Improve Thumb multiplication performance a bit (#67)

This commit is contained in:
Ken MacKay
2016-04-23 22:50:41 -07:00
parent d671b1cc22
commit e862f7c5a1
+20 -20
View File
@@ -554,10 +554,10 @@ uECC_VLI_API void uECC_vli_mult(uECC_word_t *result,
"subs %[r7], %[r0] \n\t" /* r7 = i = k - (num_words - 1) (times 4) */
"3: \n\t" /* inner loop */
"push {%[r6]} \n\t"
"push {%[r5]} \n\t"
"push {%[r4]} \n\t"
"push {%[r3]} \n\t" /* push things, r3 (c0) is at the top of stack. */
"mov r10, %[r3] \n\t"
"mov r11, %[r4] \n\t"
"mov r12, %[r5] \n\t"
"mov r14, %[r6] \n\t"
"subs %[r0], %[r6], %[r7] \n\t" /* r0 = k - i */
"ldr %[r4], [%[r2], %[r0]] \n\t" /* r4 = right[k - i] */
@@ -586,15 +586,15 @@ uECC_VLI_API void uECC_vli_mult(uECC_word_t *result,
"adds %[r0], %[r4] \n\t"
"adcs %[r6], %[r5] \n\t"
"pop {%[r3]} \n\t" /* r3 = c0 */
"pop {%[r4]} \n\t" /* r4 = c1 */
"pop {%[r5]} \n\t" /* r5 = c2 */
"mov %[r3], r10\n\t"
"mov %[r4], r11\n\t"
"mov %[r5], r12\n\t"
"adds %[r3], %[r0] \n\t" /* add low word to c0 */
"adcs %[r4], %[r6] \n\t" /* add high word to c1, including carry */
"movs %[r0], #0 \n\t" /* r0 = 0 (does not affect carry bit) */
"adcs %[r5], %[r0] \n\t" /* add carry to c2 */
"pop {%[r6]} \n\t" /* r6 = k */
"mov %[r6], r14\n\t" /* r6 = k */
"adds %[r7], #4 \n\t" /* i += 4 */
"cmp %[r7], r8 \n\t" /* i > (num_words - 1) (times 4)? */
@@ -624,7 +624,7 @@ uECC_VLI_API void uECC_vli_mult(uECC_word_t *result,
: [r3] "+l" (num_words), [r4] "=&l" (r4),
[r5] "=&l" (r5), [r6] "=&l" (r6), [r7] "=&l" (r7)
: [r0] "l" (result), [r1] "l" (left), [r2] "l" (right)
: "r8", "r9", "cc", "memory"
: "r8", "r9", "r10", "r11", "r12", "r14", "cc", "memory"
);
#endif
}
@@ -713,7 +713,6 @@ uECC_VLI_API void uECC_vli_square(uECC_word_t *result,
"lsls %[r2], #1 \n\t" /* r2 = (num_words - 1) * 8 */
"mov r9, %[r2] \n\t" /* r9 = (num_words - 1) * 8 */
"movs %[r2], #0 \n\t" /* c0 = 0 */
"mov r10, %[r2] \n\t" /* r10 = 0 */
"movs %[r3], #0 \n\t" /* c1 = 0 */
"movs %[r4], #0 \n\t" /* c2 = 0 */
"movs %[r5], #0 \n\t" /* k = 0 */
@@ -730,10 +729,10 @@ uECC_VLI_API void uECC_vli_square(uECC_word_t *result,
"subs %[r6], %[r0] \n\t" /* r6 = i = k - (num_words - 1) (times 4) */
"3: \n\t" /* inner loop */
"push {%[r5]} \n\t"
"push {%[r4]} \n\t"
"push {%[r3]} \n\t"
"push {%[r2]} \n\t" /* push things, r2 (c0) is at the top of stack. */
"mov r10, %[r2] \n\t"
"mov r11, %[r3] \n\t"
"mov r12, %[r4] \n\t"
"mov r14, %[r5] \n\t"
"subs %[r7], %[r5], %[r6] \n\t" /* r7 = k - i */
"ldr %[r3], [%[r1], %[r7]] \n\t" /* r3 = left[k - i] */
@@ -763,22 +762,23 @@ uECC_VLI_API void uECC_vli_square(uECC_word_t *result,
"adcs %[r5], %[r4] \n\t"
/* Add to acc, doubling if necessary */
"pop {%[r2]} \n\t" /* r2 = c0 */
"pop {%[r3]} \n\t" /* r3 = c1 */
"pop {%[r4]} \n\t" /* r4 = c2 */
"mov %[r2], r10\n\t"
"mov %[r3], r11\n\t"
"mov %[r4], r12\n\t"
"cmp %[r6], %[r7] \n\t" /* (i < k - i) ? */
"mov %[r7], r10 \n\t" /* r7 = 0 (does not affect flags) */
"bge 4f \n\t" /* if i >= k - i, skip */
"movs %[r7], #0 \n\t" /* r7 = 0 */
"adds %[r2], %[r0] \n\t" /* add low word to c0 */
"adcs %[r3], %[r5] \n\t" /* add high word to c1, including carry */
"adcs %[r4], %[r7] \n\t" /* add carry to c2 */
"4: \n\t"
"movs %[r7], #0 \n\t" /* r7 = 0 */
"adds %[r2], %[r0] \n\t" /* add low word to c0 */
"adcs %[r3], %[r5] \n\t" /* add high word to c1, including carry */
"adcs %[r4], %[r7] \n\t" /* add carry to c2 */
"pop {%[r5]} \n\t" /* r5 = k */
"mov %[r5], r14\n\t" /* r5 = k */
"adds %[r6], #4 \n\t" /* i += 4 */
"cmp %[r6], %[r5] \n\t" /* i >= k? */
@@ -809,7 +809,7 @@ uECC_VLI_API void uECC_vli_square(uECC_word_t *result,
: [r2] "+l" (num_words), [r3] "=&l" (r3), [r4] "=&l" (r4),
[r5] "=&l" (r5), [r6] "=&l" (r6), [r7] "=&l" (r7)
: [r0] "l" (result), [r1] "l" (left)
: "r8", "r9", "r10", "cc", "memory"
: "r8", "r9", "r10", "r11", "r12", "r14", "cc", "memory"
);
#endif
}