umaal.c

uECC_VLI_API void uECC_vli_mult(uint32_t *result, const uint32_t *left, const uint32_t *right) {
    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"
        "ldmia  r2!, {r3, r4, r5, r6} \n\t" // Load right number

        "ldr    r14, [r1], #4 \n\t"         // Load left digit
        "umull  r7, r8, r3, r14 \n\t"
        "str    r7, [r0], #4 \n\t"          // Store result digit
        
        "mov    r9, #0 \n\t"                // Carry prepare
        "umaal  r8, r9, r4, r14 \n\t"       // Multiply
        
        "mov    r10, #0 \n\t"               // Carry prepare
        "umaal  r9, r10, r5, r14 \n\t"      // Multiply
                                            
        "mov    r11, #0 \n\t"               // Carry prepare
        "umaal  r10, r11, r6, r14 \n\t"     // Multiply
        
        REPEAT(7,
            "ldr    r14, [r1], #4 \n\t"         // Load left digit
            "mov    r7, #0 \n\t"                // Carry prepare
            "umaal  r7, r8, r3, r14 \n\t"       // Multiply
            "str    r7, [r0], #4 \n\t"          // Store partial result digit
        
            "umaal  r8, r9, r4, r14 \n\t"       // Multiply
            "umaal  r9, r10, r5, r14 \n\t"      // Multiply
            "umaal  r10, r11, r6, r14 \n\t"     // Multiply
        )
        
        "str    r8, [r0], #4 \n\t"          // Store partial result digit
        "str    r9, [r0], #4 \n\t"          // Store partial result digit
        "str    r10, [r0], #4 \n\t"         // Store partial result digit
        "str    r11, [r0], #4 \n\t"         // Store partial result digit
        
        "sub r0, #32 \n\t"
        "sub r1, #32 \n\t"
        "ldmia  r2!, {r3, r4, r5, r6} \n\t" // Load right number
        
        "ldr    r14, [r1], #4 \n\t"         // Load left digit
        "ldr    r7, [r0] \n\t"              // Load partial result digit
        "mov    r8, #0 \n\t"                // Carry prepare
        "umaal  r7, r8, r3, r14 \n\t"       // Multiply
        "str    r7, [r0], #4 \n\t"          // Store result digit
        
        "mov    r9, #0 \n\t"                // Carry prepare
        "umaal  r8, r9, r4, r14 \n\t"       // Multiply
        
        "mov    r10, #0 \n\t"               // Carry prepare
        "umaal  r9, r10, r5, r14 \n\t"      // Multiply
        
        "mov    r11, #0 \n\t"               // Carry prepare
        "umaal  r10, r11, r6, r14 \n\t"     // Multiply
        
        REPEAT(7,
            "ldr    r14, [r1], #4 \n\t"         // Load left digit
            "ldr    r7, [r0] \n\t"              // Load partial result digit
            "umaal  r7, r8, r3, r14 \n\t"       // Multiply
            "str    r7, [r0], #4 \n\t"          // Store result digit
        
            "umaal  r8, r9, r4, r14 \n\t"       // Multiply
            "umaal  r9, r10, r5, r14 \n\t"      // Multiply
            "umaal  r10, r11, r6, r14 \n\t"     // Multiply
        )
        
        "str    r8, [r0], #4 \n\t"          // Store result digit
        "str    r9, [r0], #4 \n\t"          // Store result digit
        "str    r10, [r0], #4 \n\t"         // Store result digit
        "str    r11, [r0], #4 \n\t"         // Store result digit

    #if (uECC_PLATFORM != uECC_arm_thumb2)
        ".syntax divided \n\t"
    #endif
        : "+r" (r0), "+r" (r1), "+r" (r2)
        :
        : "r3", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11", "r12", "r14", "cc", "memory"
    );
}