Fix ceil(), hypot() and $clog2() codegen

[sai-v-ch]

Three math builtins — ceil(), hypot() and $clog2() — cannot be used at all on mob. This is independent of the VAMS-2023 work in #19; I found it while checking that my own new builtins were declared correctly in CodegenCx::intrinsic, and it's the same root cause I flagged in #20, so I'd suggest taking this one first.

Each fails differently, but all three for one reason: no model in the suite calls them, so the code path is never executed in CI.

What happens today

Minimal model, ceil:

error: internal error: entered unreachable code: intrinsic llvm.ceil.f64 not found

Opcode::Ceil requests llvm.ceil.f64, but it was never declared alongside llvm.floor.f64, and Builder::intrinsic turns a missing declaration into unreachable!().

hypot:

Incorrect number of arguments passed to called function!
  %60 = call double @hypot(double %25, double %29)

It is declared double(double) while Opcode::Hypot passes two arguments, so the module fails the LLVM verifier.

$clog2 is broken twice over. The spec-conformant single-argument call is rejected up front:

error: invalid argument count: expected 2 arguments but found 1

because it is typed INT_MATH_2. Passing two arguments gets past the type checker and then silently discards the second, since hir_lower only reads args[0]. Fixing the signature then exposes the second problem:

Intrinsic name not mangled correctly for type arguments! Should be: llvm.ctlz.i32
ptr @llvm.ctlz

ctlz is overloaded, so it must carry the type suffix.

The fix

Four small changes: declare llvm.ceil.f64, give hypot its second parameter, retype $clog2 to one integer argument, and use the mangled llvm.ctlz.i32. INT_MATH_2 had no other user, so it becomes INT_MATH_1 rather than adding a second signature next to it.

Why a new integration model

I'd rather not fix these and leave the next one to be found by a user, so this adds MATH_INTRINSICS, which calls every builtin that Builder::intrinsic routes to an LLVM intrinsic or a libm symbol.

Two things it has to get right to be worth anything:

• arguments derive from a branch voltage. With literal arguments, constant folding evaluates the call and codegen never sees it, so the model would pass whether or not the declarations are correct.
• arguments stay in domain — unit = x/sqrt(1+x²) for asin/acos/atanh, upper = 1+x² for ln/log/acosh — so the test fails on a real codegen problem rather than on a NaN.

Reverting any one of the four fixes makes it fail.

Testing

Against LLVM 18.1.8:

• cargo test -p osdi — 27/27 (26 existing models plus the new one)
• cargo test -p openvaf --test integration — 71/71, which compiles the new model to a shared library, dlopens it and processes its parameters
• full workspace, RUN_SLOW_TESTS=1 RUN_DEV_TESTS=1 — green
• cargo fmt --all -- --check — clean

Two pre-existing failures are unrelated and reproduce on a clean mob checkout: sourcegen osdi::gen_osdi_structs panics at sourcegen/src/osdi.rs:200, and the verilogae crate does not compile (CallBackKind::QueryPastState no longer exists). Neither is built by rust-ci.yml.

One note on scope: $clog2 now type-checks one argument, so any model that was passing two to work around the old signature would start erroring. That code was already getting its second argument ignored, so it cannot have been relied on for anything.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.