Twitter troll task to remove if-else from grading function. Compile with clang/gcc -Wpedantic -Wall -Wextra -O3 -msse -mpopcnt

check_grades.c

#include <immintrin.h>

typedef unsigned int u32;
typedef unsigned char u8;

char check_grade(u8 score) {
  score = score > 100 ? 100 : score;

  __m128i scores = _mm_set1_epi8(score);
  __m128i brackets = _mm_set_epi8(89, 79, 69, 59, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F);
  __m128i res_mask = _mm_cmpgt_epi8(scores, brackets);
  u32 res = _mm_movemask_epi8(res_mask);
  int index = _mm_popcnt_u32(res);
  return "FDCBA"[index];
}

void check_grade4(char out_grades[restrict 4], u8 score[restrict 4]) {
  __m128i scores = _mm_set_epi8(
        score[0], score[0], score[0], score[0],
        score[1], score[1], score[1], score[1],
        score[2], score[2], score[2], score[2],
        score[3], score[3], score[3], score[3]
    );

  __m128i max_scores = _mm_set1_epi8(100);
  scores = _mm_min_epu8(scores, max_scores);

  __m128i brackets = _mm_set_epi8(
        89, 79, 69, 59,
        89, 79, 69, 59,
        89, 79, 69, 59,
        89, 79, 69, 59
    );
  __m128i res_masks = _mm_cmpgt_epi8(scores, brackets);
  u32 res = _mm_movemask_epi8(res_masks);
  
  static const char * const LUT = "FDCBA";
  int index0 = _mm_popcnt_u32(res & 0xF000);
  int index1 = _mm_popcnt_u32(res & 0x0F00);
  int index2 = _mm_popcnt_u32(res & 0x00F0);
  int index3 = _mm_popcnt_u32(res & 0x000F);
  out_grades[0] = LUT[index0];
  out_grades[1] = LUT[index1];
  out_grades[2] = LUT[index2];
  out_grades[3] = LUT[index3];
}

// Test
#include <stdio.h>

#define ALIGNED(bytes) __attribute__ ((aligned (bytes)))

int main(void) {
  enum {SIZE = 128};
  ALIGNED(16) u8 scores[SIZE];
  ALIGNED(16) char grades[SIZE];
  ALIGNED(16) char grades4[SIZE];

  for (u32 i = 0; i < SIZE; ++i) {
    scores[i] = i;
  }

  for (u32 i = 0; i < SIZE; ++i) {
    grades[i] = check_grade(scores[i]);
  }
  for (u32 i = 0; i < SIZE; i += 4) {
    check_grade4(grades4 + i, scores + i);
  }

  puts("Score | check_grade | check_grade4'");
  puts("-----------------------------------");
  for (u32 i = 0; i < SIZE; ++i) {
    printf("%-3u   | '%c'         | '%c'\n", i, grades[i], grades4[i]);
    fflush(stdout);
    if (grades[i] != grades4[i])
    {
      __builtin_trap();
    }
  }

  return 0;
}