Attachment 198890 Details for Bug 232911 – c++ testcase

c++ testcase

clang-test.cpp (text/plain), 3.97 KB, created by Yuri Victorovich on 2018-11-02 16:15:34 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Yuri Victorovich

Created: 2018-11-02 16:15:34 UTC

Size: 3.97 KB

patch

obsolete

>#include <sys/types.h>
>#include <stdint.h>
>
>#define ARCH_X86_ASM(...)       __asm__ __volatile__ ( __VA_ARGS__ )
>#define __ASM_EMIT(code)                        code "\n\t"
>
>namespace sse
>{
>    void fill(float *dst, float value, size_t count)
>    {
>        #define FILL_CORE   \
>            /* Blocks of 32 items */ \
>            __ASM_EMIT("sub         $0x20, %[count]")   \
>            __ASM_EMIT("jb          4f")   \
>            __ASM_EMIT("movaps      %%xmm0, %%xmm4") \
>            __ASM_EMIT("movaps      %%xmm1, %%xmm5") \
>            __ASM_EMIT("movaps      %%xmm2, %%xmm6") \
>            __ASM_EMIT("movaps      %%xmm3, %%xmm7") \
>            __ASM_EMIT("3:") \
>            __ASM_EMIT("movaps      %%xmm0, 0x00(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm1, 0x10(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm2, 0x20(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm3, 0x30(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm4, 0x40(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm5, 0x50(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm6, 0x60(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm7, 0x70(%[dst])") \
>            __ASM_EMIT("add         $0x80, %[dst]") \
>            __ASM_EMIT("sub         $0x20, %[count]") \
>            __ASM_EMIT("jae         3b")   \
>            /* Blocks of 16 items */ \
>            __ASM_EMIT("4:") \
>            __ASM_EMIT("add         $0x20, %[count]") \
>            __ASM_EMIT("test        $0x10, %[count]") \
>            __ASM_EMIT("jz          5f") \
>            __ASM_EMIT("movaps      %%xmm0, 0x00(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm1, 0x10(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm2, 0x20(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm3, 0x30(%[dst])") \
>            __ASM_EMIT("add         $0x40, %[dst]") \
>            /* Blocks of 8 items */ \
>            __ASM_EMIT("5:") \
>            __ASM_EMIT("test        $0x08, %[count]") \
>            __ASM_EMIT("jz          6f") \
>            __ASM_EMIT("movaps      %%xmm0, 0x00(%[dst])") \
>            __ASM_EMIT("movaps      %%xmm1, 0x10(%[dst])") \
>            __ASM_EMIT("add         $0x20, %[dst]") \
>            /* Blocks of 4 items */ \
>            __ASM_EMIT("6:") \
>            __ASM_EMIT("test        $0x04, %[count]") \
>            __ASM_EMIT("jz          7f") \
>            __ASM_EMIT("movaps      %%xmm0, 0x00(%[dst])") \
>            __ASM_EMIT("add         $0x10, %[dst]") \
>            /* Tail */ \
>            __ASM_EMIT("7:") \
>            __ASM_EMIT("and         $0x03, %[count]") \
>            __ASM_EMIT("jz          2000f") \
>            __ASM_EMIT("8:") \
>            __ASM_EMIT("movss       %%xmm0, 0x00(%[dst])") \
>            __ASM_EMIT("add         $0x04, %[dst]") \
>            __ASM_EMIT("dec         %[count]") \
>            __ASM_EMIT("jnz         8b")
>
>        ARCH_X86_ASM
>        (
>            __ASM_EMIT("test        %[count], %[count]")
>            __ASM_EMIT("jz          2000f")
>            __ASM_EMIT("shufps      $0x00, %%xmm0, %%xmm0")
>
>            /* Align destination */
>            __ASM_EMIT("1:")
>            __ASM_EMIT("test        $0x0f, %[dst]")
>            __ASM_EMIT("jz          2f")
>            __ASM_EMIT("movss       %%xmm0, 0x00(%[dst])")
>            __ASM_EMIT("add         $0x4, %[dst]")
>            __ASM_EMIT("dec         %[count]")
>            __ASM_EMIT("jnz         1b")
>            __ASM_EMIT("jmp         2000f")
>
>            __ASM_EMIT("2:")
>            __ASM_EMIT("movaps      %%xmm0, %%xmm1")
>            __ASM_EMIT("movaps      %%xmm0, %%xmm2")
>            __ASM_EMIT("movaps      %%xmm1, %%xmm3")
>
>            /* Destination is aligned */
>            FILL_CORE
>
>            __ASM_EMIT("2000:")
>
>            : [dst] "+r"(dst), [count] "+r" (count), "+Yz"(value) :
>            : "cc", "memory",
>              "%xmm1", "%xmm2", "%xmm3",
>              "%xmm4", "%xmm5", "%xmm6", "%xmm7"
>        );
>
>        #undef FILL_CORE
>    }
>}
>
>float test[0x100];
>
>int main()
>{
>    sse::fill(test, 1.0f, sizeof(test)/sizeof(float));
>}

Actions: View

Attachments on bug 232911: 198890