Merge pull request #1409 from martin-frbg/issue1292-2

martin-frbg · web-flow · commit b81656936fca · 2017-12-31T20:18:48.000+01:00
Tag %1 and %2 as both input and output operands
diff --git a/kernel/x86_64/cgemv_n_microk_bulldozer-4.c b/kernel/x86_64/cgemv_n_microk_bulldozer-4.c
@@ -158,9 +158,9 @@ static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vzeroupper			 \n\t"
 
 	:
-        : 
-          "r" (i),	// 0	
-	  "r" (n1),  	// 1
+          "+r" (i),	// 0	
+	  "+r" (n1)  	// 1
+	:
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -282,9 +282,9 @@ static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vzeroupper			 \n\t"
 
 	:
-        : 
-          "r" (i),	// 0	
-	  "r" (n1),  	// 1
+          "+r" (i),	// 0	
+	  "+r" (n1)  	// 1
+	:
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -389,9 +389,9 @@ static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 	"vzeroupper			 \n\t"
 
 	:
-        : 
-          "r" (i),	// 0	
-	  "r" (n1),  	// 1
+          "+r" (i),	// 0	
+	  "+r" (n1)  	// 1
+	:
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap),     // 4
@@ -519,9 +519,9 @@ static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT a
 	"vzeroupper			 \n\t"
 
 	:
-        : 
-          "r" (i),	      // 0	
-	  "r" (n1),  	      // 1
+          "+r" (i),	      // 0	
+	  "+r" (n1)  	      // 1
+	:
           "r" (src),          // 2
           "r" (dest),         // 3
           "r" (&alpha_r),     // 4
diff --git a/kernel/x86_64/cgemv_t_microk_bulldozer-4.c b/kernel/x86_64/cgemv_t_microk_bulldozer-4.c
@@ -231,9 +231,9 @@ static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -392,9 +392,9 @@ static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -521,9 +521,9 @@ static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap),     // 4
diff --git a/kernel/x86_64/dgemv_n_microk_nehalem-4.c b/kernel/x86_64/dgemv_n_microk_nehalem-4.c
@@ -149,9 +149,9 @@ static void dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"movups         %%xmm5 ,  -16(%3,%0,8)	       \n\t"	// 2 * y
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
diff --git a/kernel/x86_64/dgemv_n_microk_piledriver-4.c b/kernel/x86_64/dgemv_n_microk_piledriver-4.c
@@ -124,9 +124,9 @@ static void dgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
 	"vzeroupper			        \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -224,9 +224,9 @@ static void dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"vzeroupper			              \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
diff --git a/kernel/x86_64/sgemv_n_microk_bulldozer-4.c b/kernel/x86_64/sgemv_n_microk_bulldozer-4.c
@@ -183,9 +183,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
 	"4:                             \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -246,9 +246,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"jnz		1b		      	\n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
diff --git a/kernel/x86_64/sgemv_n_microk_nehalem-4.c b/kernel/x86_64/sgemv_n_microk_nehalem-4.c
@@ -106,9 +106,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
 	"jnz		1b		       \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -181,9 +181,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"jnz		1b		       \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
diff --git a/kernel/x86_64/sgemv_n_microk_sandy-4.c b/kernel/x86_64/sgemv_n_microk_sandy-4.c
@@ -196,9 +196,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
 	"vzeroupper			          \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -345,9 +345,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"vzeroupper			          \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
diff --git a/kernel/x86_64/sgemv_t_microk_bulldozer-4.c b/kernel/x86_64/sgemv_t_microk_bulldozer-4.c
@@ -126,9 +126,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vmovss		%%xmm7,  12(%3)		\n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
diff --git a/kernel/x86_64/sgemv_t_microk_nehalem-4.c b/kernel/x86_64/sgemv_t_microk_nehalem-4.c
@@ -78,9 +78,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
         "movss         %%xmm7,  12(%3)         \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
diff --git a/kernel/x86_64/sgemv_t_microk_sandy-4.c b/kernel/x86_64/sgemv_t_microk_sandy-4.c
@@ -152,9 +152,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
diff --git a/kernel/x86_64/zgemv_n_microk_bulldozer-4.c b/kernel/x86_64/zgemv_n_microk_bulldozer-4.c
@@ -125,9 +125,9 @@ static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"2:					\n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -231,9 +231,9 @@ static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"2:					\n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -321,9 +321,9 @@ static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -393,9 +393,9 @@ static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap)      // 4
@@ -493,9 +493,9 @@ static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT a
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	      // 0	
+	  "+r" (n)  	      // 1
         : 
-          "r" (i),	      // 0	
-	  "r" (n),  	      // 1
           "r" (src),          // 2
           "r" (dest),         // 3
           "r" (&alpha_r),     // 4
diff --git a/kernel/x86_64/zgemv_n_microk_sandy-4.c b/kernel/x86_64/zgemv_n_microk_sandy-4.c
@@ -127,9 +127,9 @@ static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -220,9 +220,9 @@ static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -295,9 +295,9 @@ static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap)      // 4
@@ -396,9 +396,9 @@ static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT a
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	      // 0	
+	  "+r" (n)  	      // 1
         : 
-          "r" (i),	      // 0	
-	  "r" (n),  	      // 1
           "r" (src),          // 2
           "r" (dest),         // 3
           "r" (&alpha_r),     // 4
diff --git a/kernel/x86_64/zgemv_t_microk_bulldozer-4.c b/kernel/x86_64/zgemv_t_microk_bulldozer-4.c
@@ -198,9 +198,9 @@ static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -333,9 +333,9 @@ static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap[0]),  // 4
@@ -437,9 +437,9 @@ static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *
 	"vzeroupper			 \n\t"
 
 	:
+          "+r" (i),	// 0	
+	  "+r" (n)  	// 1
         : 
-          "r" (i),	// 0	
-	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
           "r" (ap),     // 4

Original file line number	Diff line number	Diff line change
`@@ -231,9 +231,9 @@ static void cgemv_kernel_4x4( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, FLOAT`
`231`	`231`	`"vzeroupper \n\t"`
`232`	`232`
`233`	`233`	`:`
	`234`	`+ "+r" (i), // 0`
	`235`	`+ "+r" (n) // 1`
`234`	`236`	`:`
`235`		`- "r" (i), // 0`
`236`		`- "r" (n), // 1`
`237`	`237`	`"r" (x), // 2`
`238`	`238`	`"r" (y), // 3`
`239`	`239`	`"r" (ap[0]), // 4`
`@@ -392,9 +392,9 @@ static void cgemv_kernel_4x2( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, FLOAT`
`392`	`392`	`"vzeroupper \n\t"`
`393`	`393`
`394`	`394`	`:`
	`395`	`+ "+r" (i), // 0`
	`396`	`+ "+r" (n) // 1`
`395`	`397`	`:`
`396`		`- "r" (i), // 0`
`397`		`- "r" (n), // 1`
`398`	`398`	`"r" (x), // 2`
`399`	`399`	`"r" (y), // 3`
`400`	`400`	`"r" (ap[0]), // 4`
`@@ -521,9 +521,9 @@ static void cgemv_kernel_4x1( BLASLONG n, FLOAT ap, FLOAT x, FLOAT y, FLOAT `
`521`	`521`	`"vzeroupper \n\t"`
`522`	`522`
`523`	`523`	`:`
	`524`	`+ "+r" (i), // 0`
	`525`	`+ "+r" (n) // 1`
`524`	`526`	`:`
`525`		`- "r" (i), // 0`
`526`		`- "r" (n), // 1`
`527`	`527`	`"r" (x), // 2`
`528`	`528`	`"r" (y), // 3`
`529`	`529`	`"r" (ap), // 4`
Original file line number	Diff line number	Diff line change
`@@ -149,9 +149,9 @@ static void dgemv_kernel_4x4( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, FLOAT`
`149`	`149`	`"movups %%xmm5 , -16(%3,%0,8) \n\t" // 2 * y`
`150`	`150`
`151`	`151`	`:`
	`152`	`+ "+r" (i), // 0`
	`153`	`+ "+r" (n) // 1`
`152`	`154`	`:`
`153`		`- "r" (i), // 0`
`154`		`- "r" (n), // 1`
`155`	`155`	`"r" (x), // 2`
`156`	`156`	`"r" (y), // 3`
`157`	`157`	`"r" (ap[0]), // 4`
Original file line number	Diff line number	Diff line change
`@@ -124,9 +124,9 @@ static void dgemv_kernel_4x8( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, BLASLO`
`124`	`124`	`"vzeroupper \n\t"`
`125`	`125`
`126`	`126`	`:`
	`127`	`+ "+r" (i), // 0`
	`128`	`+ "+r" (n) // 1`
`127`	`129`	`:`
`128`		`- "r" (i), // 0`
`129`		`- "r" (n), // 1`
`130`	`130`	`"r" (x), // 2`
`131`	`131`	`"r" (y), // 3`
`132`	`132`	`"r" (ap[0]), // 4`
`@@ -224,9 +224,9 @@ static void dgemv_kernel_4x4( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, FLOAT`
`224`	`224`	`"vzeroupper \n\t"`
`225`	`225`
`226`	`226`	`:`
	`227`	`+ "+r" (i), // 0`
	`228`	`+ "+r" (n) // 1`
`227`	`229`	`:`
`228`		`- "r" (i), // 0`
`229`		`- "r" (n), // 1`
`230`	`230`	`"r" (x), // 2`
`231`	`231`	`"r" (y), // 3`
`232`	`232`	`"r" (ap[0]), // 4`
Original file line number	Diff line number	Diff line change
`@@ -183,9 +183,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, BLASLO`
`183`	`183`	`"4: \n\t"`
`184`	`184`
`185`	`185`	`:`
	`186`	`+ "+r" (i), // 0`
	`187`	`+ "+r" (n) // 1`
`186`	`188`	`:`
`187`		`- "r" (i), // 0`
`188`		`- "r" (n), // 1`
`189`	`189`	`"r" (x), // 2`
`190`	`190`	`"r" (y), // 3`
`191`	`191`	`"r" (ap[0]), // 4`
`@@ -246,9 +246,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, FLOAT`
`246`	`246`	`"jnz 1b \n\t"`
`247`	`247`
`248`	`248`	`:`
	`249`	`+ "+r" (i), // 0`
	`250`	`+ "+r" (n) // 1`
`249`	`251`	`:`
`250`		`- "r" (i), // 0`
`251`		`- "r" (n), // 1`
`252`	`252`	`"r" (x), // 2`
`253`	`253`	`"r" (y), // 3`
`254`	`254`	`"r" (ap[0]), // 4`
Original file line number	Diff line number	Diff line change
`@@ -106,9 +106,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, BLASLO`
`106`	`106`	`"jnz 1b \n\t"`
`107`	`107`
`108`	`108`	`:`
	`109`	`+ "+r" (i), // 0`
	`110`	`+ "+r" (n) // 1`
`109`	`111`	`:`
`110`		`- "r" (i), // 0`
`111`		`- "r" (n), // 1`
`112`	`112`	`"r" (x), // 2`
`113`	`113`	`"r" (y), // 3`
`114`	`114`	`"r" (ap[0]), // 4`
`@@ -181,9 +181,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, FLOAT`
`181`	`181`	`"jnz 1b \n\t"`
`182`	`182`
`183`	`183`	`:`
	`184`	`+ "+r" (i), // 0`
	`185`	`+ "+r" (n) // 1`
`184`	`186`	`:`
`185`		`- "r" (i), // 0`
`186`		`- "r" (n), // 1`
`187`	`187`	`"r" (x), // 2`
`188`	`188`	`"r" (y), // 3`
`189`	`189`	`"r" (ap[0]), // 4`
Original file line number	Diff line number	Diff line change
`@@ -196,9 +196,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, BLASLO`
`196`	`196`	`"vzeroupper \n\t"`
`197`	`197`
`198`	`198`	`:`
	`199`	`+ "+r" (i), // 0`
	`200`	`+ "+r" (n) // 1`
`199`	`201`	`:`
`200`		`- "r" (i), // 0`
`201`		`- "r" (n), // 1`
`202`	`202`	`"r" (x), // 2`
`203`	`203`	`"r" (y), // 3`
`204`	`204`	`"r" (ap[0]), // 4`
`@@ -345,9 +345,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y, FLOAT`
`345`	`345`	`"vzeroupper \n\t"`
`346`	`346`
`347`	`347`	`:`
	`348`	`+ "+r" (i), // 0`
	`349`	`+ "+r" (n) // 1`
`348`	`350`	`:`
`349`		`- "r" (i), // 0`
`350`		`- "r" (n), // 1`
`351`	`351`	`"r" (x), // 2`
`352`	`352`	`"r" (y), // 3`
`353`	`353`	`"r" (ap[0]), // 4`
Original file line number	Diff line number	Diff line change
`@@ -126,9 +126,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y)`
`126`	`126`	`"vmovss %%xmm7, 12(%3) \n\t"`
`127`	`127`
`128`	`128`	`:`
	`129`	`+ "+r" (i), // 0`
	`130`	`+ "+r" (n) // 1`
`129`	`131`	`:`
`130`		`- "r" (i), // 0`
`131`		`- "r" (n), // 1`
`132`	`132`	`"r" (x), // 2`
`133`	`133`	`"r" (y), // 3`
`134`	`134`	`"r" (ap[0]), // 4`
Original file line number	Diff line number	Diff line change
`@@ -78,9 +78,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y)`
`78`	`78`	`"movss %%xmm7, 12(%3) \n\t"`
`79`	`79`
`80`	`80`	`:`
	`81`	`+ "+r" (i), // 0`
	`82`	`+ "+r" (n) // 1`
`81`	`83`	`:`
`82`		`- "r" (i), // 0`
`83`		`- "r" (n), // 1`
`84`	`84`	`"r" (x), // 2`
`85`	`85`	`"r" (y), // 3`
`86`	`86`	`"r" (ap[0]), // 4`
Original file line number	Diff line number	Diff line change
`@@ -152,9 +152,9 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT *ap, FLOAT x, FLOAT *y)`
`152`	`152`	`"vzeroupper \n\t"`
`153`	`153`
`154`	`154`	`:`
	`155`	`+ "+r" (i), // 0`
	`156`	`+ "+r" (n) // 1`
`155`	`157`	`:`
`156`		`- "r" (i), // 0`
`157`		`- "r" (n), // 1`
`158`	`158`	`"r" (x), // 2`
`159`	`159`	`"r" (y), // 3`
`160`	`160`	`"r" (ap[0]), // 4`