Faster memory allocs

3 files changed, 39 insertions, 1 deletions

diff --git a/TODO b/TODO
index ee7575d..6dadc64 100644
--- a/TODO
+++ b/TODO
@@ -1,3 +1,6 @@
+* optimize the free-block scanning loop
+* why is vp3view crashing?
+* apply the trace extractor to non-tail situations
 * utf8
 * regex substitution
 * asm-based regex engine
@@ -14,6 +17,7 @@
 * consider polymorphic call-site caches instead of deffs_t_ / deffs?
 
 * eliminate all FIXMEs
+** unite the /EQ-LT-OR-GE... things in lib/sys/opt.ey
 * stack underflow protection
 * input prompt
 * better interactive error handling when interpreting from stdin
diff --git a/compiler/elymasAsmLib.ey b/compiler/elymasAsmLib.ey
index 19513c7..8a0d139 100644
--- a/compiler/elymasAsmLib.ey
+++ b/compiler/elymasAsmLib.ey
@@ -525,6 +525,20 @@
       # rsi > 0  =>  currently counting block extent of free block
 
       @searchFreeBlockStart
+      /rbx /rax :movqRegReg
+      6 /rax :shrqImm8Reg
+      3 /rax :shlqImm8Reg
+      /rax /r8 /rdx :movqMemIndexReg
+      /rax /r9 /rax :movqMemIndexReg
+      /rdx :notqReg
+      /rdx /rax :andqRegReg
+      /rbx /rcx :movqRegReg
+      63 /rcx :andqImm8Reg
+      /rax :shrqClReg
+      /rax :shlqClReg
+      /rax /rax :bsfqRegReg # rax == lowest bit which denotes free block start
+      /immediateHit :jnzLbl8
+
       /rbx /r8 :btqRegMem # test block bitmap
       /nonFreeBlockStartFound :jcLbl8 # block not free
       /rbx /r9 :btqRegMem # test mark bitmap
@@ -539,6 +553,10 @@
       /searchFreeBlockStart :jaLbl8
       /noFreeBlockAvailable :jmpLbl32
 
+      @immediateHit
+      %C0 /rbx :andqImm8Reg
+      /rax /rbx :addqRegReg
+
       @freeBlockStartFound
       /rbx /rdx :movqRegReg
       @freeBlockContinues
diff --git a/compiler/elymasAsmOps.ey b/compiler/elymasAsmOps.ey
index 8a3c2a2..d397f73 100644
--- a/compiler/elymasAsmOps.ey
+++ b/compiler/elymasAsmOps.ey
@@ -65,7 +65,7 @@
 } /modrm00 deff
 
 { ==mem ==reg
-  mem [ /spl /sp /esp /rsp ] eq any {
+  mem [ /spl /sp /esp /rsp /r12b /r12w /r12d /r12 ] eq any {
     # actually encode sib with rsp index register (to get it ignored)
                     %40
     /sib regno      %07 band add
@@ -1222,6 +1222,14 @@ memoryAddressingVariants keys { ==variant memoryAddressingVariants variant . =*p
 } /nop deff
 
 { ==reg
+  reg bit8assert
+
+  reg regno %07 gt reg rexreqbyte or { 0 /none /none reg rex } rep
+  %F6
+  /two reg modrm11
+} /notbReg deff
+
+{ ==reg
   reg bit64assert
 
   1 /none /none reg rex
@@ -1370,6 +1378,14 @@ memoryAddressingVariants keys { ==variant memoryAddressingVariants variant . =*p
   reg bit64assert
 
   1 /none /none reg rex
+  %D1
+  /four reg modrm11
+} /shlq1Reg deff
+
+{ ==reg
+  reg bit64assert
+
+  1 /none /none reg rex
   %D3
   /four reg modrm11
 } /shlqClReg deff