SSE2 assembly to intrinsics convertor. Unfinished, but still useful for porting.

main.c

#include <stdlib.h>
#include <stdio.h>
#include <string.h>

static int is_mem(const char *s, int n) {
	int a, i;
	for (i = 0; i < n; i++) {
		a = s[i];
		if (a == '[' || a == ']') return 1;
	}
	return 0;
}

// TODO:
// xor,sub same reg = _mm_setzero_si128();
// cmpeq same reg = _mm_set1_?(-1);

int main() {
	static const char *repl[][2] = {
		{ "movq", "\0_mm_loadl_epi64\0_mm_storel_epi64" },
		{ "movdqa", "\0_mm_load_si128\0_mm_store_si128" },
		{ "movdqu", "\0_mm_loadu_si128\0_mm_storeu_si128" },

		{ "pand", "_mm_and_si128" },
		{ "pandn", "_mm_andnot_si128" },
		{ "pxor", "_mm_xor_si128" },
		{ "por", "_mm_or_si128" },

		{ "paddb", "_mm_add_epi8" },
		{ "paddw", "_mm_add_epi16" },
		{ "paddd", "_mm_add_epi32" },
		{ "psubb", "_mm_sub_epi8" },
		{ "psubw", "_mm_sub_epi16" },
		{ "psubd", "_mm_sub_epi32" },
		{ "paddsb", "_mm_adds_epi8" },
		{ "paddsw", "_mm_adds_epi16" },
		{ "psubsb", "_mm_subs_epi8" },
		{ "psubsw", "_mm_subs_epi16" },
		{ "paddusb", "_mm_adds_epu8" },
		{ "paddusw", "_mm_adds_epu16" },
		{ "psubusb", "_mm_subs_epu8" },
		{ "psubusw", "_mm_subs_epu16" },

		{ "psllw", "_mm_slli_epi16" },
		{ "pslld", "_mm_slli_epi32" },
		{ "psllq", "_mm_slli_epi64" },
		{ "psrlw", "_mm_srli_epi16" },
		{ "psrld", "_mm_srli_epi32" },
		{ "psrlq", "_mm_srli_epi64" },
		{ "psraw", "_mm_srai_epi16" },
		{ "psrad", "_mm_srai_epi32" },
		{ "pslldq", "_mm_bslli_si128" },
		{ "psrldq", "_mm_bsrli_si128" },

		{ "packsswb", "_mm_packs_epi16" },
		{ "packssdw", "_mm_packs_epi32" },
		{ "packuswb", "_mm_packus_epi16" },
		{ "packusdw", "_mm_packus_epi32" },
		{ "punpcklbw", "_mm_unpacklo_epi8" },
		{ "punpckhbw", "_mm_unpackhi_epi8" },
		{ "punpcklwd", "_mm_unpacklo_epi16" },
		{ "punpckhwd", "_mm_unpackhi_epi16" },
		{ "punpckldq", "_mm_unpacklo_epi32" },
		{ "punpckhdq", "_mm_unpackhi_epi32" },
		{ "punpcklqdq", "_mm_unpacklo_epi64" },
		{ "punpckhqdq", "_mm_unpackhi_epi64" },

		{ "pcmpeqb", "_mm_cmpeq_epi8" },
		{ "pcmpeqw", "_mm_cmpeq_epi16" },
		{ "pcmpeqd", "_mm_cmpeq_epi32" },
		{ "pcmpgtb", "_mm_cmpgt_epi8" },
		{ "pcmpgtw", "_mm_cmpgt_epi16" },
		{ "pcmpgtd", "_mm_cmpgt_epi32" },

		{ "pshufb", "_mm_shuffle_epi8" },
		{ "pshufd", "_mm_shuffle_epi32" },
		{ "pshuflw", "_mm_shufflelo_epi16" },
		{ "pshufhw", "_mm_shufflehi_epi16" },

		{ "pmullw", "_mm_mullo_epi16" },
		{ "pmulhw", "_mm_mulhi_epi16" },
		{ "pmulhuw", "_mm_mulhi_epu16" },

		{ "pminub", "_mm_min_epu8" },
		{ "pmaxub", "_mm_max_epu8" },
		{ "pminsw", "_mm_min_epi16" },
		{ "pmaxsw", "_mm_max_epi16" },

		{ "paddsb", "_mm_adds_epi8" },
		{ "paddsw", "_mm_adds_epi16" },
		{ "pmaddubsw", "_mm_maddubs_epi16" },
		{ "pmaddwd", "_mm_madd_epi16" }
	};

	char buf[256];

	for (;;) {
		int a, e, i0 = 0, i1, i2, i3, i4, i5, i6, i7, i8, k, n = 0, found;
		do {
			e = getchar();
			buf[n++] = e;
		} while (n < sizeof(buf) && e != '\n' && e != '\r' && e != ';' && e != EOF);
		buf[--n] = 0;

#define SKIPWS(i, i0) \
	for (i = i0; ; i++) { \
		a = buf[i]; if (a != ' ' && a != '\t') break; \
	}
#define SKIPWORD(i, i0) \
	for (i = i0; ; i++) { \
		a = buf[i]; if (!a || a == ' ' || a == '\t' || a == ',') break; \
	}
#define SKIPARG(j, i, i0) \
	for (i = j = i0; ; j++) { \
		a = buf[j]; \
		if (!a || a == ',') break; \
		if (a != ' ' && a != '\t') i = j + 1; \
	}
#define FINDREPL(r) \
	for (k = 0; k < sizeof(r) / sizeof(*r); k++) \
		if (buf[i1] == **r[k]) { \
			int l = strlen(r[k][0]); \
			if (i2 - i1 == l && !memcmp(buf + i1, *r[k], l)) break; \
		} \
	found = k < sizeof(r) / sizeof(*r);

		do {
			SKIPWS(i1, 0) if (!a) break;
			SKIPWORD(i2, i1) if (!a || a == ',') break;
			FINDREPL(repl)
			if (!found) break;
			if (!*repl[k][1]) {
				SKIPWS(i3, i2) if (!a) break;
				SKIPARG(i5, i4, i3) if (!a) break;
				if (a != ',') break;
				i5++;
				SKIPWS(i5, i5) if (!a) break;
				SKIPARG(i7, i6, i5)
				if (is_mem(buf + i3, i4 - i3)) {
					printf("\t%s(%.*s, %.*s);",
							repl[k][1] + strlen(repl[k][1] + 1) + 2, i4 - i3, buf + i3, i6 - i5, buf + i5);
				} else if (is_mem(buf + i5, i6 - i5)) {
					printf("\t%.*s = %s(%.*s);",
							i4 - i3, buf + i3, repl[k][1] + 1, i6 - i5, buf + i5);
				} else {
					printf("\t%.*s = %.*s;",
							i4 - i3, buf + i3, i6 - i5, buf + i5);
				}
				i0 = i6;
				break;
			}
			SKIPWS(i3, i2) if (!a) break;
			SKIPARG(i5, i4, i3) if (!a) break;
			if (a != ',') break;
			i5++;
			SKIPWS(i5, i5) if (!a) break;
			SKIPARG(i7, i6, i5)
			if (a != ',') {
				printf("\t%.*s = %s(%.*s, %.*s);",
						i4 - i3, buf + i3, repl[k][1],
						i4 - i3, buf + i3, i6 - i5, buf + i5);
				i0 = i6;
			} else {
				i7++;
				SKIPWS(i7, i7) if (!a) break;
				SKIPWORD(i8, i7)
				printf("\t%.*s = %s(%.*s, %.*s);",
						i4 - i3, buf + i3, repl[k][1],
						i6 - i5, buf + i5, i8 - i7, buf + i7);
				i0 = i8;
			}
		} while (0);

		printf("%s", buf + i0);
		if (e == EOF) break;
		if (e == ';') printf("//");
		else putchar(e);
		while (e != '\n' && e != '\r') {
			e = getchar();
			if (e == EOF) return 0;
			putchar(e);
		}
	}
}