Fix Thumb-1 assembly
diff --git a/asm_arm.inc b/asm_arm.inc
index f71e89d..e9d0995 100644
--- a/asm_arm.inc
+++ b/asm_arm.inc
@@ -985,95 +985,99 @@
);
#else /* Thumb-1 */
+ uint32_t r4, r5, r6, r7;
- register uint32_t *r0 __asm__("r0") = result;
- register const uint32_t *r1 __asm__("r1") = left;
- register const uint32_t *r2 __asm__("r2") = right;
-
__asm__ volatile (
".syntax unified \n\t"
- "movs r3, #0 \n\t" /* c0 = 0 */
- "movs r4, #0 \n\t" /* c1 = 0 */
- "movs r5, #0 \n\t" /* c2 = 0 */
- "movs r6, #0 \n\t" /* k = 0 */
+ "subs %[r3], #1 \n\t" /* r3 = num_words - 1 */
+ "lsls %[r3], #2 \n\t" /* r3 = (num_words - 1) * 4 */
+ "mov r8, %[r3] \n\t" /* r8 = (num_words - 1) * 4 */
+ "lsls %[r3], #1 \n\t" /* r3 = (num_words - 1) * 8 */
+ "mov r9, %[r3] \n\t" /* r9 = (num_words - 1) * 8 */
+ "movs %[r3], #0 \n\t" /* c0 = 0 */
+ "movs %[r4], #0 \n\t" /* c1 = 0 */
+ "movs %[r5], #0 \n\t" /* c2 = 0 */
+ "movs %[r6], #0 \n\t" /* k = 0 */
- "push {r0} \n\t" /* keep result on the stack */
+ "push {%[r0]} \n\t" /* keep result on the stack */
"1: \n\t" /* outer loop (k < num_words) */
- "movs r7, #0 \n\t" /* r7 = i = 0 */
+ "movs %[r7], #0 \n\t" /* r7 = i = 0 */
"b 3f \n\t"
"2: \n\t" /* outer loop (k >= num_words) */
- "movs r7, r6 \n\t" /* r7 = k */
- "subs r7, %[last_word] \n\t" /* r7 = i = k - (num_words - 1) (times 4) */
+ "movs %[r7], %[r6] \n\t" /* r7 = k */
+ "mov %[r0], r8 \n\t" /* r0 = (num_words - 1) * 4 */
+ "subs %[r7], %[r0] \n\t" /* r7 = i = k - (num_words - 1) (times 4) */
"3: \n\t" /* inner loop */
- "push {r3, r4, r5, r6} \n\t" /* push things, r3 (c0) is at the top of stack. */
- "subs r0, r6, r7 \n\t" /* r0 = k - i */
+ "push {%[r3], %[r4], %[r5], %[r6]} \n\t" /* push things, r3 (c0) is at the top of stack. */
+ "subs %[r0], %[r6], %[r7] \n\t" /* r0 = k - i */
- "ldr r4, [r2, r0] \n\t" /* r4 = right[k - i] */
- "ldr r0, [r1, r7] \n\t" /* r0 = left[i] */
+ "ldr %[r4], [%[r2], %[r0]] \n\t" /* r4 = right[k - i] */
+ "ldr %[r0], [%[r1], %[r7]] \n\t" /* r0 = left[i] */
- "lsrs r3, r0, #16 \n\t" /* r3 = a1 */
- "uxth r0, r0 \n\t" /* r0 = a0 */
+ "lsrs %[r3], %[r0], #16 \n\t" /* r3 = a1 */
+ "uxth %[r0], %[r0] \n\t" /* r0 = a0 */
- "lsrs r5, r4, #16 \n\t" /* r5 = b1 */
- "uxth r4, r4 \n\t" /* r4 = b0 */
+ "lsrs %[r5], %[r4], #16 \n\t" /* r5 = b1 */
+ "uxth %[r4], %[r4] \n\t" /* r4 = b0 */
- "movs r6, r3 \n\t" /* r6 = a1 */
- "muls r6, r5, r6 \n\t" /* r6 = a1 * b1 */
- "muls r3, r4, r3 \n\t" /* r3 = b0 * a1 */
- "muls r5, r0, r5 \n\t" /* r5 = a0 * b1 */
- "muls r0, r4, r0 \n\t" /* r0 = a0 * b0 */
+ "movs %[r6], %[r3] \n\t" /* r6 = a1 */
+ "muls %[r6], %[r5], %[r6] \n\t" /* r6 = a1 * b1 */
+ "muls %[r3], %[r4], %[r3] \n\t" /* r3 = b0 * a1 */
+ "muls %[r5], %[r0], %[r5] \n\t" /* r5 = a0 * b1 */
+ "muls %[r0], %[r4], %[r0] \n\t" /* r0 = a0 * b0 */
- "movs r4, #0 \n\t" /* r4 = 0 */
- "adds r3, r5 \n\t" /* r3 = b0 * a1 + a0 * b1 */
- "adcs r4, r4 \n\t" /* r4 = carry */
- "lsls r4, #16 \n\t" /* r4 = carry << 16 */
- "adds r6, r4 \n\t" /* r6 = a1 * b1 + carry */
+ "movs %[r4], #0 \n\t" /* r4 = 0 */
+ "adds %[r3], %[r5] \n\t" /* r3 = b0 * a1 + a0 * b1 */
+ "adcs %[r4], %[r4] \n\t" /* r4 = carry */
+ "lsls %[r4], #16 \n\t" /* r4 = carry << 16 */
+ "adds %[r6], %[r4] \n\t" /* r6 = a1 * b1 + carry */
- "lsls r4, r3, #16 \n\t" /* r4 = (b0 * a1 + a0 * b1) << 16 */
- "lsrs r3, #16 \n\t" /* r3 = (b0 * a1 + a0 * b1) >> 16 */
- "adds r0, r4 \n\t" /* r0 = low word = a0 * b0 + ((b0 * a1 + a0 * b1) << 16) */
- "adcs r6, r3 \n\t" /* r6 = high word = a1 * b1 + carry + ((b0 * a1 + a0 * b1) >> 16) */
+ "lsls %[r4], %[r3], #16 \n\t" /* r4 = (b0 * a1 + a0 * b1) << 16 */
+ "lsrs %[r3], #16 \n\t" /* r3 = (b0 * a1 + a0 * b1) >> 16 */
+ "adds %[r0], %[r4] \n\t" /* r0 = low word = a0 * b0 + ((b0 * a1 + a0 * b1) << 16) */
+ "adcs %[r6], %[r3] \n\t" /* r6 = high word =
+ a1 * b1 + carry + ((b0 * a1 + a0 * b1) >> 16) */
- "pop {r3, r4, r5} \n\t" /* r3 = c0, r4 = c1, r5 = c2 */
- "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 {%[r3], %[r4], %[r5]} \n\t" /* r3 = c0, r4 = c1, r5 = c2 */
+ "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 */
+ "pop {%[r6]} \n\t" /* r6 = k */
- "adds r7, #4 \n\t" /* i += 4 */
- "cmp r7, %[last_word] \n\t" /* i > (num_words - 1) (times 4)? */
- "bgt 4f \n\t" /* if so, exit the loop */
- "cmp r7, r6 \n\t" /* i <= k? */
- "ble 3b \n\t" /* if so, continue looping */
+ "adds %[r7], #4 \n\t" /* i += 4 */
+ "cmp %[r7], r8 \n\t" /* i > (num_words - 1) (times 4)? */
+ "bgt 4f \n\t" /* if so, exit the loop */
+ "cmp %[r7], %[r6] \n\t" /* i <= k? */
+ "ble 3b \n\t" /* if so, continue looping */
"4: \n\t" /* end inner loop */
- "ldr r0, [sp, #0] \n\t" /* r0 = result */
+ "ldr %[r0], [sp, #0] \n\t" /* r0 = result */
- "str r3, [r0, r6] \n\t" /* result[k] = c0 */
- "mov r3, r4 \n\t" /* c0 = c1 */
- "mov r4, r5 \n\t" /* c1 = c2 */
- "movs r5, #0 \n\t" /* c2 = 0 */
- "adds r6, #4 \n\t" /* k += 4 */
- "cmp r6, %[last_word] \n\t" /* k <= (num_words - 1) (times 4) ? */
- "ble 1b \n\t" /* if so, loop back, start with i = 0 */
- "cmp r6, %[lw2] \n\t" /* k <= (num_words * 2 - 2) (times 4) ? */
- "ble 2b \n\t" /* if so, loop back, start with i = (k + 1) - num_words */
+ "str %[r3], [%[r0], %[r6]] \n\t" /* result[k] = c0 */
+ "mov %[r3], %[r4] \n\t" /* c0 = c1 */
+ "mov %[r4], %[r5] \n\t" /* c1 = c2 */
+ "movs %[r5], #0 \n\t" /* c2 = 0 */
+ "adds %[r6], #4 \n\t" /* k += 4 */
+ "cmp %[r6], r8 \n\t" /* k <= (num_words - 1) (times 4) ? */
+ "ble 1b \n\t" /* if so, loop back, start with i = 0 */
+ "cmp %[r6], r9 \n\t" /* k <= (num_words * 2 - 2) (times 4) ? */
+ "ble 2b \n\t" /* if so, loop back, with i = (k + 1) - num_words */
/* end outer loop */
- "str r3, [r0, r6] \n\t" /* result[num_words * 2 - 1] = c0 */
- "pop {r0} \n\t" /* pop result off the stack */
+ "str %[r3], [%[r0], %[r6]] \n\t" /* result[num_words * 2 - 1] = c0 */
+ "pop {%[r0]} \n\t" /* pop result off the stack */
".syntax divided \n\t"
- :
- : [r0] "l" (r0), [r1] "l" (r1), [r2] "l" (r2),
- [last_word] "r" ((num_words - 1) * 4), [lw2] "r" ((num_words - 1) * 4 * 2)
- : "r3", "r4", "r5", "r6", "r7", "cc", "memory"
+ : [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"
);
#endif
}
@@ -1154,105 +1158,111 @@
);
#else
+ uint32_t r3, r4, r5, r6, r7;
- register uint32_t *r0 __asm__("r0") = result;
- register const uint32_t *r1 __asm__("r1") = left;
-
__asm__ volatile (
".syntax unified \n\t"
- "movs r2, #0 \n\t" /* c0 = 0 */
- "movs r3, #0 \n\t" /* c1 = 0 */
- "movs r4, #0 \n\t" /* c2 = 0 */
- "movs r5, #0 \n\t" /* k = 0 */
+ "subs %[r2], #1 \n\t" /* r2 = num_words - 1 */
+ "lsls %[r2], #2 \n\t" /* r2 = (num_words - 1) * 4 */
+ "mov r8, %[r2] \n\t" /* r8 = (num_words - 1) * 4 */
+ "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 */
+ "movs %[r3], #0 \n\t" /* c1 = 0 */
+ "movs %[r4], #0 \n\t" /* c2 = 0 */
+ "movs %[r5], #0 \n\t" /* k = 0 */
- "push {r0} \n\t" /* keep result on the stack */
+ "push {%[r0]} \n\t" /* keep result on the stack */
"1: \n\t" /* outer loop (k < num_words) */
- "movs r6, #0 \n\t" /* r6 = i = 0 */
+ "movs %[r6], #0 \n\t" /* r6 = i = 0 */
"b 3f \n\t"
"2: \n\t" /* outer loop (k >= num_words) */
- "movs r6, r5 \n\t" /* r6 = k */
- "subs r6, %[last_word] \n\t" /* r6 = i = k - (num_words - 1) (times 4) */
+ "movs %[r6], %[r5] \n\t" /* r6 = k */
+ "mov %[r0], r8 \n\t" /* r0 = (num_words - 1) * 4 */
+ "subs %[r6], %[r0] \n\t" /* r6 = i = k - (num_words - 1) (times 4) */
"3: \n\t" /* inner loop */
- "push {r2, r3, r4, r5} \n\t" /* push things, r2 (c0) is at the top of stack. */
- "subs r7, r5, r6 \n\t" /* r7 = k - i */
+ "push {%[r2], %[r3], %[r4], %[r5]} \n\t" /* push things, r2 (c0) is at the top of stack. */
+ "subs %[r7], %[r5], %[r6] \n\t" /* r7 = k - i */
- "ldr r3, [r1, r7] \n\t" /* r3 = left[k - i] */
- "ldr r0, [r1, r6] \n\t" /* r0 = left[i] */
+ "ldr %[r3], [%[r1], %[r7]] \n\t" /* r3 = left[k - i] */
+ "ldr %[r0], [%[r1], %[r6]] \n\t" /* r0 = left[i] */
- "lsrs r2, r0, #16 \n\t" /* r2 = a1 */
- "uxth r0, r0 \n\t" /* r0 = a0 */
+ "lsrs %[r2], %[r0], #16 \n\t" /* r2 = a1 */
+ "uxth %[r0], %[r0] \n\t" /* r0 = a0 */
- "lsrs r4, r3, #16 \n\t" /* r4 = b1 */
- "uxth r3, r3 \n\t" /* r3 = b0 */
+ "lsrs %[r4], %[r3], #16 \n\t" /* r4 = b1 */
+ "uxth %[r3], %[r3] \n\t" /* r3 = b0 */
- "movs r5, r2 \n\t" /* r5 = a1 */
- "muls r5, r4, r5 \n\t" /* r5 = a1 * b1 */
- "muls r2, r3, r2 \n\t" /* r2 = b0 * a1 */
- "muls r4, r0, r4 \n\t" /* r4 = a0 * b1 */
- "muls r0, r3, r0 \n\t" /* r0 = a0 * b0 */
+ "movs %[r5], %[r2] \n\t" /* r5 = a1 */
+ "muls %[r5], %[r4], %[r5] \n\t" /* r5 = a1 * b1 */
+ "muls %[r2], %[r3], %[r2] \n\t" /* r2 = b0 * a1 */
+ "muls %[r4], %[r0], %[r4] \n\t" /* r4 = a0 * b1 */
+ "muls %[r0], %[r3], %[r0] \n\t" /* r0 = a0 * b0 */
- "movs r3, #0 \n\t" /* r3 = 0 */
- "adds r2, r4 \n\t" /* r2 = b0 * a1 + a0 * b1 */
- "adcs r3, r3 \n\t" /* r3 = carry */
- "lsls r3, #16 \n\t" /* r3 = carry << 16 */
- "adds r5, r3 \n\t" /* r5 = a1 * b1 + carry */
+ "movs %[r3], #0 \n\t" /* r3 = 0 */
+ "adds %[r2], %[r4] \n\t" /* r2 = b0 * a1 + a0 * b1 */
+ "adcs %[r3], %[r3] \n\t" /* r3 = carry */
+ "lsls %[r3], #16 \n\t" /* r3 = carry << 16 */
+ "adds %[r5], %[r3] \n\t" /* r5 = a1 * b1 + carry */
- "lsls r3, r2, #16 \n\t" /* r3 = (b0 * a1 + a0 * b1) << 16 */
- "lsrs r2, #16 \n\t" /* r2 = (b0 * a1 + a0 * b1) >> 16 */
- "adds r0, r3 \n\t" /* r0 = low word = a0 * b0 + ((b0 * a1 + a0 * b1) << 16) */
- "adcs r5, r2 \n\t" /* r5 = high word = a1 * b1 + carry + ((b0 * a1 + a0 * b1) >> 16) */
+ "lsls %[r3], %[r2], #16 \n\t" /* r3 = (b0 * a1 + a0 * b1) << 16 */
+ "lsrs %[r2], #16 \n\t" /* r2 = (b0 * a1 + a0 * b1) >> 16 */
+ "adds %[r0], %[r3] \n\t" /* r0 = low word = a0 * b0 + ((b0 * a1 + a0 * b1) << 16) */
+ "adcs %[r5], %[r2] \n\t" /* r5 = high word =
+ a1 * b1 + carry + ((b0 * a1 + a0 * b1) >> 16) */
- "movs r3, #0 \n\t" /* r3 = 0 */
- "cmp r6, r7 \n\t" /* (i < k - i) ? */
- "mov r7, r3 \n\t" /* r7 = 0 (does not affect condition)*/
- "bge 4f \n\t" /* if i >= k - i, skip */
- "lsls r5, #1 \n\t" /* high word << 1 */
- "adcs r7, r3 \n\t" /* r7 = carry bit for c2 */
- "lsls r0, #1 \n\t" /* low word << 1 */
- "adcs r5, r3 \n\t" /* add carry from shift to high word */
+ "movs %[r3], #0 \n\t" /* r3 = 0 */
+ "cmp %[r6], %[r7] \n\t" /* (i < k - i) ? */
+ "mov %[r7], %[r3] \n\t" /* r7 = 0 (does not affect condition) */
+ "bge 4f \n\t" /* if i >= k - i, skip */
+ "lsls %[r5], #1 \n\t" /* high word << 1 */
+ "adcs %[r7], %[r3] \n\t" /* r7 = carry bit for c2 */
+ "lsls %[r0], #1 \n\t" /* low word << 1 */
+ "adcs %[r5], %[r3] \n\t" /* add carry from shift to high word */
"4: \n\t"
- "pop {r2, r3, r4} \n\t" /* r2 = c0, r3 = c1, r4 = c2 */
- "adds r2, r0 \n\t" /* add low word to c0 */
- "adcs r3, r5 \n\t" /* add high word to c1, including carry */
- "movs r0, #0 \n\t" /* r0 = 0 (does not affect carry bit) */
- "adcs r4, r0 \n\t" /* add carry to c2 */
- "adds r4, r7 \n\t" /* add carry from doubling (if any) */
+ "pop {%[r2], %[r3], %[r4]} \n\t" /* r2 = c0, r3 = c1, r4 = c2 */
+ "adds %[r2], %[r0] \n\t" /* add low word to c0 */
+ "adcs %[r3], %[r5] \n\t" /* add high word to c1, including carry */
+ "movs %[r0], #0 \n\t" /* r0 = 0 (does not affect carry bit) */
+ "adcs %[r4], %[r0] \n\t" /* add carry to c2 */
+ "adds %[r4], %[r7] \n\t" /* add carry from doubling (if any) */
- "pop {r5} \n\t" /* r5 = k */
+ "pop {%[r5]} \n\t" /* r5 = k */
- "adds r6, #4 \n\t" /* i += 4 */
- "cmp r6, r5 \n\t" /* i >= k? */
- "bge 5f \n\t" /* if so, exit the loop */
- "subs r7, r5, r6 \n\t" /* r7 = k - i */
- "cmp r6, r7 \n\t" /* i <= k - i? */
- "ble 3b \n\t" /* if so, continue looping */
+ "adds %[r6], #4 \n\t" /* i += 4 */
+ "cmp %[r6], %[r5] \n\t" /* i >= k? */
+ "bge 5f \n\t" /* if so, exit the loop */
+ "subs %[r7], %[r5], %[r6] \n\t" /* r7 = k - i */
+ "cmp %[r6], %[r7] \n\t" /* i <= k - i? */
+ "ble 3b \n\t" /* if so, continue looping */
"5: \n\t" /* end inner loop */
- "ldr r0, [sp, #0] \n\t" /* r0 = result */
+ "ldr %[r0], [sp, #0] \n\t" /* r0 = result */
- "str r2, [r0, r5] \n\t" /* result[k] = c0 */
- "mov r2, r3 \n\t" /* c0 = c1 */
- "mov r3, r4 \n\t" /* c1 = c2 */
- "movs r4, #0 \n\t" /* c2 = 0 */
- "adds r5, #4 \n\t" /* k += 4 */
- "cmp r5, %[last_word] \n\t" /* k <= (num_words - 1) (times 4) ? */
- "ble 1b \n\t" /* if so, loop back, start with i = 0 */
- "cmp r5, %[lw2] \n\t" /* k <= (num_words * 2 - 2) (times 4) ? */
- "ble 2b \n\t" /* if so, loop back, start with i = (k + 1) - num_words */
+ "str %[r2], [%[r0], %[r5]] \n\t" /* result[k] = c0 */
+ "mov %[r2], %[r3] \n\t" /* c0 = c1 */
+ "mov %[r3], %[r4] \n\t" /* c1 = c2 */
+ "movs %[r4], #0 \n\t" /* c2 = 0 */
+ "adds %[r5], #4 \n\t" /* k += 4 */
+ "cmp %[r5], r8 \n\t" /* k <= (num_words - 1) (times 4) ? */
+ "ble 1b \n\t" /* if so, loop back, start with i = 0 */
+ "cmp %[r5], r9 \n\t" /* k <= (num_words * 2 - 2) (times 4) ? */
+ "ble 2b \n\t" /* if so, loop back, with i = (k + 1) - num_words */
/* end outer loop */
- "str r2, [r0, r5] \n\t" /* result[num_words * 2 - 1] = c0 */
- "pop {r0} \n\t" /* pop result off the stack */
+ "str %[r2], [%[r0], %[r5]] \n\t" /* result[num_words * 2 - 1] = c0 */
+ "pop {%[r0]} \n\t" /* pop result off the stack */
".syntax divided \n\t"
- : [r0] "+l" (r0), [r1] "+l" (r1)
- : [last_word] "r" ((num_words - 1) * 4), [lw2] "r" ((num_words - 1) * 4 * 2)
- : "r2", "r3", "r4", "r5", "r6", "r7", "cc", "memory"
+ : [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", "cc", "memory"
);
#endif
}
