Cute Clang bug.

sse2_mul_shift.cpp

#include <emmintrin.h>

static const __m128i magic = _mm_setr_epi16(1 << 14, 1 << 12, 1 << 10, 1 << 8, 1 << 6, 1 << 4, 1 << 2, 1 << 0);

__m128i good_unpack_2bits_to_16(unsigned short x)
{
    __m128i shifted = _mm_mullo_epi16(_mm_set1_epi16(x), magic);
    return _mm_srli_epi16(shifted, 14);
}

__m128i bad_unpack_2bits_to_16(unsigned short x)
{
    __m128i magic = _mm_setr_epi16(1 << 14, 1 << 12, 1 << 10, 1 << 8, 1 << 6, 1 << 4, 1 << 2, 1 << 0);
    __m128i shifted = _mm_mullo_epi16(_mm_set1_epi16(x), magic);
    return _mm_srli_epi16(shifted, 14);
}

/*

C:\devel\llvm\repro>..\bin\clang++ --version
clang version 3.4 (198054)
Target: i686-pc-mingw32
Thread model: posix
   
C:\devel\llvm\repro>..\bin\clang++ -O2 -S sse2_mul_shift.cpp

	.def	 @feat.00;
	.scl	3;
	.type	0;
	.endef
	.globl	@feat.00
@feat.00 = 1
	.def	 __Z23good_unpack_2bits_to_16t;
	.scl	2;
	.type	32;
	.endef
	.text
	.globl	__Z23good_unpack_2bits_to_16t
	.align	16, 0x90
__Z23good_unpack_2bits_to_16t:
	pushl	%ebp
	movl	%esp, %ebp
	movd	8(%ebp), %xmm0
	punpcklwd	%xmm0, %xmm0
	pshufd	$0, %xmm0, %xmm0
	pmullw	__ZL5magic, %xmm0
	psrlw	$14, %xmm0
	popl	%ebp
	ret

	.def	 __Z22bad_unpack_2bits_to_16t;
	.scl	2;
	.type	32;
	.endef
	.globl	__Z22bad_unpack_2bits_to_16t
	.align	16, 0x90
__Z22bad_unpack_2bits_to_16t:
	pushl	%ebp
	movl	%esp, %ebp
	movzwl	8(%ebp), %eax
	movd	%eax, %xmm0
	movl	%eax, %ecx
	shll	$8, %ecx
	movd	%ecx, %xmm1
	punpcklwd	%xmm0, %xmm1
	movl	%eax, %ecx
	shll	$4, %ecx
	movd	%ecx, %xmm0
	movl	%eax, %ecx
	shll	$12, %ecx
	movd	%ecx, %xmm2
	punpcklwd	%xmm0, %xmm2
	punpcklwd	%xmm1, %xmm2
	leal	(,%eax,4), %ecx
	movd	%ecx, %xmm0
	movl	%eax, %ecx
	shll	$10, %ecx
	movd	%ecx, %xmm1
	punpcklwd	%xmm0, %xmm1
	movl	%eax, %ecx
	shll	$6, %ecx
	movd	%ecx, %xmm3
	shll	$14, %eax
	movd	%eax, %xmm0
	punpcklwd	%xmm3, %xmm0
	punpcklwd	%xmm1, %xmm0
	punpcklwd	%xmm2, %xmm0
	psrlw	$14, %xmm0
	popl	%ebp
	ret

	.def	 __GLOBAL__I_a;
	.scl	3;
	.type	32;
	.endef
	.section	.rdata,"r"
	.align	16
LCPI2_0:
	.long	268451840
	.long	16778240
	.long	1048640
	.long	65540
	.text
	.align	16, 0x90
__GLOBAL__I_a:
	pushl	%ebp
	movl	%esp, %ebp
	movaps	LCPI2_0, %xmm0
	movaps	%xmm0, __ZL5magic
	popl	%ebp
	ret

	.lcomm	__ZL5magic,16,16
	.section	.ctors,"w"
	.align	4
	.long	__GLOBAL__I_a
*/