Fix multiversioning issues caused by the parallel llvm work (#52194)

So after struggling with this for a long while it seems there were two different issues. The first one we lacked coverage over, but the other was a very subtle issue when we sorted the fptrs. ~I still need to add test that does multiversioning where we call between multiversioned functions~ Fixes #52079
JuliaLang · Nov 24, 2023 · a386cd1 · a386cd1
1 parent 187e8c2
commit a386cd1
Show file tree

Hide file tree

Showing 4 changed files with 151 additions and 3 deletions.
diff --git a/src/llvm-multiversioning.cpp b/src/llvm-multiversioning.cpp
@@ -532,7 +532,7 @@ void CloneCtx::clone_decls()
             new_F->setVisibility(F->getVisibility());
             new_F->setDSOLocal(true);
             auto base_func = F;
-            if (specs[i].flags & JL_TARGET_CLONE_ALL)
+            if (!(specs[i].flags & JL_TARGET_CLONE_ALL))
                 base_func = static_cast<Group*>(linearized[specs[i].base])->base_func(F);
             (*linearized[i]->vmap)[base_func] = new_F;
         }
@@ -587,7 +587,7 @@ void CloneCtx::clone_bodies()
             }
             for (auto &target : groups[i].clones) {
                 prepare_vmap(*target.vmap);
-                auto target_F = cast_or_null<Function>(map_get(*target.vmap, F));
+                auto target_F = cast_or_null<Function>(map_get(*target.vmap, group_F));
                 if (target_F) {
                     if (!F->isDeclaration()) {
                         clone_function(group_F, target_F, *target.vmap);

diff --git a/src/processor.cpp b/src/processor.cpp
@@ -786,7 +786,10 @@ static inline jl_image_t parse_sysimg(void *hdl, F &&callback)
 
     if (!clones.empty()) {
         assert(!fvars.empty());
-        std::sort(clones.begin(), clones.end());
+        std::sort(clones.begin(), clones.end(),
+            [](const std::pair<uint32_t, const char *> &a, const std::pair<uint32_t, const char *> &b) {
+                return (a.first & jl_sysimg_val_mask) < (b.first & jl_sysimg_val_mask);
+        });
         auto clone_offsets = (int32_t *) malloc(sizeof(int32_t) * clones.size());
         auto clone_idxs = (uint32_t *) malloc(sizeof(uint32_t) * clones.size());
         for (size_t i = 0; i < clones.size(); i++) {

diff --git a/test/compiler/codegen.jl b/test/compiler/codegen.jl
@@ -860,3 +860,16 @@ foo50964(1) # Shouldn't assert!
 # https://github.com/JuliaLang/julia/issues/51233
 obj51233 = (1,)
 @test_throws ErrorException obj51233.x
+
+# Very specific test for multiversioning
+if Sys.ARCH === :x86_64
+    foo52079() = Core.Intrinsics.have_fma(Float64)
+    if foo52079() == true
+        let io = IOBuffer()
+            code_native(io,^,(Float64,Float64), dump_module=false)
+            str = String(take!(io))
+            @test !occursin("fma_emulated", str)
+            @test occursin("vfmadd", str)
+        end
+    end
+end
diff --git a/test/llvmpasses/multiversioning-x86.ll b/test/llvmpasses/multiversioning-x86.ll
@@ -0,0 +1,132 @@
+; This file is a part of Julia. License is MIT: https://julialang.org/license
+
+; RUN: opt -enable-new-pm=1 --opaque-pointers=0 --load-pass-plugin=libjulia-codegen%shlibext -passes='JuliaMultiVersioning,CPUFeatures' -S %s | FileCheck %s
+
+; RUN: opt -enable-new-pm=1 --opaque-pointers=1 --load-pass-plugin=libjulia-codegen%shlibext -passes='JuliaMultiVersioning,CPUFeatures' -S %s | FileCheck %s
+
+; COM: This test checks that multiversioning actually happens from start to finish
+; COM: We need the fvars for a proper test
+
+
+
+; CHECK: @jl_fvar_idxs = hidden constant [5 x i32] [i32 0, i32 1, i32 2, i32 3, i32 4], align 16
+; CHECK: @jl_gvar_idxs = hidden constant [0 x i32] zeroinitializer, align 16
+; TYPED: @simd_test.reloc_slot = hidden global i32 (<4 x i32>)* null
+; OPAQUE: @simd_test.reloc_slot = hidden global ptr null
+; TYPED: @jl_fvar_offsets = hidden constant [6 x i32] [i32 5, i32 0, i32 trunc (i64 sub (i64 ptrtoint (float (float, float)* @fastmath_test to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (i32 (i32)* @loop_test to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (i32 (<4 x i32>)* @simd_test to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (i32 (<4 x i32>)* @simd_test_call to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32)]
+; OPAQUE: @jl_fvar_offsets = hidden constant [6 x i32] [i32 5, i32 0, i32 trunc (i64 sub (i64 ptrtoint (ptr @fastmath_test to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @loop_test to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @simd_test to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @simd_test_call to i64), i64 ptrtoint (ptr @boring to i64)) to i32)]
+; CHECK: @jl_gvar_base = hidden constant i64 0
+; CHECK: @jl_gvar_offsets = hidden constant [1 x i32] zeroinitializer
+; TYPED: @jl_clone_slots = hidden constant [3 x i32] [i32 1, i32 3, i32 trunc (i64 sub (i64 ptrtoint (i32 (<4 x i32>)** @simd_test.reloc_slot to i64), i64 ptrtoint (i64* @jl_gvar_base to i64)) to i32)]
+; OPAQUE: @jl_clone_slots = hidden constant [3 x i32] [i32 1, i32 3, i32 trunc (i64 sub (i64 ptrtoint (ptr @simd_test.reloc_slot to i64), i64 ptrtoint (ptr @jl_gvar_base to i64)) to i32)]
+; CHECK: @jl_clone_idxs = hidden constant [10 x i32] [i32 -2147483647, i32 3, i32 -2147483647, i32 3, i32 4, i32 1, i32 1, i32 2, i32 -2147483645, i32 4]
+; TYPED: @jl_clone_offsets = hidden constant [9 x i32] [i32 trunc (i64 sub (i64 ptrtoint (i32 (i32)* @boring.1 to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (float (float, float)* @fastmath_test.1 to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (i32 (i32)* @loop_test.1 to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (i32 (<4 x i32>)* @simd_test.1 to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (i32 (<4 x i32>)* @simd_test_call.1 to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (float (float, float)* @fastmath_test.2 to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (i32 (i32)* @loop_test.2 to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (i32 (<4 x i32>)* @simd_test.2 to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (i32 (<4 x i32>)* @simd_test_call.2 to i64), i64 ptrtoint (i32 (i32)* @boring to i64)) to i32)]
+; OPAQUE: @jl_clone_offsets = hidden constant [9 x i32] [i32 trunc (i64 sub (i64 ptrtoint (ptr @boring.1 to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @fastmath_test.1 to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @loop_test.1 to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @simd_test.1 to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @simd_test_call.1 to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @fastmath_test.2 to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @loop_test.2 to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @simd_test.2 to i64), i64 ptrtoint (ptr @boring to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @simd_test_call.2 to i64), i64 ptrtoint (ptr @boring to i64)) to i32)]
+; TYPED: @jl_fvar_base = hidden alias i64, bitcast (i32 (i32)* @boring to i64*)
+; OPAQUE: @jl_fvar_base = hidden alias i64, ptr @boring
+
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128-ni:10:11:12:13"
+target triple = "x86_64-linux-gnu"
+
+@jl_fvars = global [5 x i64*] [i64* bitcast (i32 (i32)* @boring to i64*),
+                               i64* bitcast (float (float, float)* @fastmath_test to i64*),
+                               i64* bitcast (i32 (i32)* @loop_test to i64*),
+                               i64* bitcast (i32 (<4 x i32>)* @simd_test to i64*),
+                               i64* bitcast (i32 (<4 x i32>)* @simd_test_call to i64*)
+                              ], align 16
+@jl_gvars = global [0 x i64*] zeroinitializer, align 16
+@jl_fvar_idxs = hidden constant [5 x i32] [i32 0, i32 1, i32 2, i32 3, i32 4], align 16
+@jl_gvar_idxs = hidden constant [0 x i32] zeroinitializer, align 16
+
+declare i1 @julia.cpu.have_fma.f32()
+
+; CHECK: @boring{{.*}}#[[BORING_BASE:[0-9]+]]
+define noundef i32 @boring(i32 noundef %0) {
+  ret i32 %0
+}
+
+; CHECK: @fastmath_test{{.*}}#[[NOT_BORING_BASE:[0-9]+]]
+; CHECK: %3 = sitofp i1 false to float
+define noundef float @fastmath_test(float noundef %0, float noundef %1) {
+  %3 = call i1 @julia.cpu.have_fma.f32()
+  %4 = sitofp i1 %3 to float
+  %5 = fadd fast float %0, %4
+  ret float %5
+}
+
+; CHECK: @loop_test{{.*}}#[[NOT_BORING_BASE:[0-9]+]]
+define noundef i32 @loop_test(i32 noundef %0) {
+  %2 = icmp sgt i32 %0, 0
+  br i1 %2, label %5, label %3
+
+3:                                                ; preds = %5, %1
+  %4 = phi i32 [ 0, %1 ], [ %9, %5 ]
+  ret i32 %4
+
+5:                                                ; preds = %1, %5
+  %6 = phi i32 [ %10, %5 ], [ 0, %1 ]
+  %7 = phi i32 [ %9, %5 ], [ 0, %1 ]
+  %8 = lshr i32 %6, 1
+  %9 = add nuw nsw i32 %8, %7
+  %10 = add nuw nsw i32 %6, 1
+  %11 = icmp eq i32 %10, %0
+  br i1 %11, label %3, label %5;, !llvm.loop -
+}
+
+; CHECK: @simd_test{{.*}}#[[SIMD_BASE_RELOC:[0-9]+]]
+define noundef i32 @simd_test(<4 x i32> noundef %0) {
+  %2 = extractelement <4 x i32> %0, i64 0
+  ret i32 %2
+}
+
+; CHECK: @simd_test_call{{.*}}#[[NOT_BORING_BASE:[0-9]+]]
+define noundef i32 @simd_test_call(<4 x i32> noundef %0) {
+  %2 = call noundef i32 @simd_test(<4 x i32> noundef %0)
+  ret i32 %2
+}
+
+; CHECK: @boring{{.*}}#[[BORING_CLONE:[0-9]+]]
+
+; CHECK: @fastmath_test{{.*}}#[[NOT_BORING_CLONE1:[0-9]+]]
+; CHECK: %3 = sitofp i1 false to float
+
+; CHECK: @fastmath_test{{.*}}#[[NOT_BORING_CLONE2:[0-9]+]]
+; CHECK: %3 = sitofp i1 true to float
+
+; CHECK: @loop_test{{.*}}#[[NOT_BORING_CLONE1:[0-9]+]]
+
+; CHECK: @loop_test{{.*}}#[[NOT_BORING_CLONE2:[0-9]+]]
+
+; CHECK: @simd_test{{.*}}#[[SIMD_CLONE1:[0-9]+]]
+
+; CHECK: @simd_test{{.*}}#[[SIMD_CLONE2:[0-9]+]]
+
+; CHECK: @simd_test_call{{.*}}#[[NOT_BORING_CLONE1:[0-9]+]]
+; TYPED: %2 = load i32 (<4 x i32>)*, i32 (<4 x i32>)** @simd_test.reloc_slot, align 8, !tbaa !8, !invariant.load !12
+; OPAQUE: %2 = load ptr, ptr @simd_test.reloc_slot, align 8, !tbaa !8, !invariant.load !12
+; CHECK: %3 = call noundef i32 %2(<4 x i32> noundef %0)
+
+; CHECK: @simd_test_call{{.*}}#[[NOT_BORING_CLONE2:[0-9]+]]
+; CHECK: %2 = call noundef i32 @simd_test.2(<4 x i32> noundef %0)
+
+; CHECK-DAG: attributes #[[BORING_BASE]] = { "julia.mv.clone"="0" "julia.mv.clones"="2" "julia.mv.fvar" "target-cpu"="x86-64" "target-features"="+cx16,-sse3,-pclmul,-ssse3,-fma,-sse4.1,-sse4.2,-movbe,-popcnt,-aes,-xsave,-avx,-f16c,-rdrnd,-fsgsbase,-bmi,-avx2,-bmi2,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-sahf,-lzcnt,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8" }
+; CHECK-DAG: attributes #[[NOT_BORING_BASE]] = { "julia.mv.clone"="0" "julia.mv.clones"="6" "julia.mv.fvar" "target-cpu"="x86-64" "target-features"="+cx16,-sse3,-pclmul,-ssse3,-fma,-sse4.1,-sse4.2,-movbe,-popcnt,-aes,-xsave,-avx,-f16c,-rdrnd,-fsgsbase,-bmi,-avx2,-bmi2,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-sahf,-lzcnt,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8" }
+; CHECK-DAG: attributes #[[SIMD_BASE_RELOC]] = { "julia.mv.clone"="0" "julia.mv.clones"="6" "julia.mv.reloc" "target-cpu"="x86-64" "target-features"="+cx16,-sse3,-pclmul,-ssse3,-fma,-sse4.1,-sse4.2,-movbe,-popcnt,-aes,-xsave,-avx,-f16c,-rdrnd,-fsgsbase,-bmi,-avx2,-bmi2,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-sahf,-lzcnt,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8" }
+; CHECK-DAG: attributes #[[BORING_CLONE]] = { "julia.mv.clone"="1" "julia.mv.clones"="2" "julia.mv.fvar" "target-cpu"="sandybridge" "target-features"="+sahf,+avx,+xsave,+popcnt,+sse4.2,+sse4.1,+cx16,+ssse3,+pclmul,+sse3,-fma,-movbe,-aes,-f16c,-rdrnd,-fsgsbase,-bmi,-avx2,-bmi2,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-lzcnt,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8" }
+; CHECK-DAG: attributes #[[NOT_BORING_CLONE1]] = { "julia.mv.clone"="1" "julia.mv.clones"="6" "julia.mv.fvar" "target-cpu"="sandybridge" "target-features"="+sahf,+avx,+xsave,+popcnt,+sse4.2,+sse4.1,+cx16,+ssse3,+pclmul,+sse3,-fma,-movbe,-aes,-f16c,-rdrnd,-fsgsbase,-bmi,-avx2,-bmi2,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-lzcnt,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8" }
+; CHECK-DAG: attributes #[[NOT_BORING_CLONE2]] =  { "julia.mv.clone"="2" "julia.mv.clones"="6" "julia.mv.fvar" "target-cpu"="haswell" "target-features"="+lzcnt,+sahf,+bmi2,+avx2,+bmi,+fsgsbase,+f16c,+avx,+xsave,+popcnt,+movbe,+sse4.2,+sse4.1,+cx16,+fma,+ssse3,+pclmul,+sse3,-aes,-rdrnd,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8" }
+; CHECK-DAG: attributes #[[SIMD_CLONE1]] = { "julia.mv.clone"="1" "julia.mv.clones"="6" "julia.mv.reloc" "target-cpu"="sandybridge" "target-features"="+sahf,+avx,+xsave,+popcnt,+sse4.2,+sse4.1,+cx16,+ssse3,+pclmul,+sse3,-fma,-movbe,-aes,-f16c,-rdrnd,-fsgsbase,-bmi,-avx2,-bmi2,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-lzcnt,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8" }
+; CHECK-DAG: attributes #[[SIMD_CLONE2]] = { "julia.mv.clone"="2" "julia.mv.clones"="6" "julia.mv.reloc" "target-cpu"="haswell" "target-features"="+lzcnt,+sahf,+bmi2,+avx2,+bmi,+fsgsbase,+f16c,+avx,+xsave,+popcnt,+movbe,+sse4.2,+sse4.1,+cx16,+fma,+ssse3,+pclmul,+sse3,-aes,-rdrnd,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8" }
+
+
+!llvm.module.flags = !{!0, !2}
+
+
+!0 = !{i32 1, !"julia.mv.enable", i32 1}
+!1 = !{!1}
+!2 = !{i32 1, !"julia.mv.specs", !3}
+!3 = !{!4, !5, !6}
+!4 = !{!"x86-64", !"+cx16,-sse3,-pclmul,-ssse3,-fma,-sse4.1,-sse4.2,-movbe,-popcnt,-aes,-xsave,-avx,-f16c,-rdrnd,-fsgsbase,-bmi,-avx2,-bmi2,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-sahf,-lzcnt,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8", i32 0, i32 0}
+!5 = !{!"sandybridge", !"+sahf,+avx,+xsave,+popcnt,+sse4.2,+sse4.1,+cx16,+ssse3,+pclmul,+sse3,-fma,-movbe,-aes,-f16c,-rdrnd,-fsgsbase,-bmi,-avx2,-bmi2,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-lzcnt,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8", i32 0, i32 2}
+!6 = !{!"haswell", !"+lzcnt,+sahf,+bmi2,+avx2,+bmi,+fsgsbase,+f16c,+avx,+xsave,+popcnt,+movbe,+sse4.2,+sse4.1,+cx16,+fma,+ssse3,+pclmul,+sse3,-aes,-rdrnd,-rtm,-avx512f,-avx512dq,-rdseed,-adx,-avx512ifma,-clflushopt,-clwb,-avx512pf,-avx512er,-avx512cd,-sha,-avx512bw,-avx512vl,-prefetchwt1,-avx512vbmi,-pku,-waitpkg,-avx512vbmi2,-shstk,-gfni,-vaes,-vpclmulqdq,-avx512vnni,-avx512bitalg,-avx512vpopcntdq,-rdpid,-cldemote,-movdiri,-movdir64b,-enqcmd,-uintr,-avx512vp2intersect,-serialize,-tsxldtrk,-pconfig,-amx-bf16,-avx512fp16,-amx-tile,-amx-int8,-sse4a,-prfchw,-xop,-fma4,-tbm,-mwaitx,-xsaveopt,-xsavec,-xsaves,-clzero,-wbnoinvd,-avxvnni,-avx512bf16,-ptwrite,+sse2,+mmx,+fxsr,+64bit,+cx8", i32 1, i32 284}