From 46e415b1405044b038586537d213e4f2f04b8536 Mon Sep 17 00:00:00 2001 From: Martin Kroeker Date: Thu, 14 Feb 2019 22:43:18 +0100 Subject: [PATCH 1/2] Save and restore input argument 8 (lda4) Fixes miscompilation with gcc9 -ftree-vectorize (related to issue #2009) --- kernel/x86_64/sgemv_n_microk_haswell-4.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/kernel/x86_64/sgemv_n_microk_haswell-4.c b/kernel/x86_64/sgemv_n_microk_haswell-4.c index 2c90f8aa9..e89a16785 100644 --- a/kernel/x86_64/sgemv_n_microk_haswell-4.c +++ b/kernel/x86_64/sgemv_n_microk_haswell-4.c @@ -26,7 +26,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. *****************************************************************************/ - #define HAVE_KERNEL_4x8 1 static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLONG lda4, FLOAT *alpha) __attribute__ ((noinline)); @@ -49,6 +48,8 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO "vbroadcastss (%9), %%ymm6 \n\t" // alpha + "movq %8, %%xmm10 \n\t" //save lda + "testq $0x04, %1 \n\t" "jz 2f \n\t" @@ -151,6 +152,7 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO "4: \n\t" "vzeroupper \n\t" + "movq %%xmm10, %8 \n\t" //restore lda : "+r" (i), // 0 @@ -170,6 +172,7 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO "%xmm4", "%xmm5", "%xmm6", "%xmm7", "%xmm8", "%xmm9", + "%xmm10", "%xmm12", "%xmm13", "%xmm14", "%xmm15", "memory" ); @@ -177,7 +180,6 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO } - #define HAVE_KERNEL_4x4 1 static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline)); @@ -196,6 +198,7 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT "vbroadcastss (%8), %%ymm6 \n\t" // alpha + "testq $0x04, %1 \n\t" "jz 2f \n\t" From 4255a58cd22d5395dbd6573683298849bd3a23b5 Mon Sep 17 00:00:00 2001 From: Martin Kroeker Date: Fri, 15 Feb 2019 10:10:04 +0100 Subject: [PATCH 2/2] Rename operands to put lda on the input/output constraint list --- kernel/x86_64/sgemv_n_microk_haswell-4.c | 126 +++++++++++------------ 1 file changed, 61 insertions(+), 65 deletions(-) diff --git a/kernel/x86_64/sgemv_n_microk_haswell-4.c b/kernel/x86_64/sgemv_n_microk_haswell-4.c index e89a16785..93e1e26e8 100644 --- a/kernel/x86_64/sgemv_n_microk_haswell-4.c +++ b/kernel/x86_64/sgemv_n_microk_haswell-4.c @@ -37,43 +37,41 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO __asm__ __volatile__ ( "vzeroupper \n\t" - "vbroadcastss (%2), %%ymm12 \n\t" // x0 - "vbroadcastss 4(%2), %%ymm13 \n\t" // x1 - "vbroadcastss 8(%2), %%ymm14 \n\t" // x2 - "vbroadcastss 12(%2), %%ymm15 \n\t" // x3 - "vbroadcastss 16(%2), %%ymm0 \n\t" // x4 - "vbroadcastss 20(%2), %%ymm1 \n\t" // x5 - "vbroadcastss 24(%2), %%ymm2 \n\t" // x6 - "vbroadcastss 28(%2), %%ymm3 \n\t" // x7 + "vbroadcastss (%3), %%ymm12 \n\t" // x0 + "vbroadcastss 4(%3), %%ymm13 \n\t" // x1 + "vbroadcastss 8(%3), %%ymm14 \n\t" // x2 + "vbroadcastss 12(%3), %%ymm15 \n\t" // x3 + "vbroadcastss 16(%3), %%ymm0 \n\t" // x4 + "vbroadcastss 20(%3), %%ymm1 \n\t" // x5 + "vbroadcastss 24(%3), %%ymm2 \n\t" // x6 + "vbroadcastss 28(%3), %%ymm3 \n\t" // x7 "vbroadcastss (%9), %%ymm6 \n\t" // alpha - "movq %8, %%xmm10 \n\t" //save lda - "testq $0x04, %1 \n\t" "jz 2f \n\t" - "vmovups (%3,%0,4), %%xmm7 \n\t" // 4 * y + "vmovups (%4,%0,4), %%xmm7 \n\t" // 4 * y "vxorps %%xmm4 , %%xmm4, %%xmm4 \n\t" "vxorps %%xmm5 , %%xmm5, %%xmm5 \n\t" - "vfmadd231ps (%4,%0,4), %%xmm12, %%xmm4 \n\t" - "vfmadd231ps (%5,%0,4), %%xmm13, %%xmm5 \n\t" - "vfmadd231ps (%6,%0,4), %%xmm14, %%xmm4 \n\t" - "vfmadd231ps (%7,%0,4), %%xmm15, %%xmm5 \n\t" + "vfmadd231ps (%5,%0,4), %%xmm12, %%xmm4 \n\t" + "vfmadd231ps (%6,%0,4), %%xmm13, %%xmm5 \n\t" + "vfmadd231ps (%7,%0,4), %%xmm14, %%xmm4 \n\t" + "vfmadd231ps (%8,%0,4), %%xmm15, %%xmm5 \n\t" - "vfmadd231ps (%4,%8,4), %%xmm0 , %%xmm4 \n\t" - "vfmadd231ps (%5,%8,4), %%xmm1 , %%xmm5 \n\t" - "vfmadd231ps (%6,%8,4), %%xmm2 , %%xmm4 \n\t" - "vfmadd231ps (%7,%8,4), %%xmm3 , %%xmm5 \n\t" + "vfmadd231ps (%5,%2,4), %%xmm0 , %%xmm4 \n\t" + "vfmadd231ps (%6,%2,4), %%xmm1 , %%xmm5 \n\t" + "vfmadd231ps (%7,%2,4), %%xmm2 , %%xmm4 \n\t" + "vfmadd231ps (%8,%2,4), %%xmm3 , %%xmm5 \n\t" "vaddps %%xmm4 , %%xmm5 , %%xmm5 \n\t" "vmulps %%xmm6 , %%xmm5 , %%xmm5 \n\t" "vaddps %%xmm7 , %%xmm5 , %%xmm5 \n\t" - "vmovups %%xmm5, (%3,%0,4) \n\t" // 4 * y + "vmovups %%xmm5, (%4,%0,4) \n\t" // 4 * y - "addq $4 , %8 \n\t" + "addq $4 , %2 \n\t" "addq $4 , %0 \n\t" "subq $4 , %1 \n\t" @@ -82,28 +80,28 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO "testq $0x08, %1 \n\t" "jz 3f \n\t" - "vmovups (%3,%0,4), %%ymm7 \n\t" // 8 * y + "vmovups (%4,%0,4), %%ymm7 \n\t" // 8 * y "vxorps %%ymm4 , %%ymm4, %%ymm4 \n\t" "vxorps %%ymm5 , %%ymm5, %%ymm5 \n\t" - "vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t" - "vfmadd231ps (%5,%0,4), %%ymm13, %%ymm5 \n\t" - "vfmadd231ps (%6,%0,4), %%ymm14, %%ymm4 \n\t" - "vfmadd231ps (%7,%0,4), %%ymm15, %%ymm5 \n\t" + "vfmadd231ps (%5,%0,4), %%ymm12, %%ymm4 \n\t" + "vfmadd231ps (%6,%0,4), %%ymm13, %%ymm5 \n\t" + "vfmadd231ps (%7,%0,4), %%ymm14, %%ymm4 \n\t" + "vfmadd231ps (%8,%0,4), %%ymm15, %%ymm5 \n\t" - "vfmadd231ps (%4,%8,4), %%ymm0 , %%ymm4 \n\t" - "vfmadd231ps (%5,%8,4), %%ymm1 , %%ymm5 \n\t" - "vfmadd231ps (%6,%8,4), %%ymm2 , %%ymm4 \n\t" - "vfmadd231ps (%7,%8,4), %%ymm3 , %%ymm5 \n\t" + "vfmadd231ps (%5,%2,4), %%ymm0 , %%ymm4 \n\t" + "vfmadd231ps (%6,%2,4), %%ymm1 , %%ymm5 \n\t" + "vfmadd231ps (%7,%2,4), %%ymm2 , %%ymm4 \n\t" + "vfmadd231ps (%8,%2,4), %%ymm3 , %%ymm5 \n\t" "vaddps %%ymm4 , %%ymm5 , %%ymm5 \n\t" "vmulps %%ymm6 , %%ymm5 , %%ymm5 \n\t" "vaddps %%ymm7 , %%ymm5 , %%ymm5 \n\t" - "vmovups %%ymm5, (%3,%0,4) \n\t" // 8 * y + "vmovups %%ymm5, (%4,%0,4) \n\t" // 8 * y - "addq $8 , %8 \n\t" + "addq $8 , %2 \n\t" "addq $8 , %0 \n\t" "subq $8 , %1 \n\t" @@ -118,53 +116,52 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO "vxorps %%ymm4 , %%ymm4, %%ymm4 \n\t" "vxorps %%ymm5 , %%ymm5, %%ymm5 \n\t" - "vmovups (%3,%0,4), %%ymm8 \n\t" // 8 * y - "vmovups 32(%3,%0,4), %%ymm9 \n\t" // 8 * y - - "vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t" - "vfmadd231ps 32(%4,%0,4), %%ymm12, %%ymm5 \n\t" - "vfmadd231ps (%5,%0,4), %%ymm13, %%ymm4 \n\t" - "vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5 \n\t" - "vfmadd231ps (%6,%0,4), %%ymm14, %%ymm4 \n\t" - "vfmadd231ps 32(%6,%0,4), %%ymm14, %%ymm5 \n\t" - "vfmadd231ps (%7,%0,4), %%ymm15, %%ymm4 \n\t" - "vfmadd231ps 32(%7,%0,4), %%ymm15, %%ymm5 \n\t" - - "vfmadd231ps (%4,%8,4), %%ymm0 , %%ymm4 \n\t" + "vmovups (%4,%0,4), %%ymm8 \n\t" // 8 * y + "vmovups 32(%4,%0,4), %%ymm9 \n\t" // 8 * y + + "vfmadd231ps (%5,%0,4), %%ymm12, %%ymm4 \n\t" + "vfmadd231ps 32(%5,%0,4), %%ymm12, %%ymm5 \n\t" + "vfmadd231ps (%6,%0,4), %%ymm13, %%ymm4 \n\t" + "vfmadd231ps 32(%6,%0,4), %%ymm13, %%ymm5 \n\t" + "vfmadd231ps (%7,%0,4), %%ymm14, %%ymm4 \n\t" + "vfmadd231ps 32(%7,%0,4), %%ymm14, %%ymm5 \n\t" + "vfmadd231ps (%8,%0,4), %%ymm15, %%ymm4 \n\t" + "vfmadd231ps 32(%8,%0,4), %%ymm15, %%ymm5 \n\t" + + "vfmadd231ps (%5,%2,4), %%ymm0 , %%ymm4 \n\t" "addq $16, %0 \n\t" - "vfmadd231ps 32(%4,%8,4), %%ymm0 , %%ymm5 \n\t" - "vfmadd231ps (%5,%8,4), %%ymm1 , %%ymm4 \n\t" - "vfmadd231ps 32(%5,%8,4), %%ymm1 , %%ymm5 \n\t" - "vfmadd231ps (%6,%8,4), %%ymm2 , %%ymm4 \n\t" - "vfmadd231ps 32(%6,%8,4), %%ymm2 , %%ymm5 \n\t" - "vfmadd231ps (%7,%8,4), %%ymm3 , %%ymm4 \n\t" - "vfmadd231ps 32(%7,%8,4), %%ymm3 , %%ymm5 \n\t" + "vfmadd231ps 32(%5,%2,4), %%ymm0 , %%ymm5 \n\t" + "vfmadd231ps (%6,%2,4), %%ymm1 , %%ymm4 \n\t" + "vfmadd231ps 32(%6,%2,4), %%ymm1 , %%ymm5 \n\t" + "vfmadd231ps (%7,%2,4), %%ymm2 , %%ymm4 \n\t" + "vfmadd231ps 32(%7,%2,4), %%ymm2 , %%ymm5 \n\t" + "vfmadd231ps (%8,%2,4), %%ymm3 , %%ymm4 \n\t" + "vfmadd231ps 32(%8,%2,4), %%ymm3 , %%ymm5 \n\t" "vfmadd231ps %%ymm6 , %%ymm4 , %%ymm8 \n\t" "vfmadd231ps %%ymm6 , %%ymm5 , %%ymm9 \n\t" - "addq $16, %8 \n\t" - "vmovups %%ymm8,-64(%3,%0,4) \n\t" // 8 * y + "addq $16, %2 \n\t" + "vmovups %%ymm8,-64(%4,%0,4) \n\t" // 8 * y "subq $16, %1 \n\t" - "vmovups %%ymm9,-32(%3,%0,4) \n\t" // 8 * y + "vmovups %%ymm9,-32(%4,%0,4) \n\t" // 8 * y "jnz 1b \n\t" "4: \n\t" "vzeroupper \n\t" - "movq %%xmm10, %8 \n\t" //restore lda : "+r" (i), // 0 - "+r" (n) // 1 + "+r" (n), // 1 + "+r" (lda4) // 2 : - "r" (x), // 2 - "r" (y), // 3 - "r" (ap[0]), // 4 - "r" (ap[1]), // 5 - "r" (ap[2]), // 6 - "r" (ap[3]), // 7 - "r" (lda4), // 8 + "r" (x), // 3 + "r" (y), // 4 + "r" (ap[0]), // 5 + "r" (ap[1]), // 6 + "r" (ap[2]), // 7 + "r" (ap[3]), // 8 "r" (alpha) // 9 : "cc", "%xmm0", "%xmm1", @@ -172,7 +169,6 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO "%xmm4", "%xmm5", "%xmm6", "%xmm7", "%xmm8", "%xmm9", - "%xmm10", "%xmm12", "%xmm13", "%xmm14", "%xmm15", "memory" );