[llvm] [RISCV] Allow non-canonicalized splats in isProfitableToSinkOperands (PR #161586)

[Philip Reames via llvm-commits]

https://github.com/preames created https://github.com/llvm/llvm-project/pull/161586

This isn't an optimization change - IR transforms should have remove the operands and replaced them with poison.  However, I noticed the non-canonical splat structure in a couple of llvm-reduce outputs. This results in us creating extremely atypical IR which is quite misleading about the true cause of what's going on. (Because the non-canonical splat doesn't get sunk, we then prone whatever was actually holding it outside the loop in the original example, eliminating insight as to the true cause of whatever issue we're debugging.)

>From a46ac7ed1dbfc98441deeed934f02f301b6d2976 Mon Sep 17 00:00:00 2001
From: Philip Reames <preames at rivosinc.com>
Date: Wed, 1 Oct 2025 13:46:26 -0700
Subject: [PATCH] [RISCV] Allow non-canonicalized splats in
 isProfitableToSinkOperands

This isn't an optimization change - IR transforms should have remove
the operands and replaced them with poison.  However, I noticed the
non-canonical splat structure in a couple of llvm-reduce outputs.
This results in us creating extremely atypical IR which is quite
misleading about the true cause of what's going on. (Because the
non-canoncal splat doesn't get sunk, we then prone whatever was
actually holding it outside the loop in the original example,
eliminating insight as to the true cause of whatever issue we're
debugging.)
---
 .../Target/RISCV/RISCVTargetTransformInfo.cpp |  4 +--
 .../CodeGen/RISCV/rvv/sink-splat-operands.ll  | 36 +++++++++++++++++++
 2 files changed, 38 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
index d4124ae9aeff0..ee25f6918de8b 100644
--- a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
@@ -3139,8 +3139,8 @@ bool RISCVTTIImpl::isProfitableToSinkOperands(
     bool IsVPSplat = match(Op, m_Intrinsic<Intrinsic::experimental_vp_splat>(
                                    m_Value(), m_Value(), m_Value()));
     if (!IsVPSplat &&
-        !match(Op, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()),
-                             m_Undef(), m_ZeroMask())))
+        !match(Op, m_Shuffle(m_InsertElt(m_Value(), m_Value(), m_ZeroInt()),
+                             m_Value(), m_ZeroMask())))
       continue;
 
     // Don't sink i1 splats.
diff --git a/llvm/test/CodeGen/RISCV/rvv/sink-splat-operands.ll b/llvm/test/CodeGen/RISCV/rvv/sink-splat-operands.ll
index 02825b2bda484..19a184148c0b6 100644
--- a/llvm/test/CodeGen/RISCV/rvv/sink-splat-operands.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/sink-splat-operands.ll
@@ -6018,3 +6018,39 @@ vector.latch:                                     ; preds = %for.body419
 for.cond.cleanup:                                 ; preds = %vector.latch
   ret void
 }
+
+;; This is exactly like sink_add_splat except that the splat has operands
+;; which haven't been converted to undef.
+define void @sink_non_canonical_splat(ptr nocapture %a, i32 signext %x) {
+; CHECK-LABEL: sink_non_canonical_splat:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    lui a2, 1
+; CHECK-NEXT:    add a2, a0, a2
+; CHECK-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
+; CHECK-NEXT:  .LBB131_1: # %vector.body
+; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    vle32.v v8, (a0)
+; CHECK-NEXT:    vadd.vx v8, v8, a1
+; CHECK-NEXT:    vse32.v v8, (a0)
+; CHECK-NEXT:    addi a0, a0, 16
+; CHECK-NEXT:    bne a0, a2, .LBB131_1
+; CHECK-NEXT:  # %bb.2: # %for.cond.cleanup
+; CHECK-NEXT:    ret
+entry:
+  %broadcast.splatinsert = insertelement <4 x i32> zeroinitializer, i32 %x, i32 0
+  %broadcast.splat = shufflevector <4 x i32> %broadcast.splatinsert, <4 x i32> zeroinitializer, <4 x i32> zeroinitializer
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %entry
+  %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds i32, ptr %a, i64 %index
+  %wide.load = load <4 x i32>, ptr %0, align 4
+  %1 = add <4 x i32> %wide.load, %broadcast.splat
+  store <4 x i32> %1, ptr %0, align 4
+  %index.next = add nuw i64 %index, 4
+  %2 = icmp eq i64 %index.next, 1024
+  br i1 %2, label %for.cond.cleanup, label %vector.body
+
+for.cond.cleanup:                                 ; preds = %vector.body
+  ret void
+}