[llvm] [RISCV] Scale i64 insert/extract element cost on RV32 with LMUL (PR #214609)

Ryan Buchner via llvm-commits llvm-commits at lists.llvm.org
Thu Aug 6 23:02:47 PDT 2026


================
@@ -2702,16 +2702,23 @@ InstructionCost RISCVTTIImpl::getVectorInstrCost(
     // vmv.x.s a1, v8
 
     // For insertelement, we need the following instructions:
-    // vsetivli zero, 2, e32, m4, ta, mu (not count)
-    // vmv.v.i v12, 0
-    // vslide1up.vx v16, v12, a1
-    // vslide1up.vx v12, v16, a0
+    // vsetivli zero, 2, e32, m4, ta, ma (don't count)
+    // vslide1down.vx v12, v8, a0
+    // vslide1down.vx v12, v12, a1
     // addi a0, a2, 1
-    // vsetvli zero, a0, e64, m4, tu, mu (not count)
+    // vsetvli zero, a0, e64, m4, tu, ma (don't count)
     // vslideup.vx v8, v12, a2
 
     // TODO: should we count these special vsetvlis?
-    BaseCost = Opcode == Instruction::InsertElement ? 3 : 4;
+    BaseCost =
+        Opcode == Instruction::InsertElement
+            ? getRISCVInstructionCost({RISCV::VSLIDE1DOWN_VX,
+                                       RISCV::VSLIDE1DOWN_VX,
+                                       RISCV::VSLIDEUP_VX},
+                                      LT.second, CostKind)
+            : getRISCVInstructionCost({RISCV::VSLIDEDOWN_VX, RISCV::VMV_X_S,
+                                       RISCV::VSRL_VX, RISCV::VMV_S_X},
----------------
bababuck wrote:

```suggestion
            : getRISCVInstructionCost({RISCV::VSLIDEDOWN_VX, RISCV::VMV_X_S,
                                       RISCV::VSRL_VX, RISCV::VMV_X_S},
```

https://github.com/llvm/llvm-project/pull/214609


More information about the llvm-commits mailing list