[llvm] [X86] combineVTRUNCSAT - don't split 128-bit concatenated vectors when folding to PACKSS/US (PR #194347)

[Simon Pilgrim via llvm-commits]

https://github.com/RKSimon updated https://github.com/llvm/llvm-project/pull/194347

>From b48d5a6e78e435862396022dcd9ee2b126fa19a1 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Mon, 27 Apr 2026 12:31:41 +0100
Subject: [PATCH] [X86] combineVTRUNCSAT - don't split 128-bit concatenated
 vectors when folding to PACKSS/US

If the VTRUNCS/US node has 128-bit src/dst types, then ensure we don't split into sub-128-bit vectors - just treat it as padded with zeros (matches VTRUNC behaviour)

Fixes #194344
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 17 ++++++++++++++---
 llvm/test/CodeGen/X86/packus.ll         | 19 +++++++++++++++++++
 2 files changed, 33 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index cbae4a97527a8..22685a28d31b9 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -55624,7 +55624,8 @@ static SDValue combineVTRUNC(SDNode *N, SelectionDAG &DAG,
 }
 
 static SDValue combineVTRUNCSAT(SDNode *N, SelectionDAG &DAG,
-                                TargetLowering::DAGCombinerInfo &DCI) {
+                                TargetLowering::DAGCombinerInfo &DCI,
+                                const X86Subtarget &Subtarget) {
   using namespace SDPatternMatch;
   unsigned Opc = N->getOpcode();
   EVT VT = N->getValueType(0);
@@ -55643,7 +55644,17 @@ static SDValue combineVTRUNCSAT(SDNode *N, SelectionDAG &DAG,
       (EltSizeInBits * 2) == Src.getScalarValueSizeInBits() &&
       isFreeToSplitVector(Src, DAG)) {
     SDLoc DL(N);
-    auto [LHS, RHS] = splitVector(Src, DAG, DL);
+    SDValue LHS, RHS;
+    if (Src.getValueSizeInBits() == VT.getSizeInBits()) {
+      assert(VT.is128BitVector() && "128-bit VTRUNC source expected");
+      LHS = Src;
+      RHS = getZeroVector(Src.getSimpleValueType(), Subtarget, DAG, DL);
+    } else {
+      std::tie(LHS, RHS) = splitVector(Src, DAG, DL);
+    }
+    assert(LHS.getValueSizeInBits() == VT.getSizeInBits() &&
+           RHS.getValueSizeInBits() == VT.getSizeInBits() &&
+           "PACK src/dst size mismatch");
     unsigned PackOpc = Opc == X86ISD::VTRUNCS ? X86ISD::PACKSS : X86ISD::PACKUS;
     SDValue Pack = DAG.getNode(PackOpc, DL, VT, LHS, RHS);
     if (VT.is128BitVector())
@@ -62367,7 +62378,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::TRUNCATE:       return combineTruncate(N, DAG, Subtarget);
   case X86ISD::VTRUNC:      return combineVTRUNC(N, DAG, DCI);
   case X86ISD::VTRUNCS:
-  case X86ISD::VTRUNCUS:    return combineVTRUNCSAT(N, DAG, DCI);
+  case X86ISD::VTRUNCUS:    return combineVTRUNCSAT(N, DAG, DCI, Subtarget);
   case X86ISD::ANDNP:       return combineAndnp(N, DAG, DCI, Subtarget);
   case X86ISD::FAND:        return combineFAnd(N, DAG, Subtarget);
   case X86ISD::FANDN:       return combineFAndn(N, DAG, Subtarget);
diff --git a/llvm/test/CodeGen/X86/packus.ll b/llvm/test/CodeGen/X86/packus.ll
index e678991fcec78..7de7caff18d6b 100644
--- a/llvm/test/CodeGen/X86/packus.ll
+++ b/llvm/test/CodeGen/X86/packus.ll
@@ -511,6 +511,25 @@ define <32 x i8> @packuswb_icmp_zero_trunc_256(<16 x i16> %a0) {
   ret <32 x i8> %4
 }
 
+define <8 x i8> @_mm_packs_pu16_manual(<4 x i16> %a, <4 x i16> %b) nounwind {
+; SSE-LABEL: _mm_packs_pu16_manual:
+; SSE:       # %bb.0:
+; SSE-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE-NEXT:    packuswb %xmm0, %xmm0
+; SSE-NEXT:    ret{{[l|q]}}
+;
+; AVX-LABEL: _mm_packs_pu16_manual:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    ret{{[l|q]}}
+  %sh  = shufflevector <4 x i16> %a, <4 x i16> %b, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %minv = tail call <8 x i16> @llvm.smax.v8i16(<8 x i16> %sh, <8 x i16> splat (i16 0))
+  %sat = tail call <8 x i16> @llvm.umin.v8i16(<8 x i16> %minv, <8 x i16> splat (i16 255))
+  %tr  = trunc nuw <8 x i16> %sat to <8 x i8>
+  ret <8 x i8> %tr
+}
+
 define <16 x i8> @_mm_packus_epi16_manual(<8 x i16> %a, <8 x i16> %b) nounwind {
 ; SSE-LABEL: _mm_packus_epi16_manual:
 ; SSE:       # %bb.0: