[llvm] [RISCV] Lower VP_SELECT constant false to use vmerge.vxm/vmerge.vim (PR #144461)

[Philip Reames via llvm-commits]

================
@@ -470,6 +470,28 @@ define <vscale x 2 x i64> @select_nxv2i64(<vscale x 2 x i1> %a, <vscale x 2 x i6
   ret <vscale x 2 x i64> %v
 }
 
+define <vscale x 2 x i64> @select_nxv2i64_constant_true(<vscale x 2 x i1> %a, <vscale x 2 x i64> %b, i32 zeroext %evl) {
+; CHECK-LABEL: select_nxv2i64_constant_true:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli zero, a0, e64, m2, ta, ma
+; CHECK-NEXT:    vmerge.vim v8, v8, -1, v0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 2 x i64> @llvm.vp.select.nxv2i64(<vscale x 2 x i1> %a, <vscale x 2 x i64> splat (i64 -1), <vscale x 2 x i64> %b, i32 %evl)
+  ret <vscale x 2 x i64> %v
+}
+
+define <vscale x 2 x i64> @select_nxv2i64_constant_false(<vscale x 2 x i1> %a, <vscale x 2 x i64> %b, i32 zeroext %evl) {
+; CHECK-LABEL: select_nxv2i64_constant_false:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli zero, a0, e64, m2, ta, ma
+; CHECK-NEXT:    vmnot.m v0, v0
----------------
preames wrote:

I believe the prior codegen here is going to be li; vmv.v.x; vmerge right?  If so, negating the condition would be profitable in this case.  Though you do have the hoisting interaction noted above.  

https://github.com/llvm/llvm-project/pull/144461