[llvm] [RISCV][Peephole] Clear kill flags for registers after fold VMV (PR #138847)

[Piyou Chen via llvm-commits]

================
@@ -0,0 +1,136 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 3
+# RUN: llc %s -o - -mtriple=riscv64 -mattr=+v -run-pass=riscv-vector-peephole \
+# RUN:  -verify-machineinstrs | FileCheck %s
+
+--- |
+  source_filename = "reduced.ll"
+  target datalayout = "e-m:e-p:64:64-i64:64-i128:128-n32:64-S128"
+  target triple = "riscv64-unknown-linux-gnu"
+
+  define i32 @main() #0 {
+  entry:
+    %0 = tail call <vscale x 4 x i16> @llvm.riscv.vmv.v.v.nxv4i16.i64(<vscale x 4 x i16> poison, <vscale x 4 x i16> zeroinitializer, i64 0)
+    %1 = tail call target("riscv.vector.tuple", <vscale x 8 x i8>, 7) @llvm.riscv.tuple.insert.triscv.vector.tuple_nxv8i8_7t.nxv4i16(target("riscv.vector.tuple", <vscale x 8 x i8>, 7) zeroinitializer, <vscale x 4 x i16> %0, i32 0)
+    %2 = tail call target("riscv.vector.tuple", <vscale x 8 x i8>, 7) @llvm.riscv.tuple.insert.triscv.vector.tuple_nxv8i8_7t.nxv4i16(target("riscv.vector.tuple", <vscale x 8 x i8>, 7) %1, <vscale x 4 x i16> zeroinitializer, i32 0)
+    call void @llvm.riscv.vsseg7.triscv.vector.tuple_nxv8i8_7t.i64(target("riscv.vector.tuple", <vscale x 8 x i8>, 7) %2, ptr null, i64 0, i64 4)
+    ret i32 0
+  }
+
+  declare <vscale x 4 x i16> @llvm.riscv.vmv.v.v.nxv4i16.i64(<vscale x 4 x i16>, <vscale x 4 x i16>, i64) #1
+
+  declare target("riscv.vector.tuple", <vscale x 8 x i8>, 7) @llvm.riscv.tuple.insert.triscv.vector.tuple_nxv8i8_7t.nxv4i16(target("riscv.vector.tuple", <vscale x 8 x i8>, 7), <vscale x 4 x i16>, i32 immarg) #2
+
+  declare void @llvm.riscv.vsseg7.triscv.vector.tuple_nxv8i8_7t.i64(target("riscv.vector.tuple", <vscale x 8 x i8>, 7), ptr captures(none), i64, i64 immarg) #3
+
+  attributes #0 = { "target-features"="+v" }
+  attributes #1 = { nocallback nofree nosync nounwind willreturn memory(none) "target-features"="+v" }
+  attributes #2 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) "target-features"="+v" }
+  attributes #3 = { nocallback nofree nosync nounwind willreturn memory(argmem: write) "target-features"="+v" }
+
+...
+---
+name:            main
+alignment:       4
+exposesReturnsTwice: false
+legalized:       false
+regBankSelected: false
+selected:        false
+failedISel:      false
+tracksRegLiveness: true
----------------
BeMg wrote:

`foldVMV_V_V` may not encounter the same situation since it only processes the hasOneUse case and replaces it with the def operand. If it is defined and killed, then it should never have any users to begin with.

I drop the `clearKillFlags` in `foldVMV_V_V`.

https://github.com/llvm/llvm-project/pull/138847