https://bugs.kde.org/show_bug.cgi?id=520753
Bug ID: 520753
Summary: Advertise LZCNT via CPUID for x86 (32-bit) clients
Classification: Developer tools
Product: valgrind
Version First 3.27 GIT
Reported In:
Platform: Other
OS: Linux
Status: REPORTED
Severity: normal
Priority: NOR
Component: vex
Assignee: [email protected]
Reporter: [email protected]
Target Milestone: ---
LZCNT is not advertised via CPUID when client is a x86 (32-bit) binary
$ ./vg-in-place cpuid32 -1 |& fgrep LZCNT
LZCNT advanced bit manipulation = false
$ cpuid32 -1 |& fgrep LZCNT
LZCNT advanced bit manipulation = true
$
Per Intel manual, LZCNT was introduced in "Intel® Xeon® processor E3/E5/E7 v3
product families, 4th Generation Intel® Core™ processor family". In Valgrind
guest_amd64_helpers.c, LZCNT is advertised via amd64g_dirtyhelper_CPUID_avx2()
ECX, bit 5:
3652 case 0x80000001:
3653 SET_ABCD(0x00000000, 0x00000000, 0x00000021, 0x2c100800);
-------------------------------------------------------^--------------- <<<
here
For 32-bit clients, Valgrind has guest_x86_helpers.c. The dirtyhelper_CPUID*
function is being selected around guest_x86_toIR.c:15516,
x86g_dirtyhelper_CPUID_sse3() being the "most modern" dirtyhelper. For LZCNT
introduction:
SSE (1999) → Pentium III
SSE2 (2000) → Pentium 4
SSE3 (2004) → Pentium 4 Prescott
SSSE3 (2006) → Core 2 ← this is what sse3 helper emulates
SSE4.1 (2006) → Core 2 (Penryn)
SSE4.2 (2008) → Core i7 (Nehalem) ← POPCNT introduced here
AVX (2011) → Sandy Bridge
BMI1/BMI2 (2013) → Haswell ← LZCNT introduced here
In Valgrind, LZCNT is not SW emulated (like POPCNT is) but it also isn't
directly run in the HW. Instead, in gen_LZCNT() it is treated as Iop_ClzNat32,
which translates to X86Instr_Bsfr32(False, ...) in host_x86_isel.c:1321, which
is the BSR (emit_X86Instr() in VEX/priv/host_x86_defs.c:2706). That's a
"baseline x86" instruction that's not even advertised via CPUID (introduced
probably in i386 in 1985).
That said, since Valgrind's LZCNT could be advertised
x86g_dirtyhelper_CPUID_sse0() technically (because BSR is so old), given that
LZCNT was introduced in Haswell, a good place for it is
x86g_dirtyhelper_CPUID_sse3(). Another option is to create a new
dirtyhelper_CPUID function for it. Tha attached patch uses the former
approach.
--
You are receiving this mail because:
You are watching all bug changes.
