Implement RVV floating-point unary instructions

[alanhc]

Implements the last of the five families in #734.

Adds vfsqrt.v, vfrsqrt7.v, vfrec7.v and vfclass.v (V 1.0 §13.8-13.14)
at the VFUNARY1 decode site (funct6=0x13, funct3=0x1), previously rejected as
reserved.

vfrsqrt7 and vfrec7 are the table-driven 7-bit estimates. The two
128-entry tables are transcribed from the specification's vfrsqrt7.adoc and
vfrec7.adoc, and kept in the specification's row layout by disabling
clang-format around them, as several existing files already do.

Both instructions were developed against the worked examples in the spec
before being written in C:

• vfrsqrt7(0x00718abc) = 0x5f080000, vfrsqrt7(0x7f765432) = 0x1f820000
• vfrec7(0x00718abc) = 0x7e900000, vfrec7(0x7f765432) = 0x00214000

All four are covered by the smoke test. Two details that cost some time and
may be worth noting for review:

• for a subnormal vfrec7 output the implicit leading one has to be
  restored before the right shift, otherwise the low bit of the result is
  lost (the second worked example catches this);
• the two published vfrsqrt7 examples both have an odd normalized
  exponent, so the floor in (3*B - 1 - e) / 2 is absorbed by integer
  division and an off-by-one there passes both of them. The smoke test adds
  an even-exponent input (4.0 -> 0x3eff0000) which does catch it.

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.

---

Summary by cubic

Implements the last of the five RVV floating-point unary instruction families in #734 by adding vfsqrt.v, vfrsqrt7.v, vfrec7.v, and vfclass.v (V 1.0 §13.8–13.14) at the VFUNARY1 decode site, which previously rejected the encoding as reserved.

• The four instructions are distinguished by the vs1 field (00000, 00100, 00101, 10000); both masked and unmasked forms are permitted.
• The 7-bit estimate instructions use 128-entry lookup tables transcribed from the spec, with clang-format disabled around them to preserve the spec's row layout.
• Smoke tests cover both spec worked examples; an even-exponent vfrsqrt7 input (4.0 → 0x3eff0000) is added because both published examples share an odd normalized exponent and would not catch a floor off-by-one in the output-exponent formula.
• Subnormal vfrec7 output restores the implicit leading one before the right shift, otherwise the low bit is lost (the second worked example exercises this path).
• vfrsqrt7 results are independent of the dynamic rounding mode; only vfrec7 overflow depends on frm, so only vfsqrt and vfrec7 invoke rvv_fp_begin_round.
• All four are non-foldable in constopt, run as V_NOP without EXT_V, and make check passes in default, EXT_V, JIT+EXT_V, and EXT_V-without-EXT_F builds.

^{Written for commit 08bdc4d. Summary will update on new commits.}

[cubic-dev-ai]

3 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="tests/rvv-smoke.S">

<violation number="1" location="tests/rvv-smoke.S:2008">
P3: fail_vfrec7 assigns exit code 0xa0, which collides with the range the block comment above reserves for 'spec-aware mask/vstart tests' (0xa1-0xa6 are all mask/vstart tests). A future regression returning 0xa0 will be misread as a mask/vstart failure. Use a code outside the 0xaN range, e.g. 0x9c, to keep the documented grouping meaningful.</violation>
</file>

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

<violation number="1" location="src/rv32_v_template.c:7039">
P2: When a masked unary instruction uses `vd=v0`, the loop overwrites mask bits while reading later lanes, making execution depend on earlier results. Reject a destination group overlapping `v0` when `vm=0` before entering the loop.</violation>

<violation number="2" location="src/rv32_v_template.c:7686">
P2: `vfrsqrt7.v` and `vfrec7.v` must not update `fflags`, but these helpers raise flags and this shared wrapper commits them after each vector. Separate the estimate instructions from the flag-committing wrapper while retaining `frm` selection for `vfrec7` overflow handling.</violation>
</file>

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

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: When a masked unary instruction uses vd=v0, the loop overwrites mask bits while reading later lanes, making execution depend on earlier results. Reject a destination group overlapping v0 when vm=0 before entering the loop.

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 7039:

<comment>When a masked unary instruction uses `vd=v0`, the loop overwrites mask bits while reading later lanes, making execution depend on earlier results. Reject a destination group overlapping `v0` when `vm=0` before entering the loop.</comment>

<file context>
@@ -6828,6 +7016,36 @@ static inline void rvv_exec_fp32_vf(riscv_t *rv,
+                rvv_set_elem(rv, dest, elem, 32, 0xFFFFFFFFU);
+            continue;
+        }
+        rvv_set_elem(rv, dest, elem, 32,
+                     op(rvv_get_elem(rv, ir->vs2, elem, 32)));
+    }
</file context>

[cubic-dev-ai]

P2: vfrsqrt7.v and vfrec7.v must not update fflags, but these helpers raise flags and this shared wrapper commits them after each vector. Separate the estimate instructions from the flag-committing wrapper while retaining frm selection for vfrec7 overflow handling.

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 7686:

<comment>`vfrsqrt7.v` and `vfrec7.v` must not update `fflags`, but these helpers raise flags and this shared wrapper commits them after each vector. Separate the estimate instructions from the flag-committing wrapper while retaining `frm` selection for `vfrec7` overflow handling.</comment>

<file context>
@@ -7448,6 +7666,27 @@ static inline void rvv_exec_vfmv_v_f(riscv_t *rv,
+    })
+
+RVV_FP32_UNARY_OP(vfsqrt_v, rvv_fp_sqrt32, true);
+RVV_FP32_UNARY_OP(vfrsqrt7_v, rvv_fp_rsqrt7_32, false);
+RVV_FP32_UNARY_OP(vfrec7_v, rvv_fp_rec7_32, true);
+RVV_FP32_UNARY_OP(vfclass_v, rvv_fp_class32, false);
</file context>

[cubic-dev-ai]

P3: fail_vfrec7 assigns exit code 0xa0, which collides with the range the block comment above reserves for 'spec-aware mask/vstart tests' (0xa1-0xa6 are all mask/vstart tests). A future regression returning 0xa0 will be misread as a mask/vstart failure. Use a code outside the 0xaN range, e.g. 0x9c, to keep the documented grouping meaningful.

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 2008:

<comment>fail_vfrec7 assigns exit code 0xa0, which collides with the range the block comment above reserves for 'spec-aware mask/vstart tests' (0xa1-0xa6 are all mask/vstart tests). A future regression returning 0xa0 will be misread as a mask/vstart failure. Use a code outside the 0xaN range, e.g. 0x9c, to keep the documented grouping meaningful.</comment>

<file context>
@@ -1911,6 +1995,18 @@ fail_vsmul4:
+    li a3, 0x9f
+    j write_and_exit_fail
+fail_vfrec7:
+    li a3, 0xa0
+    j write_and_exit_fail
 fail_vwmulu:
</file context>

Suggested change

	li a3, 0xa0
	li a3, 0x9c