[commit: ghc] master: Code size micro-optimizations in the X86 backend (bdb0c43)
git at git.haskell.org
git at git.haskell.org
Tue Oct 7 19:25:22 UTC 2014
Repository : ssh://git@git.haskell.org/ghc
On branch : master
Link : http://ghc.haskell.org/trac/ghc/changeset/bdb0c43c7991da3856e3a89db57c9ea78d61f55f/ghc
>---------------------------------------------------------------
commit bdb0c43c7991da3856e3a89db57c9ea78d61f55f
Author: Reid Barton <rwbarton at gmail.com>
Date: Tue Oct 7 09:24:15 2014 -0400
Code size micro-optimizations in the X86 backend
Summary:
Carter Schonwald suggested looking for opportunities to replace
instructions in GHC's output by equivalent ones that are shorter,
as recommended by the Intel optimization manuals.
This patch reduces the module sizes as reported by nofib
by about 1.5% on x86_64.
Test Plan:
Built an i386 cross-compiler and ran the test suite; the same
(rather large) set of tests failed before and after this commit.
Will let Harbormaster validate on x86_64.
Reviewers: austin
Subscribers: thomie, carter, ezyang, simonmar
Differential Revision: https://phabricator.haskell.org/D320
>---------------------------------------------------------------
bdb0c43c7991da3856e3a89db57c9ea78d61f55f
compiler/nativeGen/X86/Ppr.hs | 35 ++++++++++++++++++++++++++++++++++-
1 file changed, 34 insertions(+), 1 deletion(-)
diff --git a/compiler/nativeGen/X86/Ppr.hs b/compiler/nativeGen/X86/Ppr.hs
index 7f8f296..fcefce3 100644
--- a/compiler/nativeGen/X86/Ppr.hs
+++ b/compiler/nativeGen/X86/Ppr.hs
@@ -522,6 +522,13 @@ pprInstr (RELOAD slot reg)
pprUserReg reg]
-}
+-- Replace 'mov $0x0,%reg' by 'xor %reg,%reg', which is smaller and cheaper.
+-- The code generator catches most of these already, but not all.
+pprInstr (MOV size (OpImm (ImmInt 0)) dst@(OpReg _))
+ = pprInstr (XOR size' dst dst)
+ where size' = case size of
+ II64 -> II32 -- 32-bit version is equivalent, and smaller
+ _ -> size
pprInstr (MOV size src dst)
= pprSizeOpOp (sLit "mov") size src dst
@@ -582,6 +589,14 @@ pprInstr (SUB_CC size src dst)
however, cannot be used to determine if the upper half of the
result is non-zero." So there.
-}
+
+-- Use a 32-bit instruction when possible as it saves a byte.
+-- Notably, extracting the tag bits of a pointer has this form.
+-- TODO: we could save a byte in a subsequent CMP instruction too,
+-- but need something like a peephole pass for this
+pprInstr (AND II64 src@(OpImm (ImmInteger mask)) dst)
+ | 0 <= mask && mask < 0xffffffff
+ = pprInstr (AND II32 src dst)
pprInstr (AND size src dst) = pprSizeOpOp (sLit "and") size src dst
pprInstr (OR size src dst) = pprSizeOpOp (sLit "or") size src dst
@@ -618,7 +633,25 @@ pprInstr (CMP size src dst)
is_float FF80 = True
is_float _ = False
-pprInstr (TEST size src dst) = pprSizeOpOp (sLit "test") size src dst
+pprInstr (TEST size src dst) = sdocWithPlatform $ \platform ->
+ let size' = case (src,dst) of
+ -- Match instructions like 'test $0x3,%esi' or 'test $0x7,%rbx'.
+ -- We can replace them by equivalent, but smaller instructions
+ -- by reducing the size of the immediate operand as far as possible.
+ -- (We could handle masks larger than a single byte too,
+ -- but it would complicate the code considerably
+ -- and tag checks are by far the most common case.)
+ (OpImm (ImmInteger mask), OpReg dstReg)
+ | 0 <= mask && mask < 256 -> minSizeOfReg platform dstReg
+ _ -> size
+ in pprSizeOpOp (sLit "test") size' src dst
+ where
+ minSizeOfReg platform (RegReal (RealRegSingle i))
+ | target32Bit platform && i <= 3 = II8 -- al, bl, cl, dl
+ | target32Bit platform && i <= 7 = II16 -- si, di, bp, sp
+ | not (target32Bit platform) && i <= 15 = II8 -- al .. r15b
+ minSizeOfReg _ _ = size -- other
+
pprInstr (PUSH size op) = pprSizeOp (sLit "push") size op
pprInstr (POP size op) = pprSizeOp (sLit "pop") size op
More information about the ghc-commits
mailing list