Add default impl of commuteInstruction
Add notes about ugly V9 code.

llvm-svn: 19684

foo:
        mov %EAX, DWORD PTR [%ESP + 4]
        mov %EDX, DWORD PTR [%ESP + 8]
        shrd %EAX, %EDX, 2
        sar %EDX, 2
        ret

instead of this:

test1:
        mov %ECX, DWORD PTR [%ESP + 4]
        shr %ECX, 2
        mov %EDX, DWORD PTR [%ESP + 8]
        mov %EAX, %EDX
        shl %EAX, 30
        or %EAX, %ECX
        sar %EDX, 2
        ret

and long << 2 to this:

foo:
        mov %EAX, DWORD PTR [%ESP + 4]
        mov %ECX, DWORD PTR [%ESP + 8]
***     mov %EDX, %EAX
        shrd %EDX, %ECX, 30
        shl %EAX, 2
        ret

instead of this:

foo:
        mov %EAX, DWORD PTR [%ESP + 4]
        mov %ECX, %EAX
        shr %ECX, 30
        mov %EDX, DWORD PTR [%ESP + 8]
        shl %EDX, 2
        or %EDX, %ECX
        shl %EAX, 2
        ret

The extra copy (marked ***) can be eliminated when I teach the code generator
that shrd32rri8 is really commutative.

llvm-svn: 19681

select operations or to shifts that are by a constant.  This automatically
implements (with no special code) all of the special cases for shift by 32,
shift by < 32 and shift by > 32.

llvm-svn: 19679

llvm-svn: 19678

range.  Either they are undefined (the default), they mask the shift amount
to the size of the register (X86, Alpha, etc), or they extend the shift (PPC).

This defaults to undefined, which is conservatively correct.

llvm-svn: 19677

llvm-svn: 19675

FP_EXTEND from!

llvm-svn: 19674

llvm-svn: 19673

don't need to even think about F32 in the X86 code anymore.

llvm-svn: 19672

of zero and sign extends.

llvm-svn: 19671

llvm-svn: 19670

do it.  This results in better code on X86 for floats (because if strict
precision is not required, we can elide some more expensive double -> float
conversions like the old isel did), and allows other targets to emit
CopyFromRegs that are not legal for arguments.

llvm-svn: 19668

llvm-svn: 19667

llvm-svn: 19661

llvm-svn: 19660

llvm-svn: 19659

* Insert some really pedantic assertions that will notice when we emit the
  same loads more than one time, exposing bugs.  This turns a miscompilation in
  bzip2 into a compile-fail.  yaay.

llvm-svn: 19658

llvm-svn: 19657

llvm-svn: 19656

match (X+Y)+(Z << 1), because we match the X+Y first, consuming the index
register, then there is no place to put the Z.

llvm-svn: 19652

llvm-svn: 19651

emitted too early.  In particular, this fixes
Regression/CodeGen/X86/regpressure.ll:regpressure3.

This also improves the 2nd basic block in 164.gzip:flush_block, which went from

.LBBflush_block_1:      # loopentry.1.i
        movzx %EAX, WORD PTR [dyn_ltree + 20]
        movzx %ECX, WORD PTR [dyn_ltree + 16]
        mov DWORD PTR [%ESP + 32], %ECX
        movzx %ECX, WORD PTR [dyn_ltree + 12]
        movzx %EDX, WORD PTR [dyn_ltree + 8]
        movzx %EBX, WORD PTR [dyn_ltree + 4]
        mov DWORD PTR [%ESP + 36], %EBX
        movzx %EBX, WORD PTR [dyn_ltree]
        add DWORD PTR [%ESP + 36], %EBX
        add %EDX, DWORD PTR [%ESP + 36]
        add %ECX, %EDX
        add DWORD PTR [%ESP + 32], %ECX
        add %EAX, DWORD PTR [%ESP + 32]
        movzx %ECX, WORD PTR [dyn_ltree + 24]
        add %EAX, %ECX
        mov %ECX, 0
        mov %EDX, %ECX

to

.LBBflush_block_1:      # loopentry.1.i
        movzx %EAX, WORD PTR [dyn_ltree]
        movzx %ECX, WORD PTR [dyn_ltree + 4]
        add %ECX, %EAX
        movzx %EAX, WORD PTR [dyn_ltree + 8]
        add %EAX, %ECX
        movzx %ECX, WORD PTR [dyn_ltree + 12]
        add %ECX, %EAX
        movzx %EAX, WORD PTR [dyn_ltree + 16]
        add %EAX, %ECX
        movzx %ECX, WORD PTR [dyn_ltree + 20]
        add %ECX, %EAX
        movzx %EAX, WORD PTR [dyn_ltree + 24]
        add %ECX, %EAX
        mov %EAX, 0
        mov %EDX, %EAX

... which results in less spilling in the function.

This change alone speeds up 164.gzip from 37.23s to 36.24s on apoc.  The
default isel takes 37.31s.

llvm-svn: 19650

llvm-svn: 19649

llvm-svn: 19647

llvm-svn: 19645

X86/reg-pressure.ll again, and allows us to do nice things in other cases.
For example, we now codegen this sort of thing:

int %loadload(int *%X, int* %Y) {
  %Z = load int* %Y
  %Y = load int* %X      ;; load between %Z and store
  %Q = add int %Z, 1
  store int %Q, int* %Y
  ret int %Y
}

Into this:

loadload:
        mov %EAX, DWORD PTR [%ESP + 4]
        mov %EAX, DWORD PTR [%EAX]
        mov %ECX, DWORD PTR [%ESP + 8]
        inc DWORD PTR [%ECX]
        ret

where we weren't able to form the 'inc [mem]' before.  This also lets the
instruction selector emit loads in any order it wants to, which can be good
for register pressure as well.

llvm-svn: 19644

 1. Fold  [mem] += (1|-1) into inc [mem]/dec [mem] to save some icache space.
 2. Do not let token factor nodes prevent forming '[mem] op= val' folds.

llvm-svn: 19643

llvm-svn: 19642

llvm-svn: 19641

the basic block that uses them if possible.  This is a big win on X86, as it
lets us fold the argument loads into instructions and reduce register pressure
(by not loading all of the arguments in the entry block).

For this (contrived to show the optimization) testcase:

int %argtest(int %A, int %B) {
        %X = sub int 12345, %A
        br label %L
L:
        %Y = add int %X, %B
        ret int %Y
}

we used to produce:

argtest:
        mov %ECX, DWORD PTR [%ESP + 4]
        mov %EAX, 12345
        sub %EAX, %ECX
        mov %EDX, DWORD PTR [%ESP + 8]
.LBBargtest_1:  # L
        add %EAX, %EDX
        ret


now we produce:

argtest:
        mov %EAX, 12345
        sub %EAX, DWORD PTR [%ESP + 4]
.LBBargtest_1:  # L
        add %EAX, DWORD PTR [%ESP + 8]
        ret

This also fixes the FIXME in the code.

BTW, this occurs in real code.  164.gzip shrinks from 8623 to 8608 lines of
.s file.  The stack frame in huft_build shrinks from 1644->1628 bytes,
inflate_codes shrinks from 116->108 bytes, and inflate_block from 2620->2612,
due to fewer spills.

Take that alkis. :-)

llvm-svn: 19639

operations.

The body of the if is less indented but unmodified in this patch.

llvm-svn: 19638

llvm-svn: 19635

int %foo(int %X) {
        %T = add int %X, 13
        %S = mul int %T, 3
        ret int %S
}

as this:

        mov %ECX, DWORD PTR [%ESP + 4]
        lea %EAX, DWORD PTR [%ECX + 2*%ECX + 39]
        ret

instead of this:

        mov %ECX, DWORD PTR [%ESP + 4]
        mov %EAX, %ECX
        add %EAX, 13
        imul %EAX, %EAX, 3
        ret

llvm-svn: 19633

llvm-svn: 19632

Do not fold a load into an operation if it will induce a cycle in the DAG.

Repeat after me: dAg.

llvm-svn: 19631

useness.

llvm-svn: 19629

Disable the xform for < > cases.  It turns out that the following is being
miscompiled:

bool %test(sbyte %S) {
        %T = cast sbyte %S to uint
        %V = setgt uint %T, 255
        ret bool %V
}

llvm-svn: 19628

The comparison will probably be folded, so this is not ok to do.
This fixed 197.parser.

llvm-svn: 19624

of the bytereg.  This fixes yacr2, 300.twolf and probably others.

llvm-svn: 19622

If we emit a load because we followed a token chain to get to it, try to
fold it into its single user if possible.

llvm-svn: 19620