// icc -msse -O2 -Wall -c icc_register_pressure_bug.c -o bug.o
// icc -v: icc version 13.0.1 (gcc version 4.7.0 compatibility)

// multiple bugs for ICC under heaver register pressure
// gcc 4.7.2 and clang 3.2 both work correctly

#include <immintrin.h>

#define XMM_MOVE(xmm_src, xmm_dst)                      \
    __asm volatile("movdqa %1, %0" :                    \
                   /* writes %0 */ "=x" (xmm_dst) :     \
                   /* reads  %1 */ "x"  (xmm_src))

// function with 14 specified registers works as expected
__m128 using_14_xmm_registers_works() {
    register __m128 X0 asm("%xmm0") = {0, 1};
    register __m128 X1 asm("%xmm1") = {1, 2};
    register __m128 X2 asm("%xmm2") = {2, 3};
    register __m128 X3 asm("%xmm3");
    register __m128 X4 asm("%xmm4");
    register __m128 X5 asm("%xmm5");
    register __m128 X6 asm("%xmm6");
    register __m128 X7 asm("%xmm7");
    register __m128 X8 asm("%xmm8");
    register __m128 X9 asm("%xmm9");
    register __m128 X10 asm("%xmm10");
    register __m128 X11 asm("%xmm11");
    register __m128 X12 asm("%xmm12");
    register __m128 X13 asm("%xmm13");

    XMM_MOVE(X0, X1);
    XMM_MOVE(X1, X2);
    XMM_MOVE(X2, X3);
    XMM_MOVE(X3, X4);
    XMM_MOVE(X4, X5);
    XMM_MOVE(X5, X6);
    XMM_MOVE(X6, X7);
    XMM_MOVE(X7, X8);
    XMM_MOVE(X8, X9);
    XMM_MOVE(X9, X10);
    XMM_MOVE(X10, X11);
    XMM_MOVE(X11, X12);
    XMM_MOVE(X12, X13);
    
    return X13;
}

// function with 15 specified XMM registers is silently wrong 
// look at asm: unitialized value of %xmm14 is returned
__m128 using_15_xmm_registers_is_silently_wrong() {
    register __m128 X0 asm("%xmm0") = {0, 1};
    register __m128 X1 asm("%xmm1") = {1, 2};
    register __m128 X2 asm("%xmm2") = {2, 3};
    register __m128 X3 asm("%xmm3");
    register __m128 X4 asm("%xmm4");
    register __m128 X5 asm("%xmm5");
    register __m128 X6 asm("%xmm6");
    register __m128 X7 asm("%xmm7");
    register __m128 X8 asm("%xmm8");
    register __m128 X9 asm("%xmm9");
    register __m128 X10 asm("%xmm10");
    register __m128 X11 asm("%xmm11");
    register __m128 X12 asm("%xmm12");
    register __m128 X13 asm("%xmm13");
    register __m128 X14 asm("%xmm14");

    XMM_MOVE(X0, X1);
    XMM_MOVE(X1, X2);
    XMM_MOVE(X2, X3);
    XMM_MOVE(X3, X4);
    XMM_MOVE(X4, X5);
    XMM_MOVE(X5, X6);
    XMM_MOVE(X6, X7);
    XMM_MOVE(X7, X8);
    XMM_MOVE(X8, X9);
    XMM_MOVE(X9, X10);
    XMM_MOVE(X10, X11);
    XMM_MOVE(X11, X12);
    XMM_MOVE(X12, X13);
    XMM_MOVE(X13, X14);
    
    return X14;
}

// add -DCATASTROPHIC_ERROR to see bug for 16 specified XMM registers
// $ icc -DCATASTROPHIC_ERROR -msse -O2 -Wall -c icc_register_pressure_bug.c
//       icc_register_pressure_bug.c(103): catastrophic error: 
//       Cannot match asm operand constraint
#ifdef CATASTROPHIC_ERROR
__m128 using_16_xmm_registers_errors_out() {
    register __m128 X0 asm("%xmm0") = {0, 1};
    register __m128 X1 asm("%xmm1") = {1, 2};
    register __m128 X2 asm("%xmm2") = {2, 3};
    register __m128 X3 asm("%xmm3");
    register __m128 X4 asm("%xmm4");
    register __m128 X5 asm("%xmm5");
    register __m128 X6 asm("%xmm6");
    register __m128 X7 asm("%xmm7");
    register __m128 X8 asm("%xmm8");
    register __m128 X9 asm("%xmm9");
    register __m128 X10 asm("%xmm10");
    register __m128 X11 asm("%xmm11");
    register __m128 X12 asm("%xmm12");
    register __m128 X13 asm("%xmm13");
    register __m128 X14 asm("%xmm14");
    register __m128 X15 asm("%xmm15");

    XMM_MOVE(X0, X1);
    XMM_MOVE(X1, X2);
    XMM_MOVE(X2, X3);
    XMM_MOVE(X3, X4);
    XMM_MOVE(X4, X5);
    XMM_MOVE(X5, X6);
    XMM_MOVE(X6, X7);
    XMM_MOVE(X7, X8);
    XMM_MOVE(X8, X9);
    XMM_MOVE(X9, X10);
    XMM_MOVE(X10, X11);
    XMM_MOVE(X11, X12);
    XMM_MOVE(X12, X13);
    XMM_MOVE(X13, X14);
    XMM_MOVE(X14, X15);
    
    return X15;
}
#endif // CATASTROPHIC_ERROR