[X86][FP16] Do not generate VBROADCAST for fp16

[pengfei]

Summary

We cannot lower VBROADCAST i16 under AVX1.

Fixes #63114

[LuoYuanke]

llvm/test/CodeGen/X86/vector-half-conversions.ll

942

vbroadcastss would access 32-bit memory and vpinsrw only access 16-bit memory?

[pengfei]

llvm/test/CodeGen/X86/vector-half-conversions.ll

942

The vbroadcastss is an optimization because the high 16-bit is undefined.
ptr %a0 is a <8 x i16> pointer, so we are free to access 32-bit memory without worrying about the size.

[LuoYuanke]

LGTM.

llvm/lib/Target/X86/X86ISelLowering.cpp

 ///
 /// For convenience, this code also bundles all of the subtarget feature set
 /// filtering. While a little annoying to re-dispatch on type here, there isn't
 /// a convenient way to factor it out.
 static SDValue lowerShuffleAsBroadcast(const SDLoc &DL, MVT VT, SDValue V1,
                                        SDValue V2, ArrayRef<int> Mask,
                                        const X86Subtarget &Subtarget,
                                        SelectionDAG &DAG) {
+  MVT EltVT = VT.getVectorElementType();
   if (!((Subtarget.hasSSE3() && VT == MVT::v2f64) ||
-        (Subtarget.hasAVX() && VT.isFloatingPoint()) ||
-        (Subtarget.hasAVX2() && VT.isInteger())))
+        (Subtarget.hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
+        (Subtarget.hasAVX2() && (VT.isInteger() || EltVT == MVT::f16))))
     return SDValue();
 
   // With MOVDDUP (v2f64) we can broadcast from a register or a load, otherwise
   // we can only broadcast from a register with AVX2.
   unsigned NumEltBits = VT.getScalarSizeInBits();
   unsigned Opcode = (VT == MVT::v2f64 && !Subtarget.hasAVX2())
                         ? X86ISD::MOVDDUP
                         : X86ISD::VBROADCAST;

llvm/test/CodeGen/X86/half.ll

 ; CHECK-I686-NEXT:    popl %esi
 ; CHECK-I686-NEXT:    popl %edi
 ; CHECK-I686-NEXT:    popl %ebx
 ; CHECK-I686-NEXT:    retl
   %3 = call fast <8 x half> @llvm.maxnum.v8f16(<8 x half> %0, <8 x half> %1)
   ret <8 x half> %3
 }
 
+define void @pr63114() {
+; CHECK-LIBCALL-LABEL: pr63114:
+; CHECK-LIBCALL:       # %bb.0:
+; CHECK-LIBCALL-NEXT:    movdqu (%rax), %xmm4
+; CHECK-LIBCALL-NEXT:    pshuflw {{.*#+}} xmm0 = xmm4[0,1,3,3,4,5,6,7]
+; CHECK-LIBCALL-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
+; CHECK-LIBCALL-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,0,65535,65535,65535,65535]
+; CHECK-LIBCALL-NEXT:    pand %xmm1, %xmm0
+; CHECK-LIBCALL-NEXT:    movdqa {{.*#+}} xmm2 = [0,0,0,15360,0,0,0,0]
+; CHECK-LIBCALL-NEXT:    por %xmm2, %xmm0
+; CHECK-LIBCALL-NEXT:    movdqa {{.*#+}} xmm3 = [65535,65535,65535,65535,65535,65535,65535,0]
+; CHECK-LIBCALL-NEXT:    pand %xmm3, %xmm0
+; CHECK-LIBCALL-NEXT:    movdqa {{.*#+}} xmm5 = [0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,60]
+; CHECK-LIBCALL-NEXT:    por %xmm5, %xmm0
+; CHECK-LIBCALL-NEXT:    pshufhw {{.*#+}} xmm6 = xmm4[0,1,2,3,4,5,7,7]
+; CHECK-LIBCALL-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
+; CHECK-LIBCALL-NEXT:    pand %xmm1, %xmm6
+; CHECK-LIBCALL-NEXT:    por %xmm2, %xmm6
+; CHECK-LIBCALL-NEXT:    pand %xmm3, %xmm6
+; CHECK-LIBCALL-NEXT:    por %xmm5, %xmm6
+; CHECK-LIBCALL-NEXT:    pshufhw {{.*#+}} xmm7 = xmm4[0,1,2,3,5,5,5,5]
+; CHECK-LIBCALL-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,3,0,3]
+; CHECK-LIBCALL-NEXT:    pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,5,5,5]
+; CHECK-LIBCALL-NEXT:    pand %xmm1, %xmm4
+; CHECK-LIBCALL-NEXT:    por %xmm2, %xmm4
+; CHECK-LIBCALL-NEXT:    pand %xmm3, %xmm4
+; CHECK-LIBCALL-NEXT:    por %xmm5, %xmm4
+; CHECK-LIBCALL-NEXT:    pand %xmm1, %xmm7
+; CHECK-LIBCALL-NEXT:    por %xmm2, %xmm7
+; CHECK-LIBCALL-NEXT:    pand %xmm3, %xmm7
+; CHECK-LIBCALL-NEXT:    por %xmm5, %xmm7
+; CHECK-LIBCALL-NEXT:    movdqu %xmm7, 0
+; CHECK-LIBCALL-NEXT:    movdqu %xmm4, 32
+; CHECK-LIBCALL-NEXT:    movdqu %xmm6, 48
+; CHECK-LIBCALL-NEXT:    movdqu %xmm0, 16
+; CHECK-LIBCALL-NEXT:    retq
+;
+; BWON-F16C-LABEL: pr63114:
+; BWON-F16C:       # %bb.0:
+; BWON-F16C-NEXT:    vmovdqu (%rax), %xmm0
+; BWON-F16C-NEXT:    vpsrld $16, %xmm0, %xmm1
+; BWON-F16C-NEXT:    vbroadcastss (%rax), %xmm2
+; BWON-F16C-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; BWON-F16C-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[0,0],xmm3[0,0]
+; BWON-F16C-NEXT:    vpinsrw $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm3
+; BWON-F16C-NEXT:    vpsllq $48, %xmm3, %xmm4
+; BWON-F16C-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1,2],xmm4[3],xmm2[4,5,6,7]
+; BWON-F16C-NEXT:    vpslldq {{.*#+}} xmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1]
+; BWON-F16C-NEXT:    vpor %xmm3, %xmm2, %xmm2
+; BWON-F16C-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[0,3],xmm1[2,0]
+; BWON-F16C-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm4[3],xmm1[4,5,6,7]
+; BWON-F16C-NEXT:    vpor %xmm3, %xmm1, %xmm1
+; BWON-F16C-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
+; BWON-F16C-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm0[0,1,3,3,4,5,6,7]
+; BWON-F16C-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,0,2,1]
+; BWON-F16C-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1,2],xmm4[3],xmm2[4,5,6,7]
+; BWON-F16C-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,6],xmm3[7]
+; BWON-F16C-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
+; BWON-F16C-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm4[3],xmm0[4,5,6,7]
+; BWON-F16C-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,6],xmm3[7]
+; BWON-F16C-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; BWON-F16C-NEXT:    vmovups %ymm0, 0
+; BWON-F16C-NEXT:    vmovups %ymm1, 32
+; BWON-F16C-NEXT:    vzeroupper
+; BWON-F16C-NEXT:    retq
+;
+; CHECK-I686-LABEL: pr63114:
+; CHECK-I686:       # %bb.0:
+; CHECK-I686-NEXT:    movdqu (%eax), %xmm6
+; CHECK-I686-NEXT:    pshuflw {{.*#+}} xmm0 = xmm6[0,1,3,3,4,5,6,7]
+; CHECK-I686-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
+; CHECK-I686-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,0,65535,65535,65535,65535]
+; CHECK-I686-NEXT:    pand %xmm1, %xmm0
+; CHECK-I686-NEXT:    movdqa {{.*#+}} xmm2 = [0,0,0,15360,0,0,0,0]
+; CHECK-I686-NEXT:    por %xmm2, %xmm0
+; CHECK-I686-NEXT:    movdqa {{.*#+}} xmm3 = [65535,65535,65535,65535,65535,65535,65535,0]
+; CHECK-I686-NEXT:    pand %xmm3, %xmm0
+; CHECK-I686-NEXT:    movdqa {{.*#+}} xmm4 = [0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,60]
+; CHECK-I686-NEXT:    por %xmm4, %xmm0
+; CHECK-I686-NEXT:    pshufhw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,5,7,7]
+; CHECK-I686-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
+; CHECK-I686-NEXT:    pand %xmm1, %xmm5
+; CHECK-I686-NEXT:    por %xmm2, %xmm5
+; CHECK-I686-NEXT:    pand %xmm3, %xmm5
+; CHECK-I686-NEXT:    por %xmm4, %xmm5
+; CHECK-I686-NEXT:    pshufhw {{.*#+}} xmm7 = xmm6[0,1,2,3,5,5,5,5]
+; CHECK-I686-NEXT:    shufps {{.*#+}} xmm6 = xmm6[0,3,0,3]
+; CHECK-I686-NEXT:    pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5]
+; CHECK-I686-NEXT:    pand %xmm1, %xmm6
+; CHECK-I686-NEXT:    por %xmm2, %xmm6
+; CHECK-I686-NEXT:    pand %xmm3, %xmm6
+; CHECK-I686-NEXT:    por %xmm4, %xmm6
+; CHECK-I686-NEXT:    pand %xmm1, %xmm7
+; CHECK-I686-NEXT:    por %xmm2, %xmm7
+; CHECK-I686-NEXT:    pand %xmm3, %xmm7
+; CHECK-I686-NEXT:    por %xmm4, %xmm7
+; CHECK-I686-NEXT:    movdqu %xmm7, 0
+; CHECK-I686-NEXT:    movdqu %xmm6, 32
+; CHECK-I686-NEXT:    movdqu %xmm5, 48
+; CHECK-I686-NEXT:    movdqu %xmm0, 16
+; CHECK-I686-NEXT:    retl
+  %1 = load <24 x half>, ptr poison, align 2
+  %2 = shufflevector <24 x half> %1, <24 x half> poison, <8 x i32> <i32 2, i32 5, i32 8, i32 11, i32 14, i32 17, i32 20, i32 23>
+  %3 = shufflevector <8 x half> %2, <8 x half> <half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00>, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %4 = shufflevector <16 x half> poison, <16 x half> %3, <32 x i32> <i32 0, i32 8, i32 16, i32 24, i32 1, i32 9, i32 17, i32 25, i32 2, i32 10, i32 18, i32 26, i32 3, i32 11, i32 19, i32 27, i32 4, i32 12, i32 20, i32 28, i32 5, i32 13, i32 21, i32 29, i32 6, i32 14, i32 22, i32 30, i32 7, i32 15, i32 23, i32 31>
+  store <32 x half> %4, ptr null, align 2
+  ret void
+}
+
 attributes #0 = { nounwind }

llvm/test/CodeGen/X86/vector-half-conversions.ll

; AVX512-NEXT:    retq
   %1 = load <8 x i16>, ptr %a0
   %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %3 = bitcast <4 x i16> %2 to <4 x half>
   %4 = fpext <4 x half> %3 to <4 x float>
   ret <4 x float> %4
 }
 
 define <8 x float> @load_cvt_8i16_to_8f32(ptr %a0) nounwind {
-; AVX-LABEL: load_cvt_8i16_to_8f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    pushq %rbx
-; AVX-NEXT:    subq $48, %rsp
-; AVX-NEXT:    movq %rdi, %rbx
-; AVX-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vpinsrw $0, 8(%rdi), %xmm0, %xmm0
-; AVX-NEXT:    callq __extendhfsf2@PLT
-; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX-NEXT:    callq __extendhfsf2@PLT
-; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vpinsrw $0, 12(%rbx), %xmm0, %xmm0
-; AVX-NEXT:    callq __extendhfsf2@PLT
-; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX-NEXT:    callq __extendhfsf2@PLT
-; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT:    callq __extendhfsf2@PLT
-; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT:    vpsrld $16, %xmm0, %xmm0
-; AVX-NEXT:    callq __extendhfsf2@PLT
-; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
-; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vpinsrw $0, 4(%rbx), %xmm0, %xmm0
-; AVX-NEXT:    callq __extendhfsf2@PLT
-; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
-; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX-NEXT:    vpsrlq $48, %xmm0, %xmm0
-; AVX-NEXT:    callq __extendhfsf2@PLT
-; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
-; AVX-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
-; AVX-NEXT:    addq $48, %rsp
-; AVX-NEXT:    popq %rbx
-; AVX-NEXT:    retq
+; AVX1-LABEL: load_cvt_8i16_to_8f32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    pushq %rbx
+; AVX1-NEXT:    subq $48, %rsp
+; AVX1-NEXT:    movq %rdi, %rbx
+; AVX1-NEXT:    vmovaps (%rdi), %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vbroadcastss 8(%rdi), %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vbroadcastss 12(%rbx), %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vbroadcastss 4(%rbx), %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX1-NEXT:    addq $48, %rsp
+; AVX1-NEXT:    popq %rbx
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: load_cvt_8i16_to_8f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    pushq %rbx
+; AVX2-NEXT:    subq $48, %rsp
+; AVX2-NEXT:    movq %rdi, %rbx
+; AVX2-NEXT:    vmovdqa (%rdi), %xmm0
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 8(%rdi), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 12(%rbx), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 4(%rbx), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX2-NEXT:    addq $48, %rsp
+; AVX2-NEXT:    popq %rbx
+; AVX2-NEXT:    retq
 ;
 ; F16C-LABEL: load_cvt_8i16_to_8f32:
 ; F16C:       # %bb.0:
 ; F16C-NEXT:    vcvtph2ps (%rdi), %ymm0
 ; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: load_cvt_8i16_to_8f32:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtph2ps (%rdi), %ymm0
 ; AVX512-NEXT:    retq
   %1 = load <8 x i16>, ptr %a0
   %2 = bitcast <8 x i16> %1 to <8 x half>
   %3 = fpext <8 x half> %2 to <8 x float>
   ret <8 x float> %3
 }
 
 define <16 x float> @load_cvt_16i16_to_16f32(ptr %a0) nounwind {
 ; AVX1-LABEL: load_cvt_16i16_to_16f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    pushq %rbx
 ; AVX1-NEXT:    subq $80, %rsp
 ; AVX1-NEXT:    movq %rdi, %rbx
-; AVX1-NEXT:    vpinsrw $0, 8(%rdi), %xmm0, %xmm0
+; AVX1-NEXT:    vbroadcastss 8(%rdi), %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
-; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa (%rbx), %xmm1
 ; AVX1-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX1-NEXT:    vmovaps 16(%rbx), %xmm0
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
 ; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX1-NEXT:    vpinsrw $0, 12(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    vbroadcastss 12(%rbx), %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
 ; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
 ; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT:    vpinsrw $0, 4(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    vbroadcastss 4(%rbx), %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
 ; AVX1-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
-; AVX1-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX1-NEXT:    vpinsrw $0, 24(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vbroadcastss 24(%rbx), %xmm0
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
-; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
 ; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX1-NEXT:    vpinsrw $0, 28(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    vbroadcastss 28(%rbx), %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
 ; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
 ; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT:    vpinsrw $0, 20(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    vbroadcastss 20(%rbx), %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload

 }
 
 define <8 x double> @load_cvt_8i16_to_8f64(ptr %a0) nounwind {
 ; AVX1-LABEL: load_cvt_8i16_to_8f64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    pushq %rbx
 ; AVX1-NEXT:    subq $80, %rsp
 ; AVX1-NEXT:    movq %rdi, %rbx
-; AVX1-NEXT:    vmovdqa (%rdi), %xmm0
-; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX1-NEXT:    vpinsrw $0, 4(%rdi), %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps (%rdi), %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vbroadcastss 4(%rdi), %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
 ; AVX1-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
-; AVX1-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX1-NEXT:    vpinsrw $0, 12(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vbroadcastss 12(%rbx), %xmm0
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
 ; AVX1-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
 ; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX1-NEXT:    vpinsrw $0, 8(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    vbroadcastss 8(%rbx), %xmm0
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX1-NEXT:    callq __extendhfsf2@PLT
 ; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload