[Git][ghc/ghc][master] Aarch ncg: Optimize immediate use for address calculations

[Marge Bot (@marge-bot)]

Marge Bot pushed to branch master at Glasgow Haskell Compiler / GHC


Commits:
8f3b3b78 by Andreas Klebinger at 2023-08-01T14:49:54-04:00
Aarch ncg: Optimize immediate use for address calculations

When the offset doesn't fit into the immediate we now just reuse the
general getRegister' code path which is well optimized to compute the
offset into a register instead of a special case for CmmRegOff.

This means we generate a lot less code under certain conditions which is
why performance metrics for these improve.

-------------------------
Metric Decrease:
    T4801
    T5321FD
    T5321Fun
-------------------------

- - - - -


1 changed file:

- compiler/GHC/CmmToAsm/AArch64/CodeGen.hs


Changes:

=====================================
compiler/GHC/CmmToAsm/AArch64/CodeGen.hs
=====================================
@@ -462,6 +462,23 @@ getBitmaskImm n w
   where
     truncated = narrowU w n
 
+-- | Load/store immediate.
+-- Depends on the width of the store to some extent.
+isOffsetImm :: Int -> Width -> Bool
+isOffsetImm off w
+  -- 8 bits + sign for unscaled offsets
+  | -256 <= off, off <= 255 = True
+  -- Offset using 12-bit positive immediate, scaled by width
+  -- LDR/STR: imm12: if reg is 32bit: 0 -- 16380 in multiples of 4
+  -- LDR/STR: imm12: if reg is 64bit: 0 -- 32760 in multiples of 8
+  -- 16-bit: 0 .. 8188, 8-bit: 0 -- 4095
+  | 0 <= off, off < 4096 * byte_width, off `mod` byte_width == 0 = True
+  | otherwise = False
+  where
+    byte_width = widthInBytes w
+
+
+
 
 -- TODO OPT: we might be able give getRegister
 --          a hint, what kind of register we want.
@@ -711,18 +728,11 @@ getRegister' config plat expr
       -> return (Fixed (cmmTypeFormat (cmmRegType reg))
                        (getRegisterReg plat reg)
                        nilOL)
-    CmmRegOff reg off | isNbitEncodeable 12 (fromIntegral off) -> do
-      getRegister' config plat $
-            CmmMachOp (MO_Add width) [CmmReg reg, CmmLit (CmmInt (fromIntegral off) width)]
-          where width = typeWidth (cmmRegType reg)
-
-    CmmRegOff reg off -> do
-      (off_r, _off_format, off_code) <- getSomeReg $ CmmLit (CmmInt (fromIntegral off) width)
-      (reg, _format, code) <- getSomeReg $ CmmReg reg
-      return $ Any (intFormat width) (\dst -> off_code `appOL` code `snocOL` ADD (OpReg width dst) (OpReg width reg) (OpReg width off_r))
-          where width = typeWidth (cmmRegType reg)
-
-
+    CmmRegOff reg off ->
+      -- If we got here we will load the address into a register either way. So we might as well just expand
+      -- and re-use the existing code path to handle "reg + off".
+      let !width = cmmRegWidth reg
+      in getRegister' config plat (CmmMachOp (MO_Add width) [CmmReg reg, CmmLit (CmmInt (fromIntegral off) width)])
 
     -- for MachOps, see GHC.Cmm.MachOp
     -- For CmmMachOp, see GHC.Cmm.Expr
@@ -1244,20 +1254,8 @@ getAmode :: Platform
 
 -- OPTIMIZATION WARNING: Addressing modes.
 -- Addressing options:
--- LDUR/STUR: imm9: -256 - 255
-getAmode platform _ (CmmRegOff reg off) | -256 <= off, off <= 255
-  = return $ Amode (AddrRegImm reg' off') nilOL
-    where reg' = getRegisterReg platform reg
-          off' = ImmInt off
--- LDR/STR: imm12: if reg is 32bit: 0 -- 16380 in multiples of 4
-getAmode platform W32 (CmmRegOff reg off)
-  | 0 <= off, off <= 16380, off `mod` 4 == 0
-  = return $ Amode (AddrRegImm reg' off') nilOL
-    where reg' = getRegisterReg platform reg
-          off' = ImmInt off
--- LDR/STR: imm12: if reg is 64bit: 0 -- 32760 in multiples of 8
-getAmode platform W64 (CmmRegOff reg off)
-  | 0 <= off, off <= 32760, off `mod` 8 == 0
+getAmode platform w (CmmRegOff reg off)
+  | isOffsetImm off w
   = return $ Amode (AddrRegImm reg' off') nilOL
     where reg' = getRegisterReg platform reg
           off' = ImmInt off
@@ -1266,15 +1264,15 @@ getAmode platform W64 (CmmRegOff reg off)
 -- CmmStore (CmmMachOp (MO_Add w) [CmmLoad expr, CmmLit (CmmInt n w')]) (expr2)
 -- E.g. a CmmStoreOff really. This can be translated to `str $expr2, [$expr, #n ]
 -- for `n` in range.
-getAmode _platform _ (CmmMachOp (MO_Add _w) [expr, CmmLit (CmmInt off _w')])
-  | -256 <= off, off <= 255
+getAmode _platform w (CmmMachOp (MO_Add _w) [expr, CmmLit (CmmInt off _w')])
+  | isOffsetImm (fromIntegral off) w
   = do (reg, _format, code) <- getSomeReg expr
        return $ Amode (AddrRegImm reg (ImmInteger off)) code
 
-getAmode _platform _ (CmmMachOp (MO_Sub _w) [expr, CmmLit (CmmInt off _w')])
-  | -256 <= -off, -off <= 255
+getAmode _platform w (CmmMachOp (MO_Sub _w) [expr, CmmLit (CmmInt off _w')])
+  | isOffsetImm (fromIntegral $ -off) w
   = do (reg, _format, code) <- getSomeReg expr
-       return $ Amode (AddrRegImm reg (ImmInteger (-off))) code
+       return $ Amode (AddrRegImm reg (ImmInteger $ -off)) code
 
 -- Generic case
 getAmode _platform _ expr



View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/8f3b3b78a8cce3bd463ed175ee933c2aabffc631

-- 
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/8f3b3b78a8cce3bd463ed175ee933c2aabffc631
You're receiving this email because of your account on gitlab.haskell.org.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-commits/attachments/20230801/e3f31d65/attachment-0001.html>