Implement RVV single-width floating-point conversions

[alanhc]

Implements the single-width conversions from #734. First of three PRs that
share the VFUNARY0 decode site (funct6=0x12, funct3=0x1); see the note at
the end.

Adds the six same-width vfcvt forms (V 1.0 §13.17). The .rtz variants
pin round-towards-zero through the _r_minMag SoftFloat entry points, while
the plain forms follow the dynamic rounding mode installed by
rvv_fp_begin_round().

The widening (vfwcvt, vs1 0x08-0x0f) and narrowing (vfncvt, vs1
0x10-0x17) members are not decoded here and continue to be rejected rather
than misdecoded; the decode tests assert that.

One note on the smoke coverage: the test uses 1.5 as an input because it
rounds to 2 under RNE but truncates to 1 under RTZ. A value such as 2.5
gives 2 under both because of ties-to-even and would not distinguish an
implementation that ignored the .rtz suffix. Signed vs unsigned is pinned
with -3, which differs in both directions.

make check passes; the build was checked in the default, EXT_V,
JIT+EXT_V and EXT_V-without-EXT_F configurations. clang-format 20.1.7
clean.

Ordering: this touches the same case 1: block as the vfwcvt and
vfncvt PRs. They are independent in substance but will conflict textually,
so they are best taken in order (this one, then vfwcvt, then vfncvt); I will
rebase the later two as each lands and update the "not implemented yet"
comments accordingly.

---

Summary by cubic

Implements the six same-width vfcvt conversion instructions (V 1.0 §13.17), which were previously rejected. The .rtz variants pin round-towards-zero; the plain forms follow the dynamic rounding mode. Widening and narrowing forms remain unimplemented and are still rejected.

• Adds smoke tests that use 1.5 to distinguish dynamic rounding from .rtz behavior.
• Shares the VFUNARY0 decode site with future vfwcvt/vfncvt PRs; those will need rebasing in sequence.

^{Written for commit 0c25662. Summary will update on new commits.}

[cubic-dev-ai]

2 issues found across 5 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/rv32_v_template.c">

<violation number="1" location="src/rv32_v_template.c:6707">
P2: Float-to-integer conversions pass `exact=false`, which suppresses SoftFloat’s inexact exception. Set `exact=true` for all four float-to-integer helpers so rounded results update `fflags.NX` as required.</violation>
</file>

<file name="tests/rvv-smoke.S">

<violation number="1" location="tests/rvv-smoke.S:1844">
P2: The `vfcvt.xu.f.v` smoke test does not verify its signedness difference: it skips the negative `-1.5` element at offset 8, so a signed `vfcvt.x.f.v` implementation would still pass this block. Assert that offset 8 is zero to test the behavior described by the comment.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: Float-to-integer conversions pass exact=false, which suppresses SoftFloat’s inexact exception. Set exact=true for all four float-to-integer helpers so rounded results update fflags.NX as required.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/rv32_v_template.c, line 6707:

<comment>Float-to-integer conversions pass `exact=false`, which suppresses SoftFloat’s inexact exception. Set `exact=true` for all four float-to-integer helpers so rounded results update `fflags.NX` as required.</comment>

<file context>
@@ -6695,6 +6695,44 @@ static inline uint32_t rvv_fp_fnmsub32(uint32_t dest,
+ */
+static inline uint32_t rvv_fp_cvt_xu_f32(uint32_t bits)
+{
+    return f32_to_ui32(rvv_fp32_from_raw(bits), softfloat_roundingMode, false);
+}
+
</file context>

[cubic-dev-ai]

P2: The vfcvt.xu.f.v smoke test does not verify its signedness difference: it skips the negative -1.5 element at offset 8, so a signed vfcvt.x.f.v implementation would still pass this block. Assert that offset 8 is zero to test the behavior described by the comment.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At tests/rvv-smoke.S, line 1844:

<comment>The `vfcvt.xu.f.v` smoke test does not verify its signedness difference: it skips the negative `-1.5` element at offset 8, so a signed `vfcvt.x.f.v` implementation would still pass this block. Assert that offset 8 is zero to test the behavior described by the comment.</comment>

<file context>
@@ -1813,6 +1821,73 @@ _start:
+    assert_reg_imm t1, 2, fail_vfcvt_x_f
+    lw t1, 8(a1)
+    assert_reg_imm t1, -2, fail_vfcvt_x_f
+    lw t1, 12(a1)
+    assert_reg_imm t1, 7, fail_vfcvt_x_f
+
</file context>

Suggested change

	lw t1, 12(a1)
	lw t1, 8(a1)
	assert_reg_imm t1, 0, fail_vfcvt_xu_f
	lw t1, 12(a1)