Sebastien, Robert,

I'd like to backport to llvm-22
https://github.com/llvm/llvm-project/commit/1b9fea021840f17c41ea980300d0fc45e7285909
which fixes a "compiler stuck" issue in libggml.

I think it should go to base version of llvm as well.

Here the diff for ports@ part.

Ok?

Index: Makefile
===================================================================
RCS file: /home/cvs/ports/devel/llvm/22/Makefile,v
diff -u -p -r1.6 Makefile
--- Makefile    6 May 2026 12:47:03 -0000       1.6
+++ Makefile    20 May 2026 11:01:28 -0000
@@ -5,6 +5,8 @@ LLVM_MAJOR =    22
 LLVM_VERSION = ${LLVM_MAJOR}.1.5
 LLVM_PKGSPEC = >=22,<23
 
+REVISION-main =        0
+
 SHARED_LIBS += LLVM            0.0 \
                LTO             0.0 \
                Remarks         0.0 \
Index: patches/patch-llvm_lib_Target_X86_X86ISelLowering_cpp
===================================================================
RCS file: patches/patch-llvm_lib_Target_X86_X86ISelLowering_cpp
diff -N patches/patch-llvm_lib_Target_X86_X86ISelLowering_cpp
--- /dev/null   1 Jan 1970 00:00:00 -0000
+++ patches/patch-llvm_lib_Target_X86_X86ISelLowering_cpp       20 May 2026 
10:59:58 -0000
@@ -0,0 +1,31 @@
+Backport 
https://github.com/llvm/llvm-project/commit/1b9fea021840f17c41ea980300d0fc45e7285909
+
+Index: llvm/lib/Target/X86/X86ISelLowering.cpp
+--- llvm/lib/Target/X86/X86ISelLowering.cpp.orig
++++ llvm/lib/Target/X86/X86ISelLowering.cpp
+@@ -18004,9 +18004,12 @@ static SDValue lowerV64I8Shuffle(const SDLoc &DL, Arra
+ 
+   // Try to create an in-lane repeating shuffle mask and then shuffle the
+   // results into the target lanes.
+-  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
+-          DL, MVT::v64i8, V1, V2, Mask, Subtarget, DAG))
+-    return V;
++  // FIXME: Avoid on VBMI targets as the post lane permute often interferes
++  // with shuffle combining (should be fixed by topological DAG sorting).
++  if (!Subtarget.hasVBMI())
++    if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
++            DL, MVT::v64i8, V1, V2, Mask, Subtarget, DAG))
++      return V;
+ 
+   if (SDValue Result = lowerShuffleAsLanePermuteAndPermute(
+           DL, MVT::v64i8, V1, V2, Mask, DAG, Subtarget))
+@@ -18041,7 +18044,8 @@ static SDValue lowerV64I8Shuffle(const SDLoc &DL, Arra
+   if (Subtarget.hasVBMI())
+     return lowerShuffleWithPERMV(DL, MVT::v64i8, Mask, V1, V2, Subtarget, 
DAG);
+ 
+-  return splitAndLowerShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG, 
/*SimpleOnly*/ false);
++  return splitAndLowerShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG,
++                              /*SimpleOnly*/ false);
+ }
+ 
+ /// High-level routine to lower various 512-bit x86 vector shuffles.


-- 
wbr, Kirill

Reply via email to