rather than fail the job
that found it.
The Windows job failed on the new step with
"no test executables found under build". Ninja prints the target names
with the platform separator, so the Windows build graph lists
tests\name_tests.exe where Linux lists tests/name_tests, and the reader
matched neither the backslash nor the .exe suffix.
Accept both separators and a .exe suffix. Verified by driving the
function with recorded Ninja output for both platforms, which now
return the same names, and by re-running the check against the Linux
tree, which still reports 362 test executables and the same five.
A test executable can be declared with add_test_executable and never
registered with add_test. It is then built and never run, and ctest
reports a full pass over what it knows, so nothing in CI says so.
Measured on main, five test executables are in this state: the driver's
mesh and recorder tests, the guest formatting tests, the video out flip
tests and the BDA device tests.
tools/check_test_registration.py compares the test executables CMake
builds, read from the build graph, against the add_test calls in the
generated CTestTestfile.cmake files, which are written at configure time
and so do not depend on the build having finished. It fails on a test
executable that nothing runs, and records those five by name so the job
stays green until they are dealt with.
Both jobs run it right after configure, before anything is built.
DecodeState and the draw precheck rejected DB_SHADER_CONTROL bit 16,
PRIMITIVE_ORDERED_PIXEL_SHADER (POPS), which runs the ordered section
of overlapping pixels' shaders in primitive order. The bit now reaches
the recompiler as a pixel stage field, and such a pixel shader declares
PixelInterlockOrderedEXT and runs its whole body between
OpBeginInvocationInterlockEXT and OpEndInvocationInterlockEXT; an
invocation that a BDA read fault kills ends the interlock first. The
POPS collision SGPRs read 0, so a title skips its POPS wait and runs
its per-collision loop once. An image_load still reads a sampled
binding, which misses an earlier overlapping primitive's image_store in
the same draw. The device enables VK_EXT_fragment_shader_interlock when
it has fragmentShaderPixelInterlock; without it the recompile throws,
as does a nonzero POPS_OVERLAP_NUM_SAMPLES. The field is part of the
recompile cache key, the draw key and the serialized request (version 10).
xsvm