cleanup boundary crossing test

travisdowns · Jan 22, 2018 · 373a782 · 373a782
1 parent dc4817a
commit 373a782
Show file tree

Hide file tree

Showing 2 changed files with 16 additions and 34 deletions.
diff --git a/misc-benches.cpp b/misc-benches.cpp
@@ -11,10 +11,8 @@ extern "C" {
 bench2_f misc_add_loop32;
 bench2_f misc_add_loop64;
 bench2_f misc_port7;
-bench2_f misc_dsb_align_16;
-bench2_f misc_dsb_align_32;
-bench2_f dsb_align_body_16;
-bench2_f dsb_align_body_32;
+bench2_f dsb_alignment_cross64;
+bench2_f dsb_alignment_nocross64;
 bench2_f bmi_tzcnt;
 bench2_f bmi_lzcnt;
 bench2_f bmi_popcnt;
@@ -36,7 +34,7 @@ void register_misc(GroupList& list) {
 
     using default_maker = BenchmarkMaker<TIMER>;
 
-    const uint32_t iters = 10*1000*1000;
+    const uint32_t iters = 10*1000;
     auto benches = std::vector<Benchmark> {
         default_maker::template make_bench<misc_add_loop32>(misc_group.get(), "add-32", "32-bit add-loop", 1,
                 []{ return nullptr; }, iters),
@@ -46,9 +44,9 @@ void register_misc(GroupList& list) {
                 []{ return nullptr; }, iters),
 
         // https://dendibakh.github.io/blog/2018/01/18/Code_alignment_issues
-        default_maker::template make_bench<dsb_align_body_16>(misc_group.get(), "dsb-align16", "Weird alignment effect 16", 1,
+        default_maker::template make_bench<dsb_alignment_cross64>(misc_group.get(), "dsb-align64-cross", "Crosses 64-byte i-boundary", 1,
                 []{ return aligned_ptr(1024, 1024); }, 1024),
-        default_maker::template make_bench<dsb_align_body_32>(misc_group.get(), "dsb-align32", "Weird alignment effect 32", 1,
+        default_maker::template make_bench<dsb_alignment_nocross64>(misc_group.get(), "dsb-align64-nocross", "No cross 64-byte i-boundary", 1,
                 []{ return aligned_ptr(1024, 1024); }, 1024),
     };
 

diff --git a/x86_methods.asm b/x86_methods.asm
@@ -512,45 +512,29 @@ nop7
 %assign i i+1
 %endrep
 
-%define ITERS 128
-
 %macro dsb_body 0
+times 2 nop8
 .outer:
-mov         rax, -(ITERS * 32)
-mov         rdx, rax
-nop
-nop
-nop
-nop
-nop
-nop
-nop
-nop
-nop6
+mov     rax, 128
 .top:
 nop9
-add         rdx, 32
-jne         .top
+dec     rax
+jne     .top
 dec     rdi
 jnz     .outer
 ret
 ud2
 %endmacro
 
-GLOBAL dsb_align_body_16,dsb_align_body_32
+GLOBAL dsb_alignment_cross64,dsb_alignment_nocross64
 
-ALIGN 32
-nop8
-nop8
-nop8
-nop8
-; the loop ends up 16-byte aligned, not 32 because the code before the loop is 16 bytes
-dsb_align_body_16:
+ALIGN 64
+times 4 nop8
+; the loop ends up crossing a 64-byte boundary
+dsb_alignment_cross64:
 dsb_body
 
-times 0 * 32 int3
-
-ALIGN 32
-dsb_align_body_32:
+ALIGN 64
+dsb_alignment_nocross64:
 dsb_body