https://gcc.gnu.org/g:b95850f0c0df9e90f7a5bad61f76627bff5501c9
commit r17-5911-gb95850f0c0df9e90f7a5bad61f76627bff5501c9 Author: Pan Li <[email protected]> Date: Mon Nov 2 10:18:01 2026 +0800 RISC-V: Fix reg overlap for vs1 of the widening float reduction Like the widening integer reduction, the vs1 of the widening float reduction (aka vfwredosum.vs and vfwredusum.vs) is read with EEW = 2 * SEW while the vs2 is read with EEW = SEW. The section 5.2 of the vector spec reserves any encoding in which one vector register supplies source operands with two or more different EEWs. Thus, the vs1 must not fall inside the vs2 register group, but we may have below asm dump when the vs1 comes from the vs2 register group. vfwredosum.vs v4,v4,v5 Take the Wn3 constraint for the operand 4 (aka vs1) of the widening float reduction. See below issue of the riscv-isa-manual for more details. https://github.com/riscv/riscv-isa-manual/issues/2350 gcc/ChangeLog: * config/riscv/vector.md: Take the Wn3 constraint for the operand 4 of the widening float reduction. Signed-off-by: Pan Li <[email protected]> Diff: --- gcc/config/riscv/vector.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/gcc/config/riscv/vector.md b/gcc/config/riscv/vector.md index 8a97f35967f5..7fbee404694a 100644 --- a/gcc/config/riscv/vector.md +++ b/gcc/config/riscv/vector.md @@ -8240,7 +8240,7 @@ (reg:SI FRM_REGNUM)] UNSPEC_VPREDICATE) (unspec:<V_EXT_LMUL1> [ (match_operand:VF_HS 3 "register_operand" " vr, vr") - (match_operand:<V_EXT_LMUL1> 4 "register_operand" " vr, vr") + (match_operand:<V_EXT_LMUL1> 4 "register_operand" " Wn3, Wn3") ] ANY_FWREDUC_SUM) (match_operand:<V_EXT_LMUL1> 2 "vector_merge_operand" " vu, 0")] UNSPEC_REDUC))] "TARGET_VECTOR"
