[llvm-branch-commits] [llvm] release/18.x: [AArch64][GISel] Don't pointlessly lower G_TRUNC (#81479) (PR #81581)

[via llvm-branch-commits]

https://github.com/llvmbot created https://github.com/llvm/llvm-project/pull/81581

Backport 070848c17c2944afa494d42d3ad42929f3379842

Requested by: @nikic

>From c9a5bed568b88196dd3d2444b827a42d29f72064 Mon Sep 17 00:00:00 2001
From: Nikita Popov <npopov at redhat.com>
Date: Tue, 13 Feb 2024 09:29:56 +0100
Subject: [PATCH] [AArch64][GISel] Don't pointlessly lower G_TRUNC (#81479)

If we have something like G_TRUNC from v2s32 to v2s16, then lowering
this to a concat of two G_TRUNC s32 to s16 followed by G_TRUNC from
v2s16 to v2s8 does not bring us any closer to legality. In fact, the
first part of that is a G_BUILD_VECTOR whose legalization will produce a
new G_TRUNC from v2s32 to v2s16, and both G_TRUNCs will then get
combined to the original, causing a legalization cycle.

Make the lowering condition more precise, by requiring that the original
vector is >128 bits, which is I believe the only case where this
specific splitting approach is useful.

Note that this doesn't actually produce a legal result (the alwaysLegal
is a lie, as before), but it will cause a proper globalisel abort
instead of an infinite legalization loop.

Fixes https://github.com/llvm/llvm-project/issues/81244.

(cherry picked from commit 070848c17c2944afa494d42d3ad42929f3379842)
---
 .../AArch64/GISel/AArch64LegalizerInfo.cpp    |  5 ++--
 .../AArch64/GlobalISel/legalize-xtn.mir       | 24 +++++++++++++++++++
 2 files changed, 26 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
index fd69a7d6c33d03..4b9d549e791142 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
@@ -622,9 +622,8 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
       .lowerIf([=](const LegalityQuery &Query) {
         LLT DstTy = Query.Types[0];
         LLT SrcTy = Query.Types[1];
-        return DstTy.isVector() && (SrcTy.getSizeInBits() > 128 ||
-                                    (DstTy.getScalarSizeInBits() * 2 <
-                                     SrcTy.getScalarSizeInBits()));
+        return DstTy.isVector() && SrcTy.getSizeInBits() > 128 &&
+               DstTy.getScalarSizeInBits() * 2 <= SrcTy.getScalarSizeInBits();
       })
 
       .alwaysLegal();
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/legalize-xtn.mir b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-xtn.mir
index 16b780a8397347..661265173ae82b 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/legalize-xtn.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-xtn.mir
@@ -529,3 +529,27 @@ body:             |
     RET_ReallyLR implicit $q0
 
 ...
+
+---
+name:            pr81244
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $d0
+    ; CHECK-LABEL: name: pr81244
+    ; CHECK: liveins: $d0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $d0
+    ; CHECK-NEXT: [[TRUNC:%[0-9]+]]:_(<2 x s8>) = G_TRUNC [[COPY]](<2 x s32>)
+    ; CHECK-NEXT: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<2 x s8>), [[TRUNC]](<2 x s8>)
+    ; CHECK-NEXT: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[CONCAT_VECTORS]](<4 x s8>)
+    ; CHECK-NEXT: $d0 = COPY [[ANYEXT]](<4 x s16>)
+    ; CHECK-NEXT: RET_ReallyLR implicit $d0
+    %0:_(<2 x s32>) = COPY $d0
+    %1:_(<2 x s8>) = G_TRUNC %0(<2 x s32>)
+    %2:_(<4 x s8>) = G_CONCAT_VECTORS %1(<2 x s8>), %1(<2 x s8>)
+    %3:_(<4 x s16>) = G_ANYEXT %2(<4 x s8>)
+    $d0 = COPY %3(<4 x s16>)
+    RET_ReallyLR implicit $d0
+
+...