GCC 9: miscompilation of saxpy_kernel_16

[marxin]

Using latest GCC one can see many test failures. I isolated a test-case for that:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88880#c0

and the problem is in:
../kernel/x86_64/saxpy_microk_haswell-2.c:

    31	static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
    32	{
    33	
    34	
    35		BLASLONG register i = 0;
    36	
    37		__asm__  __volatile__
    38		(
    39		"vbroadcastss		(%4), %%ymm0		    \n\t"  // alpha	
    40	
    41		".p2align 4				            \n\t"
    42		"1:				            \n\t"
    43	
    44	        "vmovups                  (%3,%0,4), %%ymm12         \n\t"  // 8 * y
    45	        "vmovups                32(%3,%0,4), %%ymm13         \n\t"  // 8 * y
    46	        "vmovups                64(%3,%0,4), %%ymm14         \n\t"  // 8 * y
    47	        "vmovups                96(%3,%0,4), %%ymm15         \n\t"  // 8 * y
    48		"vfmadd231ps       (%2,%0,4), %%ymm0  , %%ymm12  	     \n\t"   // y += alpha * x
    49		"vfmadd231ps     32(%2,%0,4), %%ymm0  , %%ymm13  	     \n\t"   // y += alpha * x
    50		"vfmadd231ps     64(%2,%0,4), %%ymm0  , %%ymm14  	     \n\t"   // y += alpha * x
    51		"vfmadd231ps     96(%2,%0,4), %%ymm0  , %%ymm15  	     \n\t"   // y += alpha * x
    52		"vmovups	%%ymm12,   (%3,%0,4)		     \n\t"
    53		"vmovups	%%ymm13, 32(%3,%0,4)		     \n\t"
    54		"vmovups	%%ymm14, 64(%3,%0,4)		     \n\t"
    55		"vmovups	%%ymm15, 96(%3,%0,4)		     \n\t"
    56	
    57		"addq		$32, %0	  	 	             \n\t"
    58		"subq	        $32, %1			             \n\t"		
    59		"jnz		1b		             \n\t"
    60		"vzeroupper				     \n\t"
    61	
    62		:
    63	        : 
    64	          "r" (i),	// 0	
    65		  "r" (n),  	// 1
    66	          "r" (x),      // 2
    67	          "r" (y),      // 3
    68	          "r" (alpha)   // 4
    69		: "cc", 
    70		  "%xmm0", 
    71		  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
    72		  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
    73		  "memory"
    74		);
    75	

Inputs 0 and 1 are marked as read-only, but are modified at lines 57 and 58.
Thus one needs something like:

diff --git a/kernel/x86_64/saxpy_microk_haswell-2.c b/kernel/x86_64/saxpy_microk_haswell-2.c
index 3a743d64..7994735b 100644
--- a/kernel/x86_64/saxpy_microk_haswell-2.c
+++ b/kernel/x86_64/saxpy_microk_haswell-2.c
@@ -60,9 +60,9 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
        "vzeroupper                                  \n\t"
 
        :
+          "+r" (i),    // 0    
+         "+r" (n)      // 1
         : 
-          "r" (i),     // 0    
-         "r" (n),      // 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (alpha)   // 4

Note that similar problem may be in other implementations (or asm statements).
Thanks

[martin-frbg]

Thanks. I vaguely remember fixing something very similar a year ago (ah yes, #1292) in response to more aggressive reuse of registers by gcc8, it seems I never got around to fixing more than the gemv kernels although I noticed that there were around 100 suspect files.

[peter-bergner]

FYI, GCC 9 will be even more aggressive at reusing registers than GCC 8.

[martin-frbg]

I guess so. I am about halfway through with fixing AXPY and DOT already, then to see what else has this defect.

[marxin]

Thanks for the fixes. When do you expect it will land in a release?

[martin-frbg]

No release date set yet, probably end of February.