[llvm] X86: Fix asserting on bfloat argument/return without sse2 (PR #93146)

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Thu May 23 00:48:50 PDT 2024


https://github.com/arsenm created https://github.com/llvm/llvm-project/pull/93146

These now get the default promote-to-float behavior, like half does.

Fixes #92899

>From 9b8131d23bb6c83a35f6a3ea361c512adeddf5b1 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Thu, 23 May 2024 07:56:23 +0200
Subject: [PATCH] X86: Fix asserting on bfloat argument/return without sse2

These now get the default promote-to-float behavior, like half does.

Fixes #92899
---
 llvm/lib/Target/X86/X86ISelLoweringCall.cpp   |   18 +-
 .../X86/bfloat-calling-conv-no-sse2.ll        | 1441 +++++++++++++++++
 2 files changed, 1452 insertions(+), 7 deletions(-)
 create mode 100644 llvm/test/CodeGen/X86/bfloat-calling-conv-no-sse2.ll

diff --git a/llvm/lib/Target/X86/X86ISelLoweringCall.cpp b/llvm/lib/Target/X86/X86ISelLoweringCall.cpp
index b107d56f8cf9b..4fddb60db5f10 100644
--- a/llvm/lib/Target/X86/X86ISelLoweringCall.cpp
+++ b/llvm/lib/Target/X86/X86ISelLoweringCall.cpp
@@ -123,12 +123,14 @@ MVT X86TargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
       !Subtarget.hasX87())
     return MVT::i32;
 
-  if (VT.isVector() && VT.getVectorElementType() == MVT::bf16)
-    return getRegisterTypeForCallingConv(Context, CC,
-                                         VT.changeVectorElementType(MVT::f16));
+  if (isTypeLegal(MVT::f16)) {
+    if (VT.isVector() && VT.getVectorElementType() == MVT::bf16)
+      return getRegisterTypeForCallingConv(
+          Context, CC, VT.changeVectorElementType(MVT::f16));
 
-  if (VT == MVT::bf16)
-    return MVT::f16;
+    if (VT == MVT::bf16)
+      return MVT::f16;
+  }
 
   return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);
 }
@@ -161,7 +163,8 @@ unsigned X86TargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
       return 3;
   }
 
-  if (VT.isVector() && VT.getVectorElementType() == MVT::bf16)
+  if (VT.isVector() && VT.getVectorElementType() == MVT::bf16 &&
+      isTypeLegal(MVT::f16))
     return getNumRegistersForCallingConv(Context, CC,
                                          VT.changeVectorElementType(MVT::f16));
 
@@ -193,7 +196,8 @@ unsigned X86TargetLowering::getVectorTypeBreakdownForCallingConv(
   }
 
   // Split vNbf16 vectors according to vNf16.
-  if (VT.isVector() && VT.getVectorElementType() == MVT::bf16)
+  if (VT.isVector() && VT.getVectorElementType() == MVT::bf16 &&
+      isTypeLegal(MVT::f16))
     VT = VT.changeVectorElementType(MVT::f16);
 
   return TargetLowering::getVectorTypeBreakdownForCallingConv(Context, CC, VT, IntermediateVT,
diff --git a/llvm/test/CodeGen/X86/bfloat-calling-conv-no-sse2.ll b/llvm/test/CodeGen/X86/bfloat-calling-conv-no-sse2.ll
new file mode 100644
index 0000000000000..05f036201f4b1
--- /dev/null
+++ b/llvm/test/CodeGen/X86/bfloat-calling-conv-no-sse2.ll
@@ -0,0 +1,1441 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=i386-linux-gnu < %s | FileCheck -check-prefixes=CHECK,NOSSE %s
+; RUN: llc -mtriple=i386-linux-gnu -mattr=+sse < %s | FileCheck -check-prefixes=CHECK,SSE %s
+
+; Make sure no assert without SSE2 and bfloat. Issue 92899
+
+define bfloat @return_arg_bf16(bfloat %x) {
+; CHECK-LABEL: return_arg_bf16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    flds {{[0-9]+}}(%esp)
+; CHECK-NEXT:    retl
+  ret bfloat %x
+}
+
+define <2 x bfloat> @return_arg_v2bf16(<2 x bfloat> %x) {
+; CHECK-LABEL: return_arg_v2bf16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    flds {{[0-9]+}}(%esp)
+; CHECK-NEXT:    flds {{[0-9]+}}(%esp)
+; CHECK-NEXT:    retl
+  ret <2 x bfloat> %x
+}
+
+define <3 x bfloat> @return_arg_v3bf16(<3 x bfloat> %x) {
+; NOSSE-LABEL: return_arg_v3bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    pushl %eax
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    .cfi_offset %esi, -12
+; NOSSE-NEXT:    .cfi_offset %edi, -8
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %esi
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    # kill: def $ax killed $ax def $eax
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movzwl %si, %edi
+; NOSSE-NEXT:    orl %eax, %edi
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, 4(%esi)
+; NOSSE-NEXT:    movl %edi, (%esi)
+; NOSSE-NEXT:    movl %esi, %eax
+; NOSSE-NEXT:    addl $4, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl $4
+;
+; SSE-LABEL: return_arg_v3bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    pushl %eax
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    .cfi_offset %esi, -12
+; SSE-NEXT:    .cfi_offset %edi, -8
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %esi
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    # kill: def $ax killed $ax def $eax
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movzwl %si, %edi
+; SSE-NEXT:    orl %eax, %edi
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, 4(%esi)
+; SSE-NEXT:    movl %edi, (%esi)
+; SSE-NEXT:    movl %esi, %eax
+; SSE-NEXT:    addl $4, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    popl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 4
+; SSE-NEXT:    retl $4
+  ret <3 x bfloat> %x
+}
+
+define <4 x bfloat> @return_arg_v4bf16(<4 x bfloat> %x) {
+; NOSSE-LABEL: return_arg_v4bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    pushl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    pushl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 20
+; NOSSE-NEXT:    subl $12, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 32
+; NOSSE-NEXT:    .cfi_offset %esi, -20
+; NOSSE-NEXT:    .cfi_offset %edi, -16
+; NOSSE-NEXT:    .cfi_offset %ebx, -12
+; NOSSE-NEXT:    .cfi_offset %ebp, -8
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %esi
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %edi
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %ebx
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %ebp
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, 6(%ebp)
+; NOSSE-NEXT:    movw %bx, 4(%ebp)
+; NOSSE-NEXT:    movw %di, 2(%ebp)
+; NOSSE-NEXT:    movw %si, (%ebp)
+; NOSSE-NEXT:    movl %ebp, %eax
+; NOSSE-NEXT:    addl $12, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 20
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    popl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    popl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl $4
+;
+; SSE-LABEL: return_arg_v4bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %ebp
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    pushl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    pushl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 20
+; SSE-NEXT:    subl $12, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 32
+; SSE-NEXT:    .cfi_offset %esi, -20
+; SSE-NEXT:    .cfi_offset %edi, -16
+; SSE-NEXT:    .cfi_offset %ebx, -12
+; SSE-NEXT:    .cfi_offset %ebp, -8
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %esi
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %edi
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %ebx
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %ebp
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, 6(%ebp)
+; SSE-NEXT:    movw %bx, 4(%ebp)
+; SSE-NEXT:    movw %di, 2(%ebp)
+; SSE-NEXT:    movw %si, (%ebp)
+; SSE-NEXT:    movl %ebp, %eax
+; SSE-NEXT:    addl $12, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 20
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    popl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    popl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    popl %ebp
+; SSE-NEXT:    .cfi_def_cfa_offset 4
+; SSE-NEXT:    retl $4
+  ret <4 x bfloat> %x
+}
+
+define <8 x bfloat> @return_arg_v8bf16(<8 x bfloat> %x) {
+; NOSSE-LABEL: return_arg_v8bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    pushl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    pushl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 20
+; NOSSE-NEXT:    subl $12, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 32
+; NOSSE-NEXT:    .cfi_offset %esi, -20
+; NOSSE-NEXT:    .cfi_offset %edi, -16
+; NOSSE-NEXT:    .cfi_offset %ebx, -12
+; NOSSE-NEXT:    .cfi_offset %ebp, -8
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %esi
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %edi
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %ebx
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %ebp
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, 14(%ebp)
+; NOSSE-NEXT:    movw %bx, 12(%ebp)
+; NOSSE-NEXT:    movw %di, 10(%ebp)
+; NOSSE-NEXT:    movw %si, 8(%ebp)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 6(%ebp)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 4(%ebp)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 2(%ebp)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, (%ebp)
+; NOSSE-NEXT:    movl %ebp, %eax
+; NOSSE-NEXT:    addl $12, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 20
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    popl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    popl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl $4
+;
+; SSE-LABEL: return_arg_v8bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %ebp
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    pushl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    pushl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 20
+; SSE-NEXT:    subl $12, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 32
+; SSE-NEXT:    .cfi_offset %esi, -20
+; SSE-NEXT:    .cfi_offset %edi, -16
+; SSE-NEXT:    .cfi_offset %ebx, -12
+; SSE-NEXT:    .cfi_offset %ebp, -8
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %esi
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %edi
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %ebx
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %ebp
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, 14(%ebp)
+; SSE-NEXT:    movw %bx, 12(%ebp)
+; SSE-NEXT:    movw %di, 10(%ebp)
+; SSE-NEXT:    movw %si, 8(%ebp)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 6(%ebp)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 4(%ebp)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 2(%ebp)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, (%ebp)
+; SSE-NEXT:    movl %ebp, %eax
+; SSE-NEXT:    addl $12, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 20
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    popl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    popl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    popl %ebp
+; SSE-NEXT:    .cfi_def_cfa_offset 4
+; SSE-NEXT:    retl $4
+  ret <8 x bfloat> %x
+}
+
+define <16 x bfloat> @return_arg_v16bf16(<16 x bfloat> %x) {
+; NOSSE-LABEL: return_arg_v16bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    pushl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    pushl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 20
+; NOSSE-NEXT:    subl $28, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 48
+; NOSSE-NEXT:    .cfi_offset %esi, -20
+; NOSSE-NEXT:    .cfi_offset %edi, -16
+; NOSSE-NEXT:    .cfi_offset %ebx, -12
+; NOSSE-NEXT:    .cfi_offset %ebp, -8
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %esi
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %ebx
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %ebp
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, 30(%edi)
+; NOSSE-NEXT:    movw %bp, 28(%edi)
+; NOSSE-NEXT:    movw %bx, 26(%edi)
+; NOSSE-NEXT:    movw %si, 24(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 22(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 20(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 18(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 16(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 14(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 12(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 10(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 8(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 6(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 4(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, 2(%edi)
+; NOSSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; NOSSE-NEXT:    movw %ax, (%edi)
+; NOSSE-NEXT:    movl %edi, %eax
+; NOSSE-NEXT:    addl $28, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 20
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    popl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    popl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl $4
+;
+; SSE-LABEL: return_arg_v16bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %ebp
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    pushl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    pushl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 20
+; SSE-NEXT:    subl $28, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 48
+; SSE-NEXT:    .cfi_offset %esi, -20
+; SSE-NEXT:    .cfi_offset %edi, -16
+; SSE-NEXT:    .cfi_offset %ebx, -12
+; SSE-NEXT:    .cfi_offset %ebp, -8
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, {{[-0-9]+}}(%e{{[sb]}}p) # 2-byte Spill
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %esi
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %ebx
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %ebp
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, 30(%edi)
+; SSE-NEXT:    movw %bp, 28(%edi)
+; SSE-NEXT:    movw %bx, 26(%edi)
+; SSE-NEXT:    movw %si, 24(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 22(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 20(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 18(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 16(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 14(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 12(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 10(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 8(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 6(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 4(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, 2(%edi)
+; SSE-NEXT:    movzwl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 2-byte Folded Reload
+; SSE-NEXT:    movw %ax, (%edi)
+; SSE-NEXT:    movl %edi, %eax
+; SSE-NEXT:    addl $28, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 20
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    popl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    popl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    popl %ebp
+; SSE-NEXT:    .cfi_def_cfa_offset 4
+; SSE-NEXT:    retl $4
+  ret <16 x bfloat> %x
+}
+
+declare bfloat @returns_bf16(bfloat)
+declare <2 x bfloat> @returns_v2bf16(<2 x bfloat>)
+declare <3 x bfloat> @returns_v3bf16(<3 x bfloat>)
+declare <4 x bfloat> @returns_v4bf16(<4 x bfloat>)
+declare <8 x bfloat> @returns_v8bf16(<8 x bfloat>)
+declare <16 x bfloat> @returns_v16bf16(<16 x bfloat>)
+
+define void @call_ret_bf16(ptr %ptr) {
+; NOSSE-LABEL: call_ret_bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    subl $8, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    .cfi_offset %esi, -8
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; NOSSE-NEXT:    movzwl (%esi), %eax
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll returns_bf16 at PLT
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, (%esi)
+; NOSSE-NEXT:    addl $8, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl
+;
+; SSE-LABEL: call_ret_bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    subl $8, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    .cfi_offset %esi, -8
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; SSE-NEXT:    movzwl (%esi), %eax
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll returns_bf16 at PLT
+; SSE-NEXT:    fstps (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, (%esi)
+; SSE-NEXT:    addl $8, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 4
+; SSE-NEXT:    retl
+  %val = load bfloat, ptr %ptr
+  %bf16 = call bfloat @returns_bf16(bfloat %val)
+  store bfloat %bf16, ptr %ptr
+  ret void
+}
+
+define void @call_ret_v2bf16(ptr %ptr) {
+; NOSSE-LABEL: call_ret_v2bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    subl $20, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 32
+; NOSSE-NEXT:    .cfi_offset %esi, -12
+; NOSSE-NEXT:    .cfi_offset %edi, -8
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; NOSSE-NEXT:    movzwl 2(%edi), %eax
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl (%edi), %eax
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll returns_v2bf16 at PLT
+; NOSSE-NEXT:    fxch %st(1)
+; NOSSE-NEXT:    fstps {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movl %eax, %esi
+; NOSSE-NEXT:    flds {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Reload
+; NOSSE-NEXT:    fstps (%esp)
+; NOSSE-NEXT:    calll __truncsfbf2
+; NOSSE-NEXT:    movw %ax, 2(%edi)
+; NOSSE-NEXT:    movw %si, (%edi)
+; NOSSE-NEXT:    addl $20, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl
+;
+; SSE-LABEL: call_ret_v2bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    subl $36, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 48
+; SSE-NEXT:    .cfi_offset %esi, -12
+; SSE-NEXT:    .cfi_offset %edi, -8
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; SSE-NEXT:    movzwl 2(%edi), %eax
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl (%edi), %eax
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, (%esp)
+; SSE-NEXT:    calll returns_v2bf16 at PLT
+; SSE-NEXT:    fxch %st(1)
+; SSE-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) # 10-byte Folded Spill
+; SSE-NEXT:    fstps (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movl %eax, %esi
+; SSE-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) # 10-byte Folded Reload
+; SSE-NEXT:    fstps (%esp)
+; SSE-NEXT:    calll __truncsfbf2
+; SSE-NEXT:    movw %ax, 2(%edi)
+; SSE-NEXT:    movw %si, (%edi)
+; SSE-NEXT:    addl $36, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    popl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 4
+; SSE-NEXT:    retl
+  %val = load <2 x bfloat>, ptr %ptr
+  %bf16 = call <2 x bfloat> @returns_v2bf16(<2 x bfloat> %val)
+  store <2 x bfloat> %bf16, ptr %ptr
+  ret void
+}
+
+define void @call_ret_v3bf16(ptr %ptr) {
+; NOSSE-LABEL: call_ret_v3bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    subl $40, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 48
+; NOSSE-NEXT:    .cfi_offset %esi, -8
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; NOSSE-NEXT:    movl (%esi), %eax
+; NOSSE-NEXT:    movl 4(%esi), %ecx
+; NOSSE-NEXT:    leal {{[0-9]+}}(%esp), %edx
+; NOSSE-NEXT:    movl %edx, (%esp)
+; NOSSE-NEXT:    shll $16, %ecx
+; NOSSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl %eax, %ecx
+; NOSSE-NEXT:    andl $-65536, %ecx # imm = 0xFFFF0000
+; NOSSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    calll returns_v3bf16 at PLT
+; NOSSE-NEXT:    subl $4, %esp
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; NOSSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; NOSSE-NEXT:    movw %cx, 4(%esi)
+; NOSSE-NEXT:    movl %eax, (%esi)
+; NOSSE-NEXT:    addl $40, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl
+;
+; SSE-LABEL: call_ret_v3bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    subl $40, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 48
+; SSE-NEXT:    .cfi_offset %esi, -8
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; SSE-NEXT:    movl (%esi), %eax
+; SSE-NEXT:    movl 4(%esi), %ecx
+; SSE-NEXT:    leal {{[0-9]+}}(%esp), %edx
+; SSE-NEXT:    movl %edx, (%esp)
+; SSE-NEXT:    shll $16, %ecx
+; SSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl %eax, %ecx
+; SSE-NEXT:    andl $-65536, %ecx # imm = 0xFFFF0000
+; SSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    calll returns_v3bf16 at PLT
+; SSE-NEXT:    subl $4, %esp
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; SSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; SSE-NEXT:    movw %cx, 4(%esi)
+; SSE-NEXT:    movl %eax, (%esi)
+; SSE-NEXT:    addl $40, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 4
+; SSE-NEXT:    retl
+  %val = load <3 x bfloat>, ptr %ptr
+  %bf16 = call <3 x bfloat> @returns_v3bf16(<3 x bfloat> %val)
+  store <3 x bfloat> %bf16, ptr %ptr
+  ret void
+}
+
+define void @call_ret_v4bf16(ptr %ptr) {
+; NOSSE-LABEL: call_ret_v4bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    pushl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    subl $48, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 64
+; NOSSE-NEXT:    .cfi_offset %esi, -16
+; NOSSE-NEXT:    .cfi_offset %edi, -12
+; NOSSE-NEXT:    .cfi_offset %ebx, -8
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; NOSSE-NEXT:    movzwl 2(%esi), %ecx
+; NOSSE-NEXT:    movl (%esi), %eax
+; NOSSE-NEXT:    movl 4(%esi), %edx
+; NOSSE-NEXT:    movzwl 6(%esi), %edi
+; NOSSE-NEXT:    leal {{[0-9]+}}(%esp), %ebx
+; NOSSE-NEXT:    movl %ebx, (%esp)
+; NOSSE-NEXT:    shll $16, %edi
+; NOSSE-NEXT:    movl %edi, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %edx
+; NOSSE-NEXT:    movl %edx, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %ecx
+; NOSSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    calll returns_v4bf16 at PLT
+; NOSSE-NEXT:    subl $4, %esp
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; NOSSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; NOSSE-NEXT:    movzwl {{[0-9]+}}(%esp), %edx
+; NOSSE-NEXT:    movw %dx, 6(%esi)
+; NOSSE-NEXT:    movw %cx, 4(%esi)
+; NOSSE-NEXT:    movl %eax, (%esi)
+; NOSSE-NEXT:    addl $48, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    popl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl
+;
+; SSE-LABEL: call_ret_v4bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    pushl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    subl $48, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 64
+; SSE-NEXT:    .cfi_offset %esi, -16
+; SSE-NEXT:    .cfi_offset %edi, -12
+; SSE-NEXT:    .cfi_offset %ebx, -8
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; SSE-NEXT:    movzwl 2(%esi), %ecx
+; SSE-NEXT:    movl (%esi), %eax
+; SSE-NEXT:    movl 4(%esi), %edx
+; SSE-NEXT:    movzwl 6(%esi), %edi
+; SSE-NEXT:    leal {{[0-9]+}}(%esp), %ebx
+; SSE-NEXT:    movl %ebx, (%esp)
+; SSE-NEXT:    shll $16, %edi
+; SSE-NEXT:    movl %edi, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %edx
+; SSE-NEXT:    movl %edx, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %ecx
+; SSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    calll returns_v4bf16 at PLT
+; SSE-NEXT:    subl $4, %esp
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; SSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; SSE-NEXT:    movzwl {{[0-9]+}}(%esp), %edx
+; SSE-NEXT:    movw %dx, 6(%esi)
+; SSE-NEXT:    movw %cx, 4(%esi)
+; SSE-NEXT:    movl %eax, (%esi)
+; SSE-NEXT:    addl $48, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    popl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    popl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 4
+; SSE-NEXT:    retl
+  %val = load <4 x bfloat>, ptr %ptr
+  %bf16 = call <4 x bfloat> @returns_v4bf16(<4 x bfloat> %val)
+  store <4 x bfloat> %bf16, ptr %ptr
+  ret void
+}
+
+define void @call_ret_v8bf16(ptr %ptr) {
+; NOSSE-LABEL: call_ret_v8bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    pushl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    pushl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 20
+; NOSSE-NEXT:    subl $108, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 128
+; NOSSE-NEXT:    .cfi_offset %esi, -20
+; NOSSE-NEXT:    .cfi_offset %edi, -16
+; NOSSE-NEXT:    .cfi_offset %ebx, -12
+; NOSSE-NEXT:    .cfi_offset %ebp, -8
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; NOSSE-NEXT:    movzwl 2(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl (%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl 4(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movzwl 6(%esi), %edi
+; NOSSE-NEXT:    movl 8(%esi), %ebx
+; NOSSE-NEXT:    movzwl 10(%esi), %ebp
+; NOSSE-NEXT:    movl 12(%esi), %ecx
+; NOSSE-NEXT:    movzwl 14(%esi), %eax
+; NOSSE-NEXT:    leal {{[0-9]+}}(%esp), %edx
+; NOSSE-NEXT:    movl %edx, (%esp)
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %ecx
+; NOSSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %ebp
+; NOSSE-NEXT:    movl %ebp, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %ebx
+; NOSSE-NEXT:    movl %ebx, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %edi
+; NOSSE-NEXT:    movl %edi, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    calll returns_v8bf16 at PLT
+; NOSSE-NEXT:    subl $4, %esp
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; NOSSE-NEXT:    movzwl {{[0-9]+}}(%esp), %edx
+; NOSSE-NEXT:    movzwl {{[0-9]+}}(%esp), %edi
+; NOSSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ebx
+; NOSSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ebp
+; NOSSE-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
+; NOSSE-NEXT:    movw %ax, 14(%esi)
+; NOSSE-NEXT:    movw %bp, 12(%esi)
+; NOSSE-NEXT:    movw %bx, 10(%esi)
+; NOSSE-NEXT:    movw %di, 8(%esi)
+; NOSSE-NEXT:    movw %dx, 6(%esi)
+; NOSSE-NEXT:    movw %cx, 4(%esi)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    movl %eax, (%esi)
+; NOSSE-NEXT:    addl $108, %esp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 20
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 16
+; NOSSE-NEXT:    popl %edi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    popl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl
+;
+; SSE-LABEL: call_ret_v8bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %ebp
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    pushl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    pushl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 20
+; SSE-NEXT:    subl $108, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 128
+; SSE-NEXT:    .cfi_offset %esi, -20
+; SSE-NEXT:    .cfi_offset %edi, -16
+; SSE-NEXT:    .cfi_offset %ebx, -12
+; SSE-NEXT:    .cfi_offset %ebp, -8
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; SSE-NEXT:    movzwl 2(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl (%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl 4(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movzwl 6(%esi), %edi
+; SSE-NEXT:    movl 8(%esi), %ebx
+; SSE-NEXT:    movzwl 10(%esi), %ebp
+; SSE-NEXT:    movl 12(%esi), %ecx
+; SSE-NEXT:    movzwl 14(%esi), %eax
+; SSE-NEXT:    leal {{[0-9]+}}(%esp), %edx
+; SSE-NEXT:    movl %edx, (%esp)
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %ecx
+; SSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %ebp
+; SSE-NEXT:    movl %ebp, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %ebx
+; SSE-NEXT:    movl %ebx, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %edi
+; SSE-NEXT:    movl %edi, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    calll returns_v8bf16 at PLT
+; SSE-NEXT:    subl $4, %esp
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; SSE-NEXT:    movzwl {{[0-9]+}}(%esp), %edx
+; SSE-NEXT:    movzwl {{[0-9]+}}(%esp), %edi
+; SSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ebx
+; SSE-NEXT:    movzwl {{[0-9]+}}(%esp), %ebp
+; SSE-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
+; SSE-NEXT:    movw %ax, 14(%esi)
+; SSE-NEXT:    movw %bp, 12(%esi)
+; SSE-NEXT:    movw %bx, 10(%esi)
+; SSE-NEXT:    movw %di, 8(%esi)
+; SSE-NEXT:    movw %dx, 6(%esi)
+; SSE-NEXT:    movw %cx, 4(%esi)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    movl %eax, (%esi)
+; SSE-NEXT:    addl $108, %esp
+; SSE-NEXT:    .cfi_def_cfa_offset 20
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    .cfi_def_cfa_offset 16
+; SSE-NEXT:    popl %edi
+; SSE-NEXT:    .cfi_def_cfa_offset 12
+; SSE-NEXT:    popl %ebx
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    popl %ebp
+; SSE-NEXT:    .cfi_def_cfa_offset 4
+; SSE-NEXT:    retl
+  %val = load <8 x bfloat>, ptr %ptr
+  %bf16 = call <8 x bfloat> @returns_v8bf16(<8 x bfloat> %val)
+  store <8 x bfloat> %bf16, ptr %ptr
+  ret void
+}
+
+define void @call_ret_v16bf16(ptr %ptr) {
+; NOSSE-LABEL: call_ret_v16bf16:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    .cfi_offset %ebp, -8
+; NOSSE-NEXT:    movl %esp, %ebp
+; NOSSE-NEXT:    .cfi_def_cfa_register %ebp
+; NOSSE-NEXT:    pushl %ebx
+; NOSSE-NEXT:    pushl %edi
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    andl $-32, %esp
+; NOSSE-NEXT:    subl $256, %esp # imm = 0x100
+; NOSSE-NEXT:    .cfi_offset %esi, -20
+; NOSSE-NEXT:    .cfi_offset %edi, -16
+; NOSSE-NEXT:    .cfi_offset %ebx, -12
+; NOSSE-NEXT:    movl 8(%ebp), %esi
+; NOSSE-NEXT:    movzwl 2(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl (%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl 4(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movzwl 6(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl 8(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movzwl 10(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl 12(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movzwl 14(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl 16(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movzwl 18(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl 20(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movzwl 22(%esi), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl 24(%esi), %edi
+; NOSSE-NEXT:    movzwl 26(%esi), %edx
+; NOSSE-NEXT:    movl 28(%esi), %ecx
+; NOSSE-NEXT:    movzwl 30(%esi), %eax
+; NOSSE-NEXT:    leal {{[0-9]+}}(%esp), %ebx
+; NOSSE-NEXT:    movl %ebx, (%esp)
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %ecx
+; NOSSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %edx
+; NOSSE-NEXT:    movl %edx, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    shll $16, %edi
+; NOSSE-NEXT:    movl %edi, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    shll $16, %eax
+; NOSSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    flds {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; NOSSE-NEXT:    calll returns_v16bf16 at PLT
+; NOSSE-NEXT:    subl $4, %esp
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; NOSSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %ebx
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; NOSSE-NEXT:    movl %edx, 28(%esi)
+; NOSSE-NEXT:    movl %eax, 24(%esi)
+; NOSSE-NEXT:    movl %ecx, 20(%esi)
+; NOSSE-NEXT:    movl %ebx, 16(%esi)
+; NOSSE-NEXT:    movl %edi, 12(%esi)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    movl %eax, 8(%esi)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    movl %eax, 4(%esi)
+; NOSSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; NOSSE-NEXT:    movl %eax, (%esi)
+; NOSSE-NEXT:    leal -12(%ebp), %esp
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    popl %edi
+; NOSSE-NEXT:    popl %ebx
+; NOSSE-NEXT:    popl %ebp
+; NOSSE-NEXT:    .cfi_def_cfa %esp, 4
+; NOSSE-NEXT:    retl
+;
+; SSE-LABEL: call_ret_v16bf16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushl %ebp
+; SSE-NEXT:    .cfi_def_cfa_offset 8
+; SSE-NEXT:    .cfi_offset %ebp, -8
+; SSE-NEXT:    movl %esp, %ebp
+; SSE-NEXT:    .cfi_def_cfa_register %ebp
+; SSE-NEXT:    pushl %ebx
+; SSE-NEXT:    pushl %edi
+; SSE-NEXT:    pushl %esi
+; SSE-NEXT:    andl $-32, %esp
+; SSE-NEXT:    subl $256, %esp # imm = 0x100
+; SSE-NEXT:    .cfi_offset %esi, -20
+; SSE-NEXT:    .cfi_offset %edi, -16
+; SSE-NEXT:    .cfi_offset %ebx, -12
+; SSE-NEXT:    movl 8(%ebp), %esi
+; SSE-NEXT:    movzwl 2(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl (%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl 4(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movzwl 6(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl 8(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movzwl 10(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl 12(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movzwl 14(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl 16(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movzwl 18(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl 20(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movzwl 22(%esi), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl 24(%esi), %edi
+; SSE-NEXT:    movzwl 26(%esi), %edx
+; SSE-NEXT:    movl 28(%esi), %ecx
+; SSE-NEXT:    movzwl 30(%esi), %eax
+; SSE-NEXT:    leal {{[0-9]+}}(%esp), %ebx
+; SSE-NEXT:    movl %ebx, (%esp)
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %ecx
+; SSE-NEXT:    movl %ecx, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %edx
+; SSE-NEXT:    movl %edx, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    shll $16, %edi
+; SSE-NEXT:    movl %edi, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    shll $16, %eax
+; SSE-NEXT:    movl %eax, {{[0-9]+}}(%esp)
+; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; SSE-NEXT:    calll returns_v16bf16 at PLT
+; SSE-NEXT:    subl $4, %esp
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; SSE-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %ebx
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; SSE-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; SSE-NEXT:    movl %edx, 28(%esi)
+; SSE-NEXT:    movl %eax, 24(%esi)
+; SSE-NEXT:    movl %ecx, 20(%esi)
+; SSE-NEXT:    movl %ebx, 16(%esi)
+; SSE-NEXT:    movl %edi, 12(%esi)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    movl %eax, 8(%esi)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    movl %eax, 4(%esi)
+; SSE-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; SSE-NEXT:    movl %eax, (%esi)
+; SSE-NEXT:    leal -12(%ebp), %esp
+; SSE-NEXT:    popl %esi
+; SSE-NEXT:    popl %edi
+; SSE-NEXT:    popl %ebx
+; SSE-NEXT:    popl %ebp
+; SSE-NEXT:    .cfi_def_cfa %esp, 4
+; SSE-NEXT:    retl
+  %val = load <16 x bfloat>, ptr %ptr
+  %bf16 = call <16 x bfloat> @returns_v16bf16(<16 x bfloat> %val)
+  store <16 x bfloat> %bf16, ptr %ptr
+  ret void
+}



More information about the llvm-commits mailing list