[llvm] aefcd59 - [RISCV] Teach RISCVInsertVSETVLI::needVSETVLI to handle mask register instructions better.

[Craig Topper via llvm-commits]

Author: Craig Topper
Date: 2021-10-29T09:49:36-07:00
New Revision: aefcd598959abbea5eddf1fe454359f53bd9d853

URL: https://github.com/llvm/llvm-project/commit/aefcd598959abbea5eddf1fe454359f53bd9d853
DIFF: https://github.com/llvm/llvm-project/commit/aefcd598959abbea5eddf1fe454359f53bd9d853.diff

LOG: [RISCV] Teach RISCVInsertVSETVLI::needVSETVLI to handle mask register instructions better.

If the VL operand of a mask register instruction comes from an
explicit vsetvli with a different VTYPE, we can still avoid needing
a vsetvli as long as the SEW/LMUL ratio is the same and policy bits
match.

Differential Revision: https://reviews.llvm.org/D112762

Added: 
    

Modified: 
    llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
    llvm/test/CodeGen/RISCV/rvv/vsetvli-insert.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp b/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
index b4b3bbaf53a9f..69d9ab4c3f01a 100644
--- a/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
@@ -178,6 +178,26 @@ class VSETVLIInfo {
     return getSEWLMULRatio() == Other.getSEWLMULRatio();
   }
 
+  bool hasCompatibleVTYPE(const VSETVLIInfo &InstrInfo, bool Strict) const {
+    // Simple case, see if full VTYPE matches.
+    if (hasSameVTYPE(InstrInfo))
+      return true;
+
+    if (Strict)
+      return false;
+
+    // If this is a mask reg operation, it only cares about VLMAX.
+    // FIXME: Mask reg operations are probably ok if "this" VLMAX is larger
+    // than "InstrInfo".
+    // FIXME: The policy bits can probably be ignored for mask reg operations.
+    if (InstrInfo.MaskRegOp && hasSameVLMAX(InstrInfo) &&
+        TailAgnostic == InstrInfo.TailAgnostic &&
+        MaskAgnostic == InstrInfo.MaskAgnostic)
+      return true;
+
+    return false;
+  }
+
   // Determine whether the vector instructions requirements represented by
   // InstrInfo are compatible with the previous vsetvli instruction represented
   // by this.
@@ -206,23 +226,15 @@ class VSETVLIInfo {
     if (!hasSameAVL(InstrInfo))
       return false;
 
-    // Simple case, see if full VTYPE matches.
-    if (hasSameVTYPE(InstrInfo))
+    if (hasCompatibleVTYPE(InstrInfo, Strict))
       return true;
 
     // Strict matches must ensure a full VTYPE match.
     if (Strict)
       return false;
 
-    // If this is a mask reg operation, it only cares about VLMAX.
-    // FIXME: Mask reg operations are probably ok if "this" VLMAX is larger
-    // than "InstrInfo".
-    if (InstrInfo.MaskRegOp && hasSameVLMAX(InstrInfo) &&
-        TailAgnostic == InstrInfo.TailAgnostic &&
-        MaskAgnostic == InstrInfo.MaskAgnostic)
-      return true;
-
     // Store instructions don't use the policy fields.
+    // TODO: Move into hasCompatibleVTYPE?
     if (InstrInfo.StoreOp && VLMul == InstrInfo.VLMul && SEW == InstrInfo.SEW)
       return true;
 
@@ -564,7 +576,7 @@ bool RISCVInsertVSETVLI::needVSETVLI(const VSETVLIInfo &Require,
   // VSETVLI here.
   if (!CurInfo.isUnknown() && Require.hasAVLReg() &&
       Require.getAVLReg().isVirtual() && !CurInfo.hasSEWLMULRatioOnly() &&
-      Require.hasSameVTYPE(CurInfo)) {
+      CurInfo.hasCompatibleVTYPE(Require, /*Strict*/ false)) {
     if (MachineInstr *DefMI = MRI->getVRegDef(Require.getAVLReg())) {
       if (DefMI->getOpcode() == RISCV::PseudoVSETVLI ||
           DefMI->getOpcode() == RISCV::PseudoVSETVLIX0 ||

diff  --git a/llvm/test/CodeGen/RISCV/rvv/vsetvli-insert.ll b/llvm/test/CodeGen/RISCV/rvv/vsetvli-insert.ll
index b99a2cad2d74e..585a819ea04a4 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vsetvli-insert.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vsetvli-insert.ll
@@ -77,13 +77,12 @@ entry:
   ret <vscale x 1 x i64> %1
 }
 
-; FIXME the second vsetvli is unnecessary.
+; Make sure we don't insert a vsetvli for the vmand instruction.
 define <vscale x 1 x i1> @test5(<vscale x 1 x i64> %0, <vscale x 1 x i64> %1, <vscale x 1 x i1> %2, i64 %avl) nounwind {
 ; CHECK-LABEL: test5:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    vsetvli a0, a0, e64, m1, ta, mu
 ; CHECK-NEXT:    vmseq.vv v8, v8, v9
-; CHECK-NEXT:    vsetvli zero, a0, e8, mf8, ta, mu
 ; CHECK-NEXT:    vmand.mm v0, v8, v0
 ; CHECK-NEXT:    ret
 entry: