fix issue with AVX codegen

PKGBUILD

@@ -4,7 +4,7 @@
 
 pkgname=('lib32-llvm' 'lib32-llvm-libs')
 pkgver=12.0.1
-pkgrel=1
+pkgrel=2
 arch=('x86_64')
 url="https://llvm.org/"
 license=('custom:Apache 2.0 with LLVM Exception')
@@ -15,11 +15,13 @@ _source_base=https://github.com/llvm/llvm-project/releases/download/llvmorg-$pkg
 source=($_source_base/llvm-$pkgver.src.tar.xz{,.sig}
         llvm-link-with-Bsymbolic-functions.patch
         add-fno-semantic-interposition.patch
+        x86-twist-shuffle-mask.patch
         no-strict-aliasing-DwarfCompileUnit.patch)
 sha256sums=('7d9a8405f557cefc5a21bf5672af73903b64749d9bc3a50322239f56f34ffddf'
             'SKIP'
             '560ce1e206c19f4b86f4c583b743db0ad47a610418999350710aafd60ae50fcd'
             'fc8c64267a5d179e9fc24fb2bc6150edef2598c83f5b2d138d14e05ce9f4e345'
+            'c51b8754f76eb3774f46d530409f6d89f5bb47d90f0d718dbfa861f716b29693'
             'd1eff24508e35aae6c26a943dbaa3ef5acb60a145b008fd1ef9ac6f6c4faa662')
 validpgpkeys+=('B6C8F98282B944E3B0D5C2530FC3042E345AD05D') # Hans Wennborg <hans@chromium.org>
 validpgpkeys+=('474E22316ABF4785A88C6E8EA2C794A986419D8A') # Tom Stellard <tstellar@redhat.com>
@@ -33,6 +35,9 @@ prepare() {
   # https://reviews.llvm.org/D102453
   patch -Np2 -i ../add-fno-semantic-interposition.patch
 
+  # https://bugs.llvm.org/show_bug.cgi?id=50823
+  patch -Np2 -i ../x86-twist-shuffle-mask.patch
+
   # https://bugs.llvm.org/show_bug.cgi?id=50611#c3
   patch -Np2 -i ../no-strict-aliasing-DwarfCompileUnit.patch
 }

x86-twist-shuffle-mask.patch

+From 2c3bca2c3f13a0a9ef71d549a90fba23e6997d44 Mon Sep 17 00:00:00 2001
+From: "Wang, Pengfei" <pengfei.wang@intel.com>
+Date: Mon, 5 Jul 2021 21:08:49 +0800
+Subject: [PATCH] Twist shuffle mask when fold HOP(SHUFFLE(X,Y),SHUFFLE(X,Y))
+ -> SHUFFLE(HOP(X,Y))
+
+This patch fixes PR50823.
+
+The shuffle mask should be twisted twice before gotten the correct one due to the difference between inner HOP and outer.
+
+Reviewed By: RKSimon
+
+Differential Revision: https://reviews.llvm.org/D104903
+
+(cherry picked from commit 9ab99f773fec7da4183495a3fdc655a797d3bea2)
+---
+ llvm/lib/Target/X86/X86ISelLowering.cpp |  7 ++---
+ llvm/test/CodeGen/X86/haddsub-undef.ll  |  4 +--
+ llvm/test/CodeGen/X86/packss.ll         |  2 +-
+ llvm/test/CodeGen/X86/pr50823.ll        | 35 +++++++++++++++++++++++++
+ 4 files changed, 42 insertions(+), 6 deletions(-)
+ create mode 100644 llvm/test/CodeGen/X86/pr50823.ll
+
+diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
+index 1e2407c7e7f6..d8b2f765e953 100644
+--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
++++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
+@@ -43194,9 +43194,10 @@ static SDValue combineHorizOpWithShuffle(SDNode *N, SelectionDAG &DAG,
+           ShuffleVectorSDNode::commuteMask(ShuffleMask1);
+         }
+         if ((Op00 == Op10) && (Op01 == Op11)) {
+-          SmallVector<int, 4> ShuffleMask;
+-          ShuffleMask.append(ShuffleMask0.begin(), ShuffleMask0.end());
+-          ShuffleMask.append(ShuffleMask1.begin(), ShuffleMask1.end());
++          const int Map[4] = {0, 2, 1, 3};
++          SmallVector<int, 4> ShuffleMask(
++            {Map[ShuffleMask0[0]], Map[ShuffleMask1[0]], Map[ShuffleMask0[1]],
++             Map[ShuffleMask1[1]]});
+           SDLoc DL(N);
+           MVT ShufVT = VT.isFloatingPoint() ? MVT::v4f64 : MVT::v4i64;
+           SDValue Res = DAG.getNode(Opcode, DL, VT, Op00, Op01);
+diff --git a/llvm/test/CodeGen/X86/haddsub-undef.ll b/llvm/test/CodeGen/X86/haddsub-undef.ll
+index 68d058433179..e7c8b84d3bc7 100644
+--- a/llvm/test/CodeGen/X86/haddsub-undef.ll
++++ b/llvm/test/CodeGen/X86/haddsub-undef.ll
+@@ -1166,7 +1166,7 @@ define <4 x double> @PR34724_add_v4f64_u123(<4 x double> %0, <4 x double> %1) {
+ ; AVX512-FAST:       # %bb.0:
+ ; AVX512-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm0
+ ; AVX512-FAST-NEXT:    vhaddpd %ymm1, %ymm0, %ymm0
+-; AVX512-FAST-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]
++; AVX512-FAST-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,0,1,3]
+ ; AVX512-FAST-NEXT:    retq
+   %3 = shufflevector <4 x double> %0, <4 x double> %1, <2 x i32> <i32 2, i32 4>
+   %4 = shufflevector <4 x double> %0, <4 x double> %1, <2 x i32> <i32 3, i32 5>
+@@ -1267,7 +1267,7 @@ define <4 x double> @PR34724_add_v4f64_01u3(<4 x double> %0, <4 x double> %1) {
+ ; AVX512-FAST-LABEL: PR34724_add_v4f64_01u3:
+ ; AVX512-FAST:       # %bb.0:
+ ; AVX512-FAST-NEXT:    vhaddpd %ymm1, %ymm0, %ymm0
+-; AVX512-FAST-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,3,1,3]
++; AVX512-FAST-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,3,3]
+ ; AVX512-FAST-NEXT:    retq
+   %3 = shufflevector <4 x double> %0, <4 x double> undef, <2 x i32> <i32 0, i32 2>
+   %4 = shufflevector <4 x double> %0, <4 x double> undef, <2 x i32> <i32 1, i32 3>
+diff --git a/llvm/test/CodeGen/X86/packss.ll b/llvm/test/CodeGen/X86/packss.ll
+index 16349ae2c7f9..ac431b7556ea 100644
+--- a/llvm/test/CodeGen/X86/packss.ll
++++ b/llvm/test/CodeGen/X86/packss.ll
+@@ -370,7 +370,7 @@ define <32 x i8> @packsswb_icmp_zero_trunc_256(<16 x i16> %a0) {
+ ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+ ; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
+ ; AVX2-NEXT:    vpacksswb %ymm0, %ymm1, %ymm0
+-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,3]
++; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]
+ ; AVX2-NEXT:    ret{{[l|q]}}
+   %1 = icmp eq <16 x i16> %a0, zeroinitializer
+   %2 = sext <16 x i1> %1 to <16 x i16>
+diff --git a/llvm/test/CodeGen/X86/pr50823.ll b/llvm/test/CodeGen/X86/pr50823.ll
+new file mode 100644
+index 000000000000..c5d5296e5c66
+--- /dev/null
++++ b/llvm/test/CodeGen/X86/pr50823.ll
+@@ -0,0 +1,35 @@
++; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
++; RUN: llc < %s -mtriple=x86_64-unknown -mcpu=core-avx2 | FileCheck %s
++
++%v8_uniform_FVector3 = type { float, float, float }
++
++declare <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float>, <8 x float>)
++
++define void @foo(%v8_uniform_FVector3* %Out, float* %In, <8 x i32> %__mask) {
++; CHECK-LABEL: foo:
++; CHECK:       # %bb.0: # %allocas
++; CHECK-NEXT:    vmovups (%rsi), %xmm0
++; CHECK-NEXT:    vhaddps 32(%rsi), %xmm0, %xmm0
++; CHECK-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,0,1,1]
++; CHECK-NEXT:    vhaddps %ymm0, %ymm0, %ymm0
++; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
++; CHECK-NEXT:    vaddss %xmm1, %xmm0, %xmm0
++; CHECK-NEXT:    vmovss %xmm0, (%rdi)
++; CHECK-NEXT:    vzeroupper
++; CHECK-NEXT:    retq
++allocas:
++  %ptr_cast_for_load = bitcast float* %In to <8 x float>*
++  %ptr_masked_load74 = load <8 x float>, <8 x float>* %ptr_cast_for_load, align 4
++  %ptr8096 = getelementptr float, float* %In, i64 8
++  %ptr_cast_for_load81 = bitcast float* %ptr8096 to <8 x float>*
++  %ptr80_masked_load82 = load <8 x float>, <8 x float>* %ptr_cast_for_load81, align 4
++  %ret_7.i.i = shufflevector <8 x float> %ptr_masked_load74, <8 x float> %ptr80_masked_load82, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11>
++  %Out_load19 = getelementptr %v8_uniform_FVector3, %v8_uniform_FVector3* %Out, i64 0, i32 0
++  %v1.i.i100 = tail call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %ret_7.i.i, <8 x float> %ret_7.i.i)
++  %v2.i.i101 = tail call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %v1.i.i100, <8 x float> %v1.i.i100)
++  %scalar1.i.i102 = extractelement <8 x float> %v2.i.i101, i32 0
++  %scalar2.i.i103 = extractelement <8 x float> %v2.i.i101, i32 4
++  %sum.i.i104 = fadd float %scalar1.i.i102, %scalar2.i.i103
++  store float %sum.i.i104, float* %Out_load19, align 4
++  ret void
++}