Repository navigation
cpu: report FP results as the MIPS FPU does - #406
Merged
Merged
Conversation
The R4400, R5000 and R10000 use the original MIPS NaN encoding: a NaN with the fraction MSB set is signalling, one with it clear is quiet. The default NaN is 0x7FBFFFFF for that reason. iris used the IEEE 754-2008 convention, the reverse, in the interpreter and in compiled code. The evidence for the original encoding is the NEC VR10000 instruction manual, U12754E, first edition, August 2000, printed page 226, table 2-2, and the results of iris's own R4400 in cpu-tests: fpu/qnan_operand feeds 0x7FC00000 to ADD and gets Invalid and the default NaN, not Unimplemented, and the signalling-predicate compares agree. Upstream issue 361 asked about it. Swap the sense of the fraction-MSB test in is_snan and is_qnan and in their compiled twins, including the compare path, and make the default NaN 0x7FBFFFFF and 0x7FF7FFFF_FFFFFFFF. With the bit set the fraction is nonzero by construction, so the interpreter's signalling test shrinks to one mask and compare, and an ordinary operand skips the subnormal and quiet-NaN checks after one exponent test. Compiled code swaps a comparison and costs the same. The tests that fix a NaN's bit pattern move with this commit. The new ones cover the default NaN as an operand (Unimplemented, no Invalid), a signalling NaN operand (Invalid, trapped and untrapped), and a compiled-versus-interpreter check over ADD, MUL, DIV, SQRT and the quiet and signalling compare predicates. Without this commit the rest of the series still builds and passes its tests; the Octane PROM's ide test then rejects the default NaN's bit pattern.
jamesbraid
force-pushed
the
ide-fpu-fix
branch
2 times, most recently
from
November 2, 2026 15:47
ad267e6 to
c968e9c
Compare
jamesbraid
marked this pull request as draft
November 2, 2026 16:29
jamesbraid
force-pushed
the
ide-fpu-fix
branch
2 times, most recently
from
November 2, 2026 18:54
4474a7f to
4bce574
Compare
The Octane PROM's ide FPU test checks results the host FPU gets wrong for MIPS: infinity minus infinity, zero times infinity, 0/0 and infinity/infinity are Invalid Operations that deliver the default NaN (0x7FBFFFFF), 0/0 is not a divide by zero, and MAX+MAX is Overflow with Inexact. iris returned the host's NaN with no flags. 12 of the 14 subtests passed. Classify those results on every CPU model, in both engines, at no cost to an ordinary result. ADD, SUB, MUL and DIV test the result's magnitude once, in the same branch as a check that neither operand has a zero exponent. That test replaces the old operand and denormal-result tests, and the top of the range, which is where Invalid and Overflow live, goes to a cold classifier shared by both engines. The operand flags those ops used to compute up front are gone: an ordinary result proves there were none, and the classifier finds a divide by zero itself. Compiled code reaches the classifier through one call per instruction. The same call takes a denormal, infinite or NaN operand and a denormal result, where the interpreter's operand checks and denormal rule run, so the instruction carries one call and one trap check, not a copy of those rules at each of two exits. Zero, as an operand or a result, is common in real code and is written without the call. Generated code for the page holding Whetstone and its transcendental kernel comes to 245 bytes per guest instruction, against 400 before this series. SQRT, RECIP, RSQRT, the multiply-add family and CVT.S.D swap the host's NaN for the default NaN and add V. They do it the same way round: one compare of the result's magnitude, and only an infinity or a NaN goes on to classify the operands. In compiled RECIP, RSQRT and the multiply-add family the compare shares the branch these ops already took on a stale Cause, so the ordinary path gains no branch. A finite result cannot come from a NaN operand, a zero divisor or a negative square root, so the signalling-NaN tests on up to three operands no longer run for ordinary arithmetic. CVT.S.D checks its source for a signalling NaN only after the conversion has already left the exact path, since a NaN is never equal to itself. Host instructions per FP op for this and the next commit together, perf stat instructions:u, ordinary operands, ship profile, against the commit before this series (slope over a 120-op region, three alternating rounds, counts do not move between rounds). R10000 numbers; R5000 and R4400 differ by a constant: compiled add.d, mul.d, div.d 36 -> 32 compiled sqrt.s, cvt.s.d no change interpreter add.d, mul.d, div.d 197-200 -> 171 interpreter madd/msub/nmadd/nmsub 168-181 -> 153-159 interpreter recip.s, rsqrt.s 145 -> 136, 153 -> 137 interpreter sqrt.s, cvt.s.d no change Every measured op, in both engines, on the R4400, R5000 and R10000, is within 0.1 instruction of the old count or cheaper. The compiled madd/msub/nmadd/nmsub, recip and rsqrt were recounted from the generated code of a 120-op run of each (R10000, host instructions on the ordinary path, before -> after): madd.d 52 -> 24, recip.s 39 -> 24, rsqrt.s 48 -> 27, with the same three conditional branches per op as before. The tests that check results run on the ordinary passthrough models in both engines, and compiled code is checked against the interpreter for each rule. This commit delivers the IEEE 754-2008 default NaN, 0x7FC00000. The Octane's R10000 delivers 0x7FBFFFFF; the next commit moves the NaN encoding, and with it the ide test's NaN checks pass.
With FS set and neither Underflow nor Inexact enabled, an R10000 replaces a denormal ADD/SUB/MUL/DIV/SQRT result with a signed zero and sets Cause and Flag U|I. The interpreter does that. Compiled code trapped with Unimplemented Operation instead, for every denormal result, whatever FS said. The trap decision negated the "FS is set" compare with bnot. A Cranelift compare yields 0 or 1, and bnot(1) is 0xFE, which still reads as true, so "FS clear or U|I enabled" was always true. The same mistake in the round-to-integer path was fixed earlier. This one uses a compare against zero too. The new equivalence test halves the smallest normal number in both formats with FS set, FS clear and FS set with Underflow enabled, and checks the compiled result against the interpreter and the manual. The trap and flush blocks are marked cold too. They hang off the cold denormal block but were not cold themselves, so Cranelift laid them out among the hot instructions of every FP op. On the benchmark suite that cost fpu/linpack its fast mode in more runs (median 565 against 619 MIPS on main) and left fpu/transcend about 1% behind main; with the blocks cold linpack matches main and transcend is 8% ahead.
jamesbraid
force-pushed
the
ide-fpu-fix
branch
4 times, most recently
from
November 3, 2026 03:32
29c9f8e to
18b942b
Compare
jamesbraid
force-pushed
the
ide-fpu-fix
branch
5 times, most recently
from
November 7, 2026 03:34
0927da0 to
11a418d
Compare
jamesbraid
changed the title
jamesbraid
force-pushed
the
ide-fpu-fix
branch
from
November 7, 2026 19:35
11a418d to
cfb5403
Compare
jamesbraid
marked this pull request as ready for review
November 7, 2026 20:14
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The FPU handlers returned host results where the R4400, R5000 and R10000 differ, and compiled code and the interpreter disagreed in places. This makes both follow the hardware, and removes a slow path in compiled FP code.
Fixes
ideFPU test needs this.Testing
cargo test --features jitv2 --libpasses on every commit (release) and on the tip in release and debug.Performance
Real clock, guest MIPS (higher is faster), median of 10 interleaved runs each on one pinned CCD. A positive change means this branch is faster:
The FP kernels are significant (p < 0.01) and the controls do not move. Main's whetstone and transcend also fall into a slow mode in some runs; this branch does not. Host-instruction counts in deterministic mode fall far more (whetstone -45%), but that counts JIT compile work that normally runs on other threads, so wall time is the better measure.