[RISCV] Use policy instead of ForceTailAgnostic for vmsbf/vmsif/vmsof pseudos.

[topperc]

ForceTailAgnostic is currently treated as an override of the policy operand. It doesn't do anything else so we can just use the policy directly.

I'm going to see about removing the only other use of ForceTailAgnostic by adding a policy operand to those pseudos.

[llvmbot]

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

Author: Craig Topper (topperc)

Changes

ForceTailAgnostic is currently treated as an override of the policy operand. It doesn't do anything else so we can just use the policy directly.

I'm going to see about removing the only other use of ForceTailAgnostic by adding a policy operand to those pseudos.

---

Full diff: https://github.com/llvm/llvm-project/pull/127535.diff

1 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td (+2-2)

diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
index 77f41e3c202c7..33c04d1c05613 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
@@ -122,6 +122,7 @@ def DecImm : SDNodeXForm<imm, [{
 
 defvar TAIL_AGNOSTIC = 1;
 defvar TU_MU = 0;
+defvar TA_MU = 1;
 defvar TA_MA = 3;
 
 //===----------------------------------------------------------------------===//
@@ -2029,7 +2030,6 @@ multiclass VPseudoVSFS_M {
       def "_M_" # mti.BX : VPseudoUnaryNoMaskNoPolicy<VR, VR, constraint>,
                            SchedUnary<"WriteVMSFSV", "ReadVMSFSV", mx,
                                       forcePassthruRead=true>;
-      let ForceTailAgnostic = true in
       def "_M_" # mti.BX # "_MASK" : VPseudoUnaryMask<VR, VR, constraint,
                                                       sewop = sew_mask>,
                                      SchedUnary<"WriteVMSFSV", "ReadVMSFSV", mx,
@@ -4019,7 +4019,7 @@ class VPatMaskUnaryMask<string intrinsic_name,
                 (!cast<Instruction>(inst#"_M_"#mti.BX#"_MASK")
                 (mti.Mask VR:$passthru),
                 (mti.Mask VR:$rs2),
-                (mti.Mask VMV0:$vm), GPR:$vl, mti.Log2SEW, TU_MU)>;
+                (mti.Mask VMV0:$vm), GPR:$vl, mti.Log2SEW, TA_MU)>;
 
 class VPatUnaryAnyMask<string intrinsic,
                        string inst,

[lukel97]

LGTM. I don't think there's much point keeping around the TSFlag/optimisation in RISCVInsertVSETVLI if none of the intrinsics can manually set the policy.

[wangpc-pp]

LGTM.
(If we remove ForceTailAgnostic, there will be a bubble in TSFlags, should we compact it?)

[topperc]

LGTM. (If we remove ForceTailAgnostic, there will be a bubble in TSFlags, should we compact it?)

If I'm able to remove ForceTailAgnostic, I will compact the TSFlags.