[PowerPC] Incorrect store alignment with __builtin_vsx_build_pair() builtin.

[nikic]

Function test2 from https://github.com/llvm/llvm-project/blob/main/clang/test/CodeGen/PowerPC/builtins-ppc-build-pair-mma.c generates this pre-optimization IR:

define dso_local void @test2(ptr noundef %vqp, ptr noundef %vpp, <16 x i8> noundef %vc1, <16 x i8> noundef %vc2, ptr noundef %resp) #0 {
entry:
  %vqp.addr = alloca ptr, align 8
  %vpp.addr = alloca ptr, align 8
  %vc1.addr = alloca <16 x i8>, align 16
  %vc2.addr = alloca <16 x i8>, align 16
  %resp.addr = alloca ptr, align 8
  %vq = alloca <512 x i1>, align 64
  %vp = alloca <256 x i1>, align 32
  %res = alloca <256 x i1>, align 32
  store ptr %vqp, ptr %vqp.addr, align 8, !tbaa !2
  store ptr %vpp, ptr %vpp.addr, align 8, !tbaa !2
  store <16 x i8> %vc1, ptr %vc1.addr, align 16, !tbaa !6
  store <16 x i8> %vc2, ptr %vc2.addr, align 16, !tbaa !6
  store ptr %resp, ptr %resp.addr, align 8, !tbaa !2
  call void @llvm.lifetime.start.p0(i64 64, ptr %vq) #3
  %0 = load ptr, ptr %vqp.addr, align 8, !tbaa !2
  %1 = load <512 x i1>, ptr %0, align 64, !tbaa !7
  store <512 x i1> %1, ptr %vq, align 64, !tbaa !7
  call void @llvm.lifetime.start.p0(i64 32, ptr %vp) #3
  %2 = load ptr, ptr %vpp.addr, align 8, !tbaa !2
  %3 = load <256 x i1>, ptr %2, align 32, !tbaa !9
  store <256 x i1> %3, ptr %vp, align 32, !tbaa !9
  call void @llvm.lifetime.start.p0(i64 32, ptr %res) #3
  %4 = load <16 x i8>, ptr %vc1.addr, align 16, !tbaa !6
  %5 = load <16 x i8>, ptr %vc2.addr, align 16, !tbaa !6
  %6 = call <256 x i1> @llvm.ppc.vsx.assemble.pair(<16 x i8> %5, <16 x i8> %4)
  store <256 x i1> %6, ptr %res, align 64
  %7 = load <256 x i1>, ptr %res, align 32, !tbaa !9
  %8 = load ptr, ptr %resp.addr, align 8, !tbaa !2
  store <256 x i1> %7, ptr %8, align 32, !tbaa !9
  call void @llvm.lifetime.end.p0(i64 32, ptr %res) #3
  call void @llvm.lifetime.end.p0(i64 32, ptr %vp) #3
  call void @llvm.lifetime.end.p0(i64 64, ptr %vq) #3
  ret void
}

Note that store <256 x i1> %6, ptr %res, align 64 performs a 64 aligned store of a 32 byte type to a 32 aligned alloca. Pretty sure this was supposed to get alignment 32.

I think this is due to the hardcoded 64 alignment in

llvm-project/clang/lib/CodeGen/CGBuiltin.cpp

Line 18219 in c2b92a4

return Builder.CreateAlignedStore(Call, Ops[0], MaybeAlign(64));

.

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: Nikita Popov (nikic)

Function test2 from https://github.com/llvm/llvm-project/blob/main/clang/test/CodeGen/PowerPC/builtins-ppc-build-pair-mma.c generates this pre-optimization IR:

define dso_local void @<!-- -->test2(ptr noundef %vqp, ptr noundef %vpp, &lt;16 x i8&gt; noundef %vc1, &lt;16 x i8&gt; noundef %vc2, ptr noundef %resp) #<!-- -->0 {
entry:
  %vqp.addr = alloca ptr, align 8
  %vpp.addr = alloca ptr, align 8
  %vc1.addr = alloca &lt;16 x i8&gt;, align 16
  %vc2.addr = alloca &lt;16 x i8&gt;, align 16
  %resp.addr = alloca ptr, align 8
  %vq = alloca &lt;512 x i1&gt;, align 64
  %vp = alloca &lt;256 x i1&gt;, align 32
  %res = alloca &lt;256 x i1&gt;, align 32
  store ptr %vqp, ptr %vqp.addr, align 8, !tbaa !2
  store ptr %vpp, ptr %vpp.addr, align 8, !tbaa !2
  store &lt;16 x i8&gt; %vc1, ptr %vc1.addr, align 16, !tbaa !6
  store &lt;16 x i8&gt; %vc2, ptr %vc2.addr, align 16, !tbaa !6
  store ptr %resp, ptr %resp.addr, align 8, !tbaa !2
  call void @<!-- -->llvm.lifetime.start.p0(i64 64, ptr %vq) #<!-- -->3
  %0 = load ptr, ptr %vqp.addr, align 8, !tbaa !2
  %1 = load &lt;512 x i1&gt;, ptr %0, align 64, !tbaa !7
  store &lt;512 x i1&gt; %1, ptr %vq, align 64, !tbaa !7
  call void @<!-- -->llvm.lifetime.start.p0(i64 32, ptr %vp) #<!-- -->3
  %2 = load ptr, ptr %vpp.addr, align 8, !tbaa !2
  %3 = load &lt;256 x i1&gt;, ptr %2, align 32, !tbaa !9
  store &lt;256 x i1&gt; %3, ptr %vp, align 32, !tbaa !9
  call void @<!-- -->llvm.lifetime.start.p0(i64 32, ptr %res) #<!-- -->3
  %4 = load &lt;16 x i8&gt;, ptr %vc1.addr, align 16, !tbaa !6
  %5 = load &lt;16 x i8&gt;, ptr %vc2.addr, align 16, !tbaa !6
  %6 = call &lt;256 x i1&gt; @<!-- -->llvm.ppc.vsx.assemble.pair(&lt;16 x i8&gt; %5, &lt;16 x i8&gt; %4)
  store &lt;256 x i1&gt; %6, ptr %res, align 64
  %7 = load &lt;256 x i1&gt;, ptr %res, align 32, !tbaa !9
  %8 = load ptr, ptr %resp.addr, align 8, !tbaa !2
  store &lt;256 x i1&gt; %7, ptr %8, align 32, !tbaa !9
  call void @<!-- -->llvm.lifetime.end.p0(i64 32, ptr %res) #<!-- -->3
  call void @<!-- -->llvm.lifetime.end.p0(i64 32, ptr %vp) #<!-- -->3
  call void @<!-- -->llvm.lifetime.end.p0(i64 64, ptr %vq) #<!-- -->3
  ret void
}

Note that store <256 x i1> %6, ptr %res, align 64 performs a 64 aligned store of a 32 byte type to a 32 aligned alloca. Pretty sure this was supposed to get alignment 32.

I think this is due to the hardcoded 64 alignment in

llvm-project/clang/lib/CodeGen/CGBuiltin.cpp

Line 18219 in c2b92a4

return Builder.CreateAlignedStore(Call, Ops[0], MaybeAlign(64));

.

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

Author: Nikita Popov (nikic)

Function test2 from https://github.com/llvm/llvm-project/blob/main/clang/test/CodeGen/PowerPC/builtins-ppc-build-pair-mma.c generates this pre-optimization IR:

define dso_local void @<!-- -->test2(ptr noundef %vqp, ptr noundef %vpp, &lt;16 x i8&gt; noundef %vc1, &lt;16 x i8&gt; noundef %vc2, ptr noundef %resp) #<!-- -->0 {
entry:
  %vqp.addr = alloca ptr, align 8
  %vpp.addr = alloca ptr, align 8
  %vc1.addr = alloca &lt;16 x i8&gt;, align 16
  %vc2.addr = alloca &lt;16 x i8&gt;, align 16
  %resp.addr = alloca ptr, align 8
  %vq = alloca &lt;512 x i1&gt;, align 64
  %vp = alloca &lt;256 x i1&gt;, align 32
  %res = alloca &lt;256 x i1&gt;, align 32
  store ptr %vqp, ptr %vqp.addr, align 8, !tbaa !2
  store ptr %vpp, ptr %vpp.addr, align 8, !tbaa !2
  store &lt;16 x i8&gt; %vc1, ptr %vc1.addr, align 16, !tbaa !6
  store &lt;16 x i8&gt; %vc2, ptr %vc2.addr, align 16, !tbaa !6
  store ptr %resp, ptr %resp.addr, align 8, !tbaa !2
  call void @<!-- -->llvm.lifetime.start.p0(i64 64, ptr %vq) #<!-- -->3
  %0 = load ptr, ptr %vqp.addr, align 8, !tbaa !2
  %1 = load &lt;512 x i1&gt;, ptr %0, align 64, !tbaa !7
  store &lt;512 x i1&gt; %1, ptr %vq, align 64, !tbaa !7
  call void @<!-- -->llvm.lifetime.start.p0(i64 32, ptr %vp) #<!-- -->3
  %2 = load ptr, ptr %vpp.addr, align 8, !tbaa !2
  %3 = load &lt;256 x i1&gt;, ptr %2, align 32, !tbaa !9
  store &lt;256 x i1&gt; %3, ptr %vp, align 32, !tbaa !9
  call void @<!-- -->llvm.lifetime.start.p0(i64 32, ptr %res) #<!-- -->3
  %4 = load &lt;16 x i8&gt;, ptr %vc1.addr, align 16, !tbaa !6
  %5 = load &lt;16 x i8&gt;, ptr %vc2.addr, align 16, !tbaa !6
  %6 = call &lt;256 x i1&gt; @<!-- -->llvm.ppc.vsx.assemble.pair(&lt;16 x i8&gt; %5, &lt;16 x i8&gt; %4)
  store &lt;256 x i1&gt; %6, ptr %res, align 64
  %7 = load &lt;256 x i1&gt;, ptr %res, align 32, !tbaa !9
  %8 = load ptr, ptr %resp.addr, align 8, !tbaa !2
  store &lt;256 x i1&gt; %7, ptr %8, align 32, !tbaa !9
  call void @<!-- -->llvm.lifetime.end.p0(i64 32, ptr %res) #<!-- -->3
  call void @<!-- -->llvm.lifetime.end.p0(i64 32, ptr %vp) #<!-- -->3
  call void @<!-- -->llvm.lifetime.end.p0(i64 64, ptr %vq) #<!-- -->3
  ret void
}

Note that store <256 x i1> %6, ptr %res, align 64 performs a 64 aligned store of a 32 byte type to a 32 aligned alloca. Pretty sure this was supposed to get alignment 32.

I think this is due to the hardcoded 64 alignment in

llvm-project/clang/lib/CodeGen/CGBuiltin.cpp

Line 18219 in c2b92a4

return Builder.CreateAlignedStore(Call, Ops[0], MaybeAlign(64));

.