[Git][ghc/ghc][wip/T23645] Improve mulmayoflo

[Jaro Reinders (@Noughtmare)]

Jaro Reinders pushed to branch wip/T23645 at Glasgow Haskell Compiler / GHC


Commits:
9513973b by Jaro Reinders at 2023-07-19T12:00:48+02:00
Improve mulmayoflo

- - - - -


3 changed files:

- compiler/GHC/CmmToAsm/X86/CodeGen.hs
- testsuite/tests/codeGen/should_run/MulMayOflo.hs
- testsuite/tests/codeGen/should_run/all.T


Changes:

=====================================
compiler/GHC/CmmToAsm/X86/CodeGen.hs
=====================================
@@ -972,21 +972,20 @@ getRegister' _ is32Bit (CmmMachOp mop [x, y]) = -- dyadic MachOps
          -- multiplication doesn't use two registers.
          --
          -- The plan is:
-         -- 1. truncate and sign-extend a and b to 8bit width
-         -- 2. multiply a' = a * b in 32bit width
-         -- 3. copy and sign-extend 8bit from a' to c
-         -- 4. compare a' and c: they are equal if there was no overflow
+         -- 1. multiply b1 = a * b where b1 has 16bit width
+         -- 2. subtract b2 = b1 - (-128)
+         -- (this has shorter machine code than b1 + 128)
+         -- 3. compare b2 > 255 unsigned
+         -- This is equivalent to checking b1 < -128 || 127 < b1
          (a_reg, a_code) <- getNonClobberedReg a
-         (b_reg, b_code) <- getNonClobberedReg b
+         b_code <- getAnyReg b
          let
-             code = a_code `appOL` b_code `appOL`
+             code = a_code `appOL` b_code eax `appOL`
                         toOL [
-                           MOVSxL II8 (OpReg a_reg) (OpReg a_reg),
-                           MOVSxL II8 (OpReg b_reg) (OpReg b_reg),
-                           IMUL II32 (OpReg b_reg) (OpReg a_reg),
-                           MOVSxL II8 (OpReg a_reg) (OpReg eax),
-                           CMP II16 (OpReg a_reg) (OpReg eax),
-                           SETCC NE (OpReg eax)
+                           IMUL2 II8 (OpReg a_reg),   -- result in %ax
+                           SUB II16 (OpImm (ImmInt (-128))) (OpReg eax),
+                           CMP II16 (OpImm (ImmInt 255)) (OpReg eax),
+                           SETCC GU (OpReg eax)
                         ]
          return (Fixed II8 eax code)
     imulMayOflo rep a b = do


=====================================
testsuite/tests/codeGen/should_run/MulMayOflo.hs
=====================================
@@ -23,4 +23,4 @@ import GHC.Exts
 foreign import prim "runCmmzh" runCmmzh# :: Word# -> Word#
 
 main :: IO ()
-main = print . show $ W# (runCmmzh# 42##)
+main = W# (runCmmzh# 42##) `seq` pure ()


=====================================
testsuite/tests/codeGen/should_run/all.T
=====================================
@@ -232,8 +232,6 @@ test('OrigThunkInfo', normal, compile_and_run, ['-forig-thunk-info'])
 # Note [MO_S_MulMayOflo significant width]) and may require fixing/adjustment.
 test('MulMayOflo_full',
      [ extra_files(['MulMayOflo.hs']),
-       when(unregisterised(), skip),
-       unless(arch('x86_64') or arch('i386'), skip),
-        ignore_stdout],
+       when(unregisterised(), skip) ],
      multi_compile_and_run,
      ['MulMayOflo', [('MulMayOflo_full.cmm', '')], ''])



View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/9513973b37c0cce03b7a57f28be124cdd54efb69

-- 
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/9513973b37c0cce03b7a57f28be124cdd54efb69
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/20230719/8077f22a/attachment-0001.html>