[X86] Possible missed optimization with truncating vector cast

[okaneco]

This function produces the following assembly
https://llvm.godbolt.org/z/q8YEhevbo

define void @cast_i16x4_to_u8x4(ptr sret(<4 x i8>) %_0, ptr %x) unnamed_addr #0 {
  %1 = load <4 x i16>, ptr %x
  %2 = trunc <4 x i16> %1 to <4 x i8>
  store <4 x i8> %2, ptr %_0
  ret void
}

cast_i16x4_to_u8x4:                     # @cast_i16x4_to_u8x4
        mov     rax, rdi
        mov     rcx, qword ptr [rsi]
        movq    xmm0, rcx
        movdqa  xmmword ptr [rsp - 24], xmm0
        movzx   ecx, cl
        movzx   edx, byte ptr [rsp - 22]
        shl     edx, 8
        or      edx, ecx
        movzx   ecx, byte ptr [rsp - 20]
        shl     ecx, 16
        or      ecx, edx
        movzx   edx, byte ptr [rsp - 18]
        shl     edx, 24
        or      edx, ecx
        mov     dword ptr [rdi], edx
        ret

but I expected something like this instead.

.LCPI0_0:
        .short  255                             # 0xff
        .short  255                             # 0xff
        .short  255                             # 0xff
        .short  255                             # 0xff
        .short  255                             # 0xff
        .short  255                             # 0xff
        .short  255                             # 0xff
        .short  255                             # 0xff
cast_i16x4_to_u8x4:                     # @cast_i16x4_to_u8x4
        mov     rax, rdi
        movq    xmm0, qword ptr [rsi]
        pand    xmm0, xmmword ptr [rip + .LCPI0_0]
        packuswb        xmm0, xmm0
        movd    dword ptr [rdi], xmm0
        ret

---

Original Rust code https://rust.godbolt.org/z/WrKdjx9sa
Originally found in Rust portable-simd code for casting an i16x4 to u8x4 rust-lang/portable-simd#369 (comment)

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Collyn O'Kane (okaneco)

This function produces the following assembly https://llvm.godbolt.org/z/q8YEhevbo ```llvm define void @cast_i16x4_to_u8x4(ptr sret(<4 x i8>) %_0, ptr %x) unnamed_addr #0 { %1 = load <4 x i16>, ptr %x %2 = trunc <4 x i16> %1 to <4 x i8> store <4 x i8> %2, ptr %_0 ret void } ``` ```asm cast_i16x4_to_u8x4: # @cast_i16x4_to_u8x4 mov rax, rdi mov rcx, qword ptr [rsi] movq xmm0, rcx movdqa xmmword ptr [rsp - 24], xmm0 movzx ecx, cl movzx edx, byte ptr [rsp - 22] shl edx, 8 or edx, ecx movzx ecx, byte ptr [rsp - 20] shl ecx, 16 or ecx, edx movzx edx, byte ptr [rsp - 18] shl edx, 24 or edx, ecx mov dword ptr [rdi], edx ret ``` but I expected something like this instead. ```asm .LCPI0_0: .short 255 # 0xff .short 255 # 0xff .short 255 # 0xff .short 255 # 0xff .short 255 # 0xff .short 255 # 0xff .short 255 # 0xff .short 255 # 0xff cast_i16x4_to_u8x4: # @cast_i16x4_to_u8x4 mov rax, rdi movq xmm0, qword ptr [rsi] pand xmm0, xmmword ptr [rip + .LCPI0_0] packuswb xmm0, xmm0 movd dword ptr [rdi], xmm0 ret ```

---

Original Rust code https://rust.godbolt.org/z/WrKdjx9sa
Originally found in Rust portable-simd code for casting an i16x4 to u8x4 rust-lang/portable-simd#369 (comment)

[RKSimon]

Even x86-64-v4 scalarizes this, despite it being able to have done a simple zextload/truncstore combo