Optimize restoring of ymm registers on x86-64.

The patch mainly reduces the code size but also avoids some jumps.

diff --git a/ChangeLog b/ChangeLog
index 87db19e..1bfdd7b 100644 (file)
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,8 @@
+2009-07-16  Ulrich Drepper  <drepper@redhat.com>
+
+       * sysdeps/x86_64/dl-trampoline.S (_dl_runtime_profile): Optimize
+       restoring of ymm registers a bit.
+
 2009-07-15  H.J. Lu  <hongjiu.lu@intel.com>
 
        * sysdeps/x86_64/memcmp.S: New file.

diff --git a/sysdeps/x86_64/dl-trampoline.S b/sysdeps/x86_64/dl-trampoline.S
index 7f20491..49d239f 100644 (file)
--- a/sysdeps/x86_64/dl-trampoline.S
+++ b/sysdeps/x86_64/dl-trampoline.S
@@ -185,81 +185,73 @@ L(no_avx1):
        movq  LR_R8_OFFSET(%rsp), %r8
        movq  LR_R9_OFFSET(%rsp), %r9
 
+       movaps              (LR_XMM_OFFSET)(%rsp), %xmm0
+       movaps   (LR_XMM_OFFSET + XMM_SIZE)(%rsp), %xmm1
+       movaps (LR_XMM_OFFSET + XMM_SIZE*2)(%rsp), %xmm2
+       movaps (LR_XMM_OFFSET + XMM_SIZE*3)(%rsp), %xmm3
+       movaps (LR_XMM_OFFSET + XMM_SIZE*4)(%rsp), %xmm4
+       movaps (LR_XMM_OFFSET + XMM_SIZE*5)(%rsp), %xmm5
+       movaps (LR_XMM_OFFSET + XMM_SIZE*6)(%rsp), %xmm6
+       movaps (LR_XMM_OFFSET + XMM_SIZE*7)(%rsp), %xmm7
+
 # ifdef HAVE_AVX_SUPPORT
        cmpl    $0, L(have_avx)(%rip)
        js      L(no_avx2)
 
        /* Check if any xmm0-xmm7 registers are changed by audit
           module.  */
-       vmovdqa (LR_XMM_OFFSET)(%rsp), %xmm0
-       vpcmpeqq (LR_SIZE)(%rsp), %xmm0, %xmm1
-       vpmovmskb %xmm1, %esi
+       vpcmpeqq (LR_SIZE)(%rsp), %xmm0, %xmm8
+       vpmovmskb %xmm8, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu                 (LR_VECTOR_OFFSET)(%rsp), %ymm0
 
-1:     vmovdqa (LR_XMM_OFFSET + XMM_SIZE)(%rsp), %xmm1
-       vpcmpeqq (LR_SIZE + XMM_SIZE)(%rsp), %xmm1, %xmm2
-       vpmovmskb %xmm2, %esi
+1:     vpcmpeqq (LR_SIZE + XMM_SIZE)(%rsp), %xmm1, %xmm8
+       vpmovmskb %xmm8, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu   (LR_VECTOR_OFFSET + VECTOR_SIZE)(%rsp), %ymm1
 
-1:     vmovdqa (LR_XMM_OFFSET + XMM_SIZE*2)(%rsp), %xmm2
-       vpcmpeqq (LR_SIZE + XMM_SIZE*2)(%rsp), %xmm2, %xmm3
-       vpmovmskb %xmm3, %esi
+1:     vpcmpeqq (LR_SIZE + XMM_SIZE*2)(%rsp), %xmm2, %xmm8
+       vpmovmskb %xmm8, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*2)(%rsp), %ymm2
 
-1:     vmovdqa (LR_XMM_OFFSET + XMM_SIZE*3)(%rsp), %xmm3
-       vpcmpeqq (LR_SIZE + XMM_SIZE*3)(%rsp), %xmm3, %xmm4
-       vpmovmskb %xmm4, %esi
+1:     vpcmpeqq (LR_SIZE + XMM_SIZE*3)(%rsp), %xmm3, %xmm8
+       vpmovmskb %xmm8, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*3)(%rsp), %ymm3
 
-1:     vmovdqa (LR_XMM_OFFSET + XMM_SIZE*4)(%rsp), %xmm4
-       vpcmpeqq (LR_SIZE + XMM_SIZE*4)(%rsp), %xmm4, %xmm5
-       vpmovmskb %xmm5, %esi
+1:     vpcmpeqq (LR_SIZE + XMM_SIZE*4)(%rsp), %xmm4, %xmm8
+       vpmovmskb %xmm8, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*4)(%rsp), %ymm4
 
-1:     vmovdqa (LR_XMM_OFFSET + XMM_SIZE*5)(%rsp), %xmm5
-       vpcmpeqq (LR_SIZE + XMM_SIZE*5)(%rsp), %xmm5, %xmm6
-       vpmovmskb %xmm6, %esi
+1:     vpcmpeqq (LR_SIZE + XMM_SIZE*5)(%rsp), %xmm5, %xmm8
+       vpmovmskb %xmm8, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*5)(%rsp), %ymm5
 
-1:     vmovdqa (LR_XMM_OFFSET + XMM_SIZE*6)(%rsp), %xmm6
-       vpcmpeqq (LR_SIZE + XMM_SIZE*6)(%rsp), %xmm6, %xmm7
-       vpmovmskb %xmm7, %esi
+1:     vpcmpeqq (LR_SIZE + XMM_SIZE*6)(%rsp), %xmm6, %xmm8
+       vpmovmskb %xmm8, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*6)(%rsp), %ymm6
 
-1:     vmovdqa (LR_XMM_OFFSET + XMM_SIZE*7)(%rsp), %xmm7
-       vpcmpeqq (LR_SIZE + XMM_SIZE*7)(%rsp), %xmm7, %xmm8
+1:     vpcmpeqq (LR_SIZE + XMM_SIZE*7)(%rsp), %xmm7, %xmm8
        vpmovmskb %xmm8, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*7)(%rsp), %ymm7
-       jmp     1f
 
 L(no_avx2):
+1:
 # endif
-       movaps              (LR_XMM_OFFSET)(%rsp), %xmm0
-       movaps   (LR_XMM_OFFSET + XMM_SIZE)(%rsp), %xmm1
-       movaps (LR_XMM_OFFSET + XMM_SIZE*2)(%rsp), %xmm2
-       movaps (LR_XMM_OFFSET + XMM_SIZE*3)(%rsp), %xmm3
-       movaps (LR_XMM_OFFSET + XMM_SIZE*4)(%rsp), %xmm4
-       movaps (LR_XMM_OFFSET + XMM_SIZE*5)(%rsp), %xmm5
-       movaps (LR_XMM_OFFSET + XMM_SIZE*6)(%rsp), %xmm6
-       movaps (LR_XMM_OFFSET + XMM_SIZE*7)(%rsp), %xmm7
-
-1:     movq 16(%rbx), %r10     # Anything in framesize?
+       movq 16(%rbx), %r10     # Anything in framesize?
        testq %r10, %r10
        jns 3f
 
@@ -358,32 +350,31 @@ L(no_avx3):
        movq LRV_RAX_OFFSET(%rsp), %rax
        movq LRV_RDX_OFFSET(%rsp), %rdx
 
+       movaps LRV_XMM0_OFFSET(%rsp), %xmm0
+       movaps LRV_XMM1_OFFSET(%rsp), %xmm1
+
 # ifdef HAVE_AVX_SUPPORT
        cmpl    $0, L(have_avx)(%rip)
        js      L(no_avx4)
 
        /* Check if xmm0/xmm1 registers are changed by audit module.  */
-       vmovdqa LRV_XMM0_OFFSET(%rsp), %xmm0
-       vpcmpeqq (LRV_SIZE)(%rsp), %xmm0, %xmm1
-       vpmovmskb %xmm1, %esi
+       vpcmpeqq (LRV_SIZE)(%rsp), %xmm0, %xmm2
+       vpmovmskb %xmm2, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu LRV_VECTOR0_OFFSET(%rsp), %ymm0
 
-1:     vmovdqa LRV_XMM1_OFFSET(%rsp), %xmm1
-       vpcmpeqq (LRV_SIZE + XMM_SIZE)(%rsp), %xmm1, %xmm2
+1:     vpcmpeqq (LRV_SIZE + XMM_SIZE)(%rsp), %xmm1, %xmm2
        vpmovmskb %xmm2, %esi
        cmpl $0xffff, %esi
        je 1f
        vmovdqu LRV_VECTOR1_OFFSET(%rsp), %ymm1
-       jmp 1f
 
 L(no_avx4):
+1:
 # endif
-       movaps LRV_XMM0_OFFSET(%rsp), %xmm0
-       movaps LRV_XMM1_OFFSET(%rsp), %xmm1
 
-1:     fldt LRV_ST1_OFFSET(%rsp)
+       fldt LRV_ST1_OFFSET(%rsp)
        fldt LRV_ST0_OFFSET(%rsp)
 
        movq %rbx, %rsp