[PATCH] D152845: [RISCV][InsertVSETVLI] Treat vmv.v.i as-if it were vmv.s.x when VL=1, and inactive lanes are undefined

[Philip Reames via Phabricator via llvm-commits]

reames updated this revision to Diff 531010.
reames edited the summary of this revision.
reames added a comment.

Don't increase LMUL, move to forward pass only (for now), and handle vmv.v.x, and vfmv.v.f.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D152845/new/

https://reviews.llvm.org/D152845

Files:
  llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
  llvm/test/CodeGen/RISCV/rvv/fixed-vectors-reduction-int-vp.ll
  llvm/test/CodeGen/RISCV/rvv/fixed-vectors-reduction-int.ll


Index: llvm/test/CodeGen/RISCV/rvv/fixed-vectors-reduction-int.ll
===================================================================
--- llvm/test/CodeGen/RISCV/rvv/fixed-vectors-reduction-int.ll
+++ llvm/test/CodeGen/RISCV/rvv/fixed-vectors-reduction-int.ll
@@ -2199,9 +2199,7 @@
 ; RV32:       # %bb.0:
 ; RV32-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
 ; RV32-NEXT:    vle64.v v8, (a0)
-; RV32-NEXT:    vsetivli zero, 1, e64, m1, ta, ma
 ; RV32-NEXT:    vmv.v.i v9, -1
-; RV32-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
 ; RV32-NEXT:    vredand.vs v8, v8, v9
 ; RV32-NEXT:    vmv.x.s a0, v8
 ; RV32-NEXT:    li a1, 32
@@ -5686,9 +5684,7 @@
 ; RV32:       # %bb.0:
 ; RV32-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
 ; RV32-NEXT:    vle64.v v8, (a0)
-; RV32-NEXT:    vsetivli zero, 1, e64, m1, ta, ma
 ; RV32-NEXT:    vmv.v.i v9, -1
-; RV32-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
 ; RV32-NEXT:    vredminu.vs v8, v8, v9
 ; RV32-NEXT:    vmv.x.s a0, v8
 ; RV32-NEXT:    li a1, 32
Index: llvm/test/CodeGen/RISCV/rvv/fixed-vectors-reduction-int-vp.ll
===================================================================
--- llvm/test/CodeGen/RISCV/rvv/fixed-vectors-reduction-int-vp.ll
+++ llvm/test/CodeGen/RISCV/rvv/fixed-vectors-reduction-int-vp.ll
@@ -1431,8 +1431,8 @@
 ; RV32-NEXT:    vmv.v.x v9, a1
 ; RV32-NEXT:    vmsne.vi v9, v9, 0
 ; RV32-NEXT:    vmand.mm v0, v9, v0
-; RV32-NEXT:    vsetvli zero, zero, e8, mf8, ta, ma
 ; RV32-NEXT:    vmv.v.i v9, 1
+; RV32-NEXT:    vsetvli zero, zero, e8, mf8, ta, ma
 ; RV32-NEXT:    vmerge.vvm v8, v9, v8, v0
 ; RV32-NEXT:    vmv.x.s a0, v8
 ; RV32-NEXT:    mv a1, a2
@@ -1452,8 +1452,8 @@
 ; RV64-NEXT:    vmv.v.x v9, a1
 ; RV64-NEXT:    vmsne.vi v9, v9, 0
 ; RV64-NEXT:    vmand.mm v0, v9, v0
-; RV64-NEXT:    vsetvli zero, zero, e8, mf8, ta, ma
 ; RV64-NEXT:    vmv.v.i v9, 1
+; RV64-NEXT:    vsetvli zero, zero, e8, mf8, ta, ma
 ; RV64-NEXT:    vmerge.vvm v8, v9, v8, v0
 ; RV64-NEXT:    vmv.x.s a0, v8
 ; RV64-NEXT:    mv a1, a2
Index: llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
===================================================================
--- llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
+++ llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
@@ -85,6 +85,17 @@
   }
 }
 
+static bool isScalarSplatInstr(const MachineInstr &MI) {
+  switch (getRVVMCOpcode(MI.getOpcode())) {
+  default:
+    return false;
+  case RISCV::VMV_V_I:
+  case RISCV::VMV_V_X:
+  case RISCV::VFMV_V_F:
+    return true;
+  }
+}
+
 static bool isVSlideInstr(const MachineInstr &MI) {
   switch (getRVVMCOpcode(MI.getOpcode())) {
   default:
@@ -911,6 +922,20 @@
     Used.TailPolicy = false;
   }
 
+  // A tail undefined vmv.v.i/x or vfmv.v.f with VL=1 can be treated in the same
+  // semantically as vmv.s.x.  This is particularly useful since we don't have an
+  // immediate form of vmv.s.x, and thus frequently use vmv.v.i in it's place.
+  // Since a splat is non-constant time in LMUL, we do need to be careful to not
+  // increase the number of active vector registers (unlikely for vmv.s.x.)
+  if (isScalarSplatInstr(MI) && Require.hasAVLImm() && Require.getAVLImm() == 1 &&
+      isLMUL1OrSmaller(CurInfo.getVLMUL()) && hasUndefinedMergeOp(MI, *MRI)) {
+    Used.LMUL = false;
+    Used.SEWLMULRatio = false;
+    Used.VLAny = false;
+    Used.SEW = DemandedFields::SEWGreaterThanOrEqual;
+    Used.TailPolicy = false;
+  }
+
   if (CurInfo.isCompatible(Used, Require, *MRI))
     return false;
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D152845.531010.patch
Type: text/x-patch
Size: 3450 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230613/cc54df7e/attachment.bin>