From 8d4998b478028b08c6bc50a1b49573517a6e973c Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Thu, 9 Mar 2023 12:32:11 +0000
Subject: [PATCH] [X86] combineConcatVectorOps - add basic TRUNCATE handling on
 AVX512 targets

This really should be in shuffle combining, but we're still struggling to handle mismatched vector sizes
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 16 ++++++++++-
 llvm/test/CodeGen/X86/vector-trunc.ll   | 47 +++++++--------------------------
 2 files changed, 25 insertions(+), 38 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 5280dfa..a8c37db 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -55893,7 +55893,7 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
   if (llvm::all_of(Ops, [Op0](SDValue Op) {
         return Op.getOpcode() == Op0.getOpcode();
       })) {
-    auto ConcatSubOperand = [&](MVT VT, ArrayRef<SDValue> SubOps, unsigned I) {
+    auto ConcatSubOperand = [&](EVT VT, ArrayRef<SDValue> SubOps, unsigned I) {
       SmallVector<SDValue> Subs;
       for (SDValue SubOp : SubOps)
         Subs.push_back(SubOp.getOperand(I));
@@ -56048,6 +56048,20 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
         }
       }
       break;
+    case ISD::TRUNCATE:
+      if (!IsSplat && NumOps == 2 && VT.is256BitVector()) {
+        EVT SrcVT = Ops[0].getOperand(0).getValueType();
+        if (SrcVT.is256BitVector() && SrcVT.isSimple() &&
+            SrcVT == Ops[1].getOperand(0).getValueType() &&
+            Subtarget.useAVX512Regs() &&
+            Subtarget.getPreferVectorWidth() >= 512 &&
+            (SrcVT.getScalarSizeInBits() > 16 || Subtarget.useBWIRegs())) {
+          EVT NewSrcVT = SrcVT.getDoubleNumVectorElementsVT(*DAG.getContext());
+          return DAG.getNode(ISD::TRUNCATE, DL, VT,
+                             ConcatSubOperand(NewSrcVT, Ops, 0));
+        }
+      }
+      break;
     case X86ISD::VSHLI:
     case X86ISD::VSRLI:
       // Special case: SHL/SRL AVX1 V4i64 by 32-bits can lower as a shuffle.
diff --git a/llvm/test/CodeGen/X86/vector-trunc.ll b/llvm/test/CodeGen/X86/vector-trunc.ll
index b3147c1..4b73e5a 100644
--- a/llvm/test/CodeGen/X86/vector-trunc.ll
+++ b/llvm/test/CodeGen/X86/vector-trunc.ll
@@ -1392,37 +1392,12 @@ define <8 x i32> @trunc2x4i64_8i32(<4 x i64> %a, <4 x i64> %b) {
 ; AVX2-FAST-PERLANE-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,2],ymm2[0,2],ymm0[4,6],ymm2[4,6]
 ; AVX2-FAST-PERLANE-NEXT:    retq
 ;
-; AVX512F-LABEL: trunc2x4i64_8i32:
-; AVX512F:       # %bb.0: # %entry
-; AVX512F-NEXT:    # kill: def $ymm1 killed $ymm1 def $zmm1
-; AVX512F-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; AVX512F-NEXT:    vpmovqd %zmm0, %ymm0
-; AVX512F-NEXT:    vpmovqd %zmm1, %ymm1
-; AVX512F-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512F-NEXT:    retq
-;
-; AVX512VL-LABEL: trunc2x4i64_8i32:
-; AVX512VL:       # %bb.0: # %entry
-; AVX512VL-NEXT:    vpmovqd %ymm0, %xmm0
-; AVX512VL-NEXT:    vpmovqd %ymm1, %xmm1
-; AVX512VL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512VL-NEXT:    retq
-;
-; AVX512BW-LABEL: trunc2x4i64_8i32:
-; AVX512BW:       # %bb.0: # %entry
-; AVX512BW-NEXT:    # kill: def $ymm1 killed $ymm1 def $zmm1
-; AVX512BW-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; AVX512BW-NEXT:    vpmovqd %zmm0, %ymm0
-; AVX512BW-NEXT:    vpmovqd %zmm1, %ymm1
-; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512BW-NEXT:    retq
-;
-; AVX512BWVL-LABEL: trunc2x4i64_8i32:
-; AVX512BWVL:       # %bb.0: # %entry
-; AVX512BWVL-NEXT:    vpmovqd %ymm0, %xmm0
-; AVX512BWVL-NEXT:    vpmovqd %ymm1, %xmm1
-; AVX512BWVL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512BWVL-NEXT:    retq
+; AVX512-LABEL: trunc2x4i64_8i32:
+; AVX512:       # %bb.0: # %entry
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
+; AVX512-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vpmovqd %zmm0, %ymm0
+; AVX512-NEXT:    retq
 entry:
   %0 = trunc <4 x i64> %a to <4 x i32>
   %1 = trunc <4 x i64> %b to <4 x i32>
@@ -1797,18 +1772,16 @@ define <32 x i8> @trunc2x16i16_32i8(<16 x i16> %a, <16 x i16> %b) {
 ;
 ; AVX512BW-LABEL: trunc2x16i16_32i8:
 ; AVX512BW:       # %bb.0: # %entry
-; AVX512BW-NEXT:    # kill: def $ymm1 killed $ymm1 def $zmm1
 ; AVX512BW-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vpmovwb %zmm0, %ymm0
-; AVX512BW-NEXT:    vpmovwb %zmm1, %ymm1
-; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512BWVL-LABEL: trunc2x16i16_32i8:
 ; AVX512BWVL:       # %bb.0: # %entry
-; AVX512BWVL-NEXT:    vpmovwb %ymm0, %xmm0
-; AVX512BWVL-NEXT:    vpmovwb %ymm1, %xmm1
-; AVX512BWVL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
+; AVX512BWVL-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
+; AVX512BWVL-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512BWVL-NEXT:    vpmovwb %zmm0, %ymm0
 ; AVX512BWVL-NEXT:    retq
 entry:
   %0 = trunc <16 x i16> %a to <16 x i8>
-- 
2.7.4