[RISCV] Remove policy and merge operand from unmasked vmsbf/vmsif/vmsof.m.

[topperc]

These instructions always update the destination under a tail agnostic policy.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Craig Topper (topperc)

Changes

These instructions always update the destination under a tail agnostic policy.

---

Full diff: https://github.com/llvm/llvm-project/pull/94530.diff

1 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td (+18-3)

diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
index 5fe5a7a5bd5cc..5db7cb4dc6180 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
@@ -1033,6 +1033,22 @@ class VPseudoUnaryNoMask<DAGOperand RetClass,
   let HasVecPolicyOp = 1;
 }
 
+class VPseudoUnaryNoMaskNoPolicy<DAGOperand RetClass,
+                                 DAGOperand OpClass,
+                                 string Constraint = "",
+                                 int TargetConstraintType = 1> :
+      Pseudo<(outs RetClass:$rd),
+             (ins OpClass:$rs2, AVL:$vl, ixlenimm:$sew), []>,
+      RISCVVPseudo {
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+  let Constraints = Constraint;
+  let TargetOverlapConstraintType = TargetConstraintType;
+  let HasVLOp = 1;
+  let HasSEWOp = 1;
+}
+
 class VPseudoUnaryNoMaskRoundingMode<DAGOperand RetClass,
                                      DAGOperand OpClass,
                                      string Constraint = "",
@@ -2056,7 +2072,7 @@ multiclass VPseudoVSFS_M {
   foreach mti = AllMasks in {
     defvar mx = mti.LMul.MX;
     let VLMul = mti.LMul.value in {
-      def "_M_" # mti.BX : VPseudoUnaryNoMask<VR, VR, constraint>,
+      def "_M_" # mti.BX : VPseudoUnaryNoMaskNoPolicy<VR, VR, constraint>,
                            SchedUnary<"WriteVMSFSV", "ReadVMSFSV", mx,
                                       forceMergeOpRead=true>;
       def "_M_" # mti.BX # "_MASK" : VPseudoUnaryMask<VR, VR, constraint>,
@@ -4078,9 +4094,8 @@ class VPatMaskUnaryNoMask<string intrinsic_name,
                 (mti.Mask VR:$rs2),
                 VLOpFrag)),
                 (!cast<Instruction>(inst#"_M_"#mti.BX)
-                (mti.Mask (IMPLICIT_DEF)),
                 (mti.Mask VR:$rs2),
-                GPR:$vl, mti.Log2SEW, TA_MA)>;
+                GPR:$vl, mti.Log2SEW)>;
 
 class VPatMaskUnaryMask<string intrinsic_name,
                         string inst,

[preames]

Why bother? Given we need the masked variants which do have a mask policy, what do we gain by having distinct pseudo shapes between unmasked and masked? Is the goal to exploit the don't care on the tail policy? If so, then why not do that directly so that the masked variants benefit as well?

[preames]

Why bother? Given we need the masked variants which do have a mask policy, what do we gain by having distinct pseudo shapes between unmasked and masked? Is the goal to exploit the don't care on the tail policy? If so, then why not do that directly so that the masked variants benefit as well?

On reflection, I realized we already have different shapes for these due to the mask operand. Ignore this please.

[preames]

LGTM

[topperc]

Why bother? Given we need the masked variants which do have a mask policy, what do we gain by having distinct pseudo shapes between unmasked and masked? Is the goal to exploit the don't care on the tail policy? If so, then why not do that directly so that the masked variants benefit as well?

Goal was just to remove dimensions that should never be used. After the issue I found with reductions last week, I started auditing uses of TAIL_AGNOSTIC and TA_MA.