[llvm] [RISCV] Preserve tail agnostic policy in some vector peepholes (PR #105788)

[Luke Lau via llvm-commits]

https://github.com/lukel97 updated https://github.com/llvm/llvm-project/pull/105788

>From 5e1d34a50bf28a60d26579389cf0ffb381368efe Mon Sep 17 00:00:00 2001
From: Luke Lau <luke at igalia.com>
Date: Fri, 23 Aug 2024 14:27:12 +0800
Subject: [PATCH 1/2] Precommit tests

---
 llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir b/llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir
index b2526c6df6939e..aa14914b3ef96c 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir
+++ b/llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir
@@ -18,3 +18,17 @@ body: |
     %y:gpr = ADDI $x0, 1
     %z:vr = PseudoVMV_V_V_M1 %passthru, %x, 4, 5 /* e32 */, 0 /* tu, mu */
 ...
+---
+name: tail_agnostic
+body: |
+  bb.0:
+    liveins: $v8
+    ; CHECK-LABEL: name: tail_agnostic
+    ; CHECK: liveins: $v8
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %passthru:vr = COPY $v8
+    ; CHECK-NEXT: %x:vr = PseudoVADD_VV_M1 %passthru, $noreg, $noreg, 4, 5 /* e32 */, 0 /* tu, mu */
+    %passthru:vr = COPY $v8
+    %x:vr = PseudoVADD_VV_M1 %passthru, $noreg, $noreg, 4, 5 /* e32 */, 0 /* tu, mu */
+    %y:vr = PseudoVMV_V_V_M1 %passthru, %x, 4, 5 /* e32 */, 1 /* ta, mu */
+...

>From 78b228703cd5d2af05533a4495b23553ff8e04f6 Mon Sep 17 00:00:00 2001
From: Luke Lau <luke at igalia.com>
Date: Fri, 23 Aug 2024 14:29:15 +0800
Subject: [PATCH 2/2] [RISCV] Preserve tail agnostic policy in some vector
 peepholes

This patch helps avoid regressions in an upcoming patch by making sure we don't accidentally lose a tail agnostic policy when doing some peepholes. There are two places this happens:

1. When converting a vmerge.vvm with an all ones mask to a vmv.v.v, if vmerge's passthru was undefined the tail will always be undef, even if the new passthru (vmerge's old false operand) isn't undef. We can use ta here.

2. When folding a vmv.v.v into its source, if the vmv.v.v had a ta policy we can copy it over, provided that we kept the same VL.
---
 llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp    | 16 ++++++++++------
 .../RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir     |  2 +-
 llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir |  2 +-
 3 files changed, 12 insertions(+), 8 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp b/llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp
index 34e5d9224f7150..16f63e54be7f17 100644
--- a/llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVectorPeephole.cpp
@@ -365,8 +365,10 @@ bool RISCVVectorPeephole::convertVMergeToVMv(MachineInstr &MI) const {
   MI.removeOperand(1);  // Passthru operand
   MI.tieOperands(0, 1); // Tie false to dest
   MI.removeOperand(3);  // Mask operand
-  MI.addOperand(
-      MachineOperand::CreateImm(RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED));
+  int64_t Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED;
+  if (PassthruReg == RISCV::NoRegister)
+    Policy |= RISCVII::TAIL_AGNOSTIC;
+  MI.addOperand(MachineOperand::CreateImm(Policy));
 
   // vmv.v.v doesn't have a mask operand, so we may be able to inflate the
   // register class for the destination and passthru operands e.g. VRNoV0 -> VR
@@ -520,10 +522,12 @@ bool RISCVVectorPeephole::foldVMV_V_V(MachineInstr &MI) {
                                               *Src->getParent()->getParent()));
   }
 
-  // Use a conservative tu,mu policy, RISCVInsertVSETVLI will relax it if
-  // passthru is undef.
-  Src->getOperand(RISCVII::getVecPolicyOpNum(Src->getDesc()))
-      .setImm(RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED);
+  // If MI had a ta policy and the VL didn't increase, we can preserve it.
+  int64_t Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED;
+  if ((MI.getOperand(5).getImm() & RISCVII::TAIL_AGNOSTIC) &&
+      isVLKnownLE(MI.getOperand(3), SrcVL))
+    Policy |= RISCVII::TAIL_AGNOSTIC;
+  Src->getOperand(RISCVII::getVecPolicyOpNum(Src->getDesc())).setImm(Policy);
 
   MRI->replaceRegWith(MI.getOperand(0).getReg(), Src->getOperand(0).getReg());
   MI.eraseFromParent();
diff --git a/llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir b/llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir
index 01fff3de0aa8bd..9080fcfb0f0420 100644
--- a/llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir
+++ b/llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-to-vmv.mir
@@ -15,7 +15,7 @@ body: |
     ; CHECK-NEXT: %avl:gprnox0 = COPY $x1
     ; CHECK-NEXT: %mask:vmv0 = PseudoVMSET_M_B8 %avl, 5 /* e32 */
     ; CHECK-NEXT: $v0 = COPY %mask
-    ; CHECK-NEXT: %x:vr = PseudoVMV_V_V_M1 %false, %true, %avl, 5 /* e32 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %x:vr = PseudoVMV_V_V_M1 %false, %true, %avl, 5 /* e32 */, 1 /* ta, mu */
     %false:vr = COPY $v8
     %true:vr = COPY $v9
     %avl:gprnox0 = COPY $x1
diff --git a/llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir b/llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir
index aa14914b3ef96c..4294f48a7a2f57 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir
+++ b/llvm/test/CodeGen/RISCV/rvv/vmv.v.v-peephole.mir
@@ -27,7 +27,7 @@ body: |
     ; CHECK: liveins: $v8
     ; CHECK-NEXT: {{  $}}
     ; CHECK-NEXT: %passthru:vr = COPY $v8
-    ; CHECK-NEXT: %x:vr = PseudoVADD_VV_M1 %passthru, $noreg, $noreg, 4, 5 /* e32 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %x:vr = PseudoVADD_VV_M1 %passthru, $noreg, $noreg, 4, 5 /* e32 */, 1 /* ta, mu */
     %passthru:vr = COPY $v8
     %x:vr = PseudoVADD_VV_M1 %passthru, $noreg, $noreg, 4, 5 /* e32 */, 0 /* tu, mu */
     %y:vr = PseudoVMV_V_V_M1 %passthru, %x, 4, 5 /* e32 */, 1 /* ta, mu */