diff options
| author | Torbjorn Granlund <tege@gmplib.org> | 2012-03-13 20:24:40 +0100 |
|---|---|---|
| committer | Torbjorn Granlund <tege@gmplib.org> | 2012-03-13 20:24:40 +0100 |
| commit | d8e5cad31f22d2b09adc4cda55d11a25d4966543 (patch) | |
| tree | 1dcd93bffb890977f8235ba2055d08f937894a02 /mpn/x86_64/core2 | |
| parent | 1761a0cf9244597affef02697f5fcfe96634191d (diff) | |
| download | gmp-d8e5cad31f22d2b09adc4cda55d11a25d4966543.tar.gz | |
Shorten critical path.
Diffstat (limited to 'mpn/x86_64/core2')
| -rw-r--r-- | mpn/x86_64/core2/gcd_1.asm | 41 |
1 files changed, 21 insertions, 20 deletions
diff --git a/mpn/x86_64/core2/gcd_1.asm b/mpn/x86_64/core2/gcd_1.asm index 765e74eaa..a4b97978c 100644 --- a/mpn/x86_64/core2/gcd_1.asm +++ b/mpn/x86_64/core2/gcd_1.asm @@ -28,16 +28,18 @@ C cycles/bit (approx) C AMD K8,K9 8.5 C AMD K10 5 C AMD bd1 5 -C AMD bobcat 11 -C Intel P4 24 -C Intel core2 5.5 -C Intel NHM 6 -C Intel SBR 6 +C AMD bobcat 10 +C Intel P4 18 +C Intel core2 4.3 +C Intel NHM 5 +C Intel SBR 5 C Intel atom 17 -C VIA nano 6.5 - +C VIA nano 5.3 C Numbers measured with: speed -CD -s1-64 mpn_gcd_1 +C TODO +C * Optimise inner-loop for specific CPUs. The code relies too much on OoO +C execution. C INPUT PARAMETERS define(`up', `%rdi') @@ -93,27 +95,26 @@ L(reduced): pop %rdx pop %r8 - test %rax, %rax + bsf %rax, %rcx - mov %rax, %rcx + test %rax, %rax jnz L(mid) - mov %rdx, %rax jmp L(done) - ALIGN(16) C K10 C2 NHM SBR -L(top): cmovc %r10, %rax C if x-y carried 0,7 0,6 0,7 0 - cmovc %rcx, %rdx C use x,y-x 0 1 1 1 -L(mid): bsf %rax, %rcx C 1 2 2 2 - mov %rdx, %r10 C 1 3 3 3 - shr R8(%rcx), %rax C 5 4 5 5 - mov %rax, %rcx C 6 5 6 7 - sub %rax, %r10 C 6 5 6 7 - sub %rdx, %rax C 6 5 6 7 + ALIGN(16) C K10 BD C2 NHM SBR +L(top): cmovc %r10, %rax C if x-y < 0 0,6 0,5 0,6 0,5 0,6 + cmovc %r9, %rdx C use x,y-x 0 0 2 1 1 + bsf %r10, %rcx C 0 0 0 0 0 +L(mid): mov %rdx, %r10 C 1 1 4 3 3 + shr R8(%rcx), %rax C 4 3 2 3 3 + mov %rax, %r9 C 5 4 3 4 5 + sub %rax, %r10 C 5 4 5 4 5 + sub %rdx, %rax C 5 4 4 4 5 jnz L(top) C - mov %rcx, %rax L(done): + mov %rdx, %rax mov %r8, %rcx shl R8(%rcx), %rax DOS64_EXIT() |