Implement RVV integer extension instructions

[alanhc]

Implements the integer-extension family from #734.

Adds vzext.vf{2,4,8} and vsext.vf{2,4,8} (V 1.0 §11.3) at the VXUNARY0
decode site, which was previously rejected with a FIXME.

The one non-obvious part is the register-overlap rule. vzext/vsext have
a destination EEW greater than the source EEW, and §5.2 requires the source
EMUL to be at least 1 and the overlap to sit in the highest-numbered
part of the destination group. The spec's own example is that at LMUL=8,
vzext.vf4 v0, v6 is legal while a source of v0, v2 or v4 is not.

That is not what rvv_cross_eew_overlap_illegal() implements — it requires
both groups to share a base register, which matches the narrowing direction
instead and gets these two cases exactly backwards. The check is therefore
spelled out in the execution helper with a comment explaining why the shared
predicate is not used.

Source EMUL is compared against 1 via rvv_lmul_ratio() rather than the
register span, because rvv_eew_reg_span() clamps a fractional EMUL to a
span of one register and the information is lost.

Encodings were checked against riscv-opcodes (funct6=0x12, funct3=0x2,
vs1 = 2..7); reserved vs1 values are rejected. Smoke coverage added to
tests/rvv-smoke.S for vf2/vf4 zero- and sign-extension plus a masked case.
make check passes with EXT_V enabled, and 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 integer-extension instructions vzext.vf{2,4,8} and vsext.vf{2,4,8} (V 1.0 §11.3), which were previously rejected as unimplemented. They are now decoded and executed, including masked forms and tail handling; the widening register-overlap rule follows §5.2 rather than the existing rvv_cross_eew_overlap_illegal() predicate.

^{Written for commit 15453a8. 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:6536">
P1: When a masked extension uses `vd = v0`, the loop overwrites the mask register while reading it for later elements, producing incorrect results. Reject destination groups overlapping `v0` when `ir->vm == 0`.</violation>
</file>

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

<violation number="1" location="tests/rvv-smoke.S:1830">
P3: The new smoke cases never exercise the extension overlap rule: every source/destination pair is disjoint. An incorrect overlap check would pass these tests, so add a legal highest-part overlap case and verify reserved overlaps where possible.</violation>
</file>

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

Re-trigger cubic</sub>

[cubic-dev-ai]

P1: When a masked extension uses vd = v0, the loop overwrites the mask register while reading it for later elements, producing incorrect results. Reject destination groups overlapping v0 when ir->vm == 0.

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 6536:

<comment>When a masked extension uses `vd = v0`, the loop overwrites the mask register while reading it for later elements, producing incorrect results. Reject destination groups overlapping `v0` when `ir->vm == 0`.</comment>

<file context>
@@ -6499,6 +6499,93 @@ RVOP(vid_v, {
+     */
+    if (src_bits < 8)
+        return rvv_trap_illegal_state(rv, 0);
+    if (!rvv_eew_reg_span(rv, src_bits, &src_span) ||
+        !rvv_validate_data_reg(rv->csr_vtype, ir->vd) ||
+        !rvv_validate_eew_reg(rv, src_bits, ir->vs2))
</file context>

Suggested change

	if (!rvv_eew_reg_span(rv, src_bits, &src_span) ||
	!rvv_validate_data_reg(rv->csr_vtype, ir->vd) ||
	!rvv_validate_eew_reg(rv, src_bits, ir->vs2))
	if (!rvv_eew_reg_span(rv, src_bits, &src_span) ||
	!rvv_validate_data_reg(rv->csr_vtype, ir->vd) ||
	!rvv_validate_eew_reg(rv, src_bits, ir->vs2) ||
	(!ir->vm && rvv_reg_spans_overlap(ir->vd, dest_span, 0, 1)))

[cubic-dev-ai]

P3: The new smoke cases never exercise the extension overlap rule: every source/destination pair is disjoint. An incorrect overlap check would pass these tests, so add a legal highest-part overlap case and verify reserved overlaps where possible.

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 1830:

<comment>The new smoke cases never exercise the extension overlap rule: every source/destination pair is disjoint. An incorrect overlap check would pass these tests, so add a legal highest-part overlap case and verify reserved overlaps where possible.</comment>

<file context>
@@ -1813,6 +1819,82 @@ _start:
+    vle8.v v20, (a1)
+
+    # vzext.vf2: 0x7f,0x80,0x01,0xff -> zero-extended to 16-bit
+    vzext.vf2 v22, v20
+    la a1, ext_out
+    vse16.v v22, (a1)
</file context>