armsx/ARMSX2
0
0
Fork 0
mirror of https://github.com/ARMSX2/ARMSX2.git synced 2026-08-24 16:50:16 -07:00
Code Issues Packages Projects Releases Wiki Activity

Tests: the guarded add/sub skip could not fire

Browse Source 
The randomized test skipped pairs whose result leaves the EE's range by
testing the model's output for an exponent of 255. The model computes in
host singles under chop rounding, which saturates rather than
overflowing, so the skip never fired: a pair whose exact difference sits
above FLT_MAX was asserted against a value the model cannot hold, and
the interpreter's top-binade answer read as a failure.

The skip now reads the saturated word, and the pair that found it is a
row of its own.
This commit is contained in:
pstef
2026-08-09 11:21:25 +02:00
parent 0a0633e551
commit 152b7a9384
1 changed files with 42 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ctest/core/recompilers/ee_fpu_guarded_addsub_console_tests.cpp
+42 -5
View File
@@ -183,6 +183,15 @@ __noinline u32 ModelGuardedAddSub(u32 a, u32 b, bool issub)
return rb;
}
// Host singles stop one binade below the EE, and chop rounding saturates rather
// than overflowing, so a result at or above 2^128 comes back +-0x7F7FFFFF. Both
// engines carry such a result further -- the interpreter into the top binade,
// the fast path to +-FLT_MAX -- which is EeFpuTopBinadeConsole's subject.
bool OutsideTheModelsRange(u32 want)
{
return (want & 0x7FFFFFFFu) >= 0x7F7FFFFFu;
}
__noinline u32 ModelPlainAddSub(u32 a, u32 b, bool issub)
{
float fa, fb;
@@ -279,7 +288,7 @@ TEST(EeFpuGuardedAddSubConsole, RandomisedAddSubMatchesTheModelOnBothEngines)
std::uniform_int_distribution<u32> pick(0, 1);
int checked = 0, masked_rows = 0, cliff_rows = 0, boundary_rows = 0;
int underflow_rows = 0;
int underflow_rows = 0, overflow_rows = 0;
for (int i = 0; i < 3000; ++i)
{
const int ea = exp_dist(rng);
@@ -292,10 +301,12 @@ TEST(EeFpuGuardedAddSubConsole, RandomisedAddSubMatchesTheModelOnBothEngines)
const u32 want_add = ModelGuardedAddSub(a, b, false);
const u32 want_sub = ModelGuardedAddSub(a, b, true);
// Skip pairs whose result leaves the EE's range: the result clamp then
// dominates and the guard mask is unobservable.
if ((want_add & 0x7F800000u) >= 0x7F800000u || (want_sub & 0x7F800000u) >= 0x7F800000u)
// Counted, see the bound below.
if (OutsideTheModelsRange(want_add) || OutsideTheModelsRange(want_sub))
{
++overflow_rows;
continue;
}
// Same at the bottom: below 2^-126 the console keeps an add/sub result's
// mantissa bits instead of flushing (EeFpuUnderflowConsole), and the
// model above computes in host floats, which FZ has already flushed.
@@ -335,10 +346,36 @@ TEST(EeFpuGuardedAddSubConsole, RandomisedAddSubMatchesTheModelOnBothEngines)
EXPECT_GT(checked, 100) << "anti-vacuity: on no pair did the guard change the "
"answer, so this test would pass with the masking "
"deleted";
// If the skip swallows more than a handful of the 3000 pairs, the
// If either skip swallows more than a handful of the 3000 pairs, the
// generator's exponent distribution moved.
EXPECT_LT(underflow_rows, 100) << "the underflow skip swallowed " << underflow_rows
<< " pairs; it is meant to be a handful";
EXPECT_LT(overflow_rows, 100) << "the out-of-range skip swallowed " << overflow_rows
<< " pairs; it is meant to be a handful";
}
// ---------------------------------------------------------------------------
// The skip's own row. sub.s of these two is 0x1.1c1bac8p+128, above every host
// single, so the model saturates where the interpreter encodes the top binade.
// |diff| is 1, so nothing is masked: the row is outside the model, not a
// guard-bit case.
// ---------------------------------------------------------------------------
TEST(EeFpuGuardedAddSubConsole, ResultsAboveTheHostSingleRangeAreSkipped)
{
const ScopedEeFpEnv fp_env;
constexpr u32 kA = 0x7ED69A1Du; // exponent 253
constexpr u32 kB = 0xFF30CE9Eu; // exponent 254, negative
float fa, fb;
std::memcpy(&fa, &kA, 4);
std::memcpy(&fb, &kB, 4);
EXPECT_GT(static_cast<double>(fa) - static_cast<double>(fb), 0x1.fffffep127);
const u32 want = ModelGuardedAddSub(kA, kB, true);
EXPECT_EQ(want, 0x7F7FFFFFu);
EXPECT_TRUE(OutsideTheModelsRange(want));
EXPECT_EQ(RunOne(FORM_SUB, kA, kB, 0, false), 0x7F8E0DD6u) << "interp: top binade";
EXPECT_EQ(RunOne(FORM_SUB, kA, kB, 0, true), 0x7F7FFFFFu) << "jit: saturates";
}
// ---------------------------------------------------------------------------