testcase like this:
int %test(int* %P, int %A) {
%Pv = load int* %P
%B = add int %A, %Pv
ret int %B
}
We now generate:
test:
mov %ECX, DWORD PTR [%ESP + 4]
mov %EAX, DWORD PTR [%ESP + 8]
add %EAX, DWORD PTR [%ECX]
ret
Instead of:
test:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, DWORD PTR [%ESP + 8]
mov %EAX, DWORD PTR [%EAX]
add %EAX, %ECX
ret
... saving one instruction, and often a register. Note that there are a lot
of other instructions that could use this, but they aren't handled. I'm not
really interested in adding them, but mul/div and all of the FP instructions
could be supported as well if someone wanted to add them.
llvm-svn: 12204
This allows pointers to aggregate objects, whose elements are only read, to
be promoted and passed in by element instead of by reference. This can
enable a LOT of subsequent optimizations in the caller function.
It's worth pointing out that this stuff happens a LOT of C++ programs, because
objects in templates are generally passed around by reference. When these
templates are instantiated on small aggregate or scalar types, however, it is
more efficient to pass them in by value than by reference.
This transformation triggers most on C++ codes (e.g. 334 times on eon), but
does happen on C codes as well. For example, on mesa it triggers 72 times,
and on gcc it triggers 35 times. this is amazingly good considering that
we are using 'basicaa' so far.
llvm-svn: 12202
The problem was that we were merging a field of a node with a value that was
deleted. Thanks to bugpoint for reducing povray to a nice small 3 function
example. :)
llvm-svn: 12116
Make sure to scope the NodeMap passed into cloneInto so that it doesn't point
to nodes that are deleted. Add some FIXME's for future performance enhancements.
llvm-svn: 12115
do it on povray. The problem is that we were not copying globals from callees to
callers unless the existed in both graphs. We should have copied them in the case
where the global pointed to a node that was copied as well.
llvm-svn: 12104
been using the default target data layout object to lower malloc instructions,
causing us to allocate more memory than we needed! This could improve the
performance of the CBE generated code substantially!
llvm-svn: 12088
(16) into certain areas of the SPARC V9 back-end. I'm fairly sure the US IIIi's
dcache has 32-byte lines, so I'm not sure where the 16 came from. However, in
the interest of not breaking things any more than they already are, I'm going
to leave the constant alone.
llvm-svn: 12043
of generating this code:
mov %EAX, 4
mov DWORD PTR [%ESP], %EAX
mov %AX, 123
movsx %EAX, %AX
mov DWORD PTR [%ESP + 4], %EAX
call Y
we now generate:
mov DWORD PTR [%ESP], 4
mov DWORD PTR [%ESP + 4], 123
call Y
Which hurts the eyes less. :)
Considering that register pressure around call sites is already high (with all
of the callee clobber registers n stuff), this may help a lot.
llvm-svn: 12028
1) For 8-bit registers try to use first the ones that are parts of the
same register (AL then AH). This way we only alias 2 16/32-bit
registers after allocating 4 8-bit variables.
2) Move EBX as the last register to allocate. This will cause less
spills to happen since we will have 8-bit registers available up to
register excaustion (assuming we use the allocation order). It
would be nice if we could push all of the 8-bit aliased registers
towards the end but we much prefer to keep callee saved register to
the end to avoid saving them on entry and exit of the function.
For example this gives a slight reduction of spills with linear scan
on 164.gzip.
Before:
11221 asm-printer - Number of machine instrs printed
975 spiller - Number of loads added
675 spiller - Number of stores added
398 spiller - Number of register spills
After:
11182 asm-printer - Number of machine instrs printed
952 spiller - Number of loads added
652 spiller - Number of stores added
386 spiller - Number of register spills
llvm-svn: 11996
their names more decriptive. A name consists of the base name, a
default operand size followed by a character per operand with an
optional special size. For example:
ADD8rr -> add, 8-bit register, 8-bit register
IMUL16rmi -> imul, 16-bit register, 16-bit memory, 16-bit immediate
IMUL16rmi8 -> imul, 16-bit register, 16-bit memory, 8-bit immediate
MOVSX32rm16 -> movsx, 32-bit register, 16-bit memory
llvm-svn: 11995
Note that this is a band-aid put over a band-aid. This just undisables
tail duplication in on very specific case that it seems to work in.
llvm-svn: 11989
parse. The name is now I (operand size)*. For example:
Im32 -> instruction with 32-bit memory operands.
Im16i8 -> instruction with 16-bit memory operands and 8 bit immediate
operands.
llvm-svn: 11970
the size of the immediate and the memory operand on instructions that
use them. This resolves problems with instructions that take both a
memory and an immediate operand but their sizes differ (i.e. ADDmi32b).
llvm-svn: 11967
an 8-bit immediate. So mark the shifts that take immediates as taking
an 8-bit argument. The rest with the implicit use of CL are marked
appropriately.
A bug still exists:
def SHLDmri32 : I2A8 <"shld", 0xA4, MRMDestMem>, TB; // [mem32] <<= [mem32],R32 imm8
The immediate in the above instruction is 8-bit but the memory
reference is 32-bit. The printer prints this as an 8-bit reference
which confuses the assembler. Same with SHRDmri32.
llvm-svn: 11931
Functions with linkonce linkage are declared with weak linkage.
Global floating point constants used to represent unprintable values
(such as NaN and infinity) are declared static so that they don't interfere
with other CBE generated translation units.
llvm-svn: 11884
1. Functions do not make things incomplete, only variables
2. Constant global variables no longer need to be marked incomplete, because
we are guaranteed that the initializer for the global will be in the
graph we are hacking on now. This makes resolution of indirect calls happen
a lot more in the bu pass, supports things like vtables and the C counterparts
(giant constant arrays of function pointers), etc...
Testcase here: test/Regression/Analysis/DSGraph/constant_globals.ll
llvm-svn: 11852
Make the incompleteness marker faster by looping directly over the globals
instead of over the scalars to find the globals
Fix a bug where we didn't mark a global incomplete if it didn't have any
outgoing edges. This wouldn't break any current clients but is still wrong.
llvm-svn: 11848
pair, and look up varargs in the execution stack every time, instead of
just pushing iterators (which can be invalidated during callFunction())
around. (union GenericValue now has a "pair of uints" member, to support
this mechanism.) Fixes Bug 234.
llvm-svn: 11845
assume that if they don't intend to write to a global variable, that they
would mark it as constant. However, there are people that don't understand
that the compiler can do nice things for them if they give it the information
it needs.
This pass looks for blatently obvious globals that are only ever read from.
Though it uses a trivially simple "alias analysis" of sorts, it is still able
to do amazing things to important benchmarks. 253.perlbmk, for example,
contains several ***GIANT*** function pointer tables that are not marked
constant and should be. Marking them constant allows the optimizer to turn
a whole bunch of indirect calls into direct calls. Note that only a link-time
optimizer can do this transformation, but perlbmk does have several strings
and other minor globals that can be marked constant by this pass when run
from GCCAS.
176.gcc has a ton of strings and large tables that are marked constant, both
at compile time (38 of them) and at link time (48 more). Other benchmarks
give similar results, though it seems like big ones have disproportionally
more than small ones.
This pass is extremely quick and does good things. I'm going to enable it
in gccas & gccld. Not bad for 50 SLOC.
llvm-svn: 11836
scaled indexes. This allows us to compile GEP's like this:
int* %test([10 x { int, { int } }]* %X, int %Idx) {
%Idx = cast int %Idx to long
%X = getelementptr [10 x { int, { int } }]* %X, long 0, long %Idx, ubyte 1, ubyte 0
ret int* %X
}
Into a single address computation:
test:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, DWORD PTR [%ESP + 8]
lea %EAX, DWORD PTR [%EAX + 8*%ECX + 4]
ret
Before it generated:
test:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, DWORD PTR [%ESP + 8]
shl %ECX, 3
add %EAX, %ECX
lea %EAX, DWORD PTR [%EAX + 4]
ret
This is useful for things like int/float/double arrays, as the indexing can be folded into
the loads&stores, reducing register pressure and decreasing the pressure on the decode unit.
With these changes, I expect our performance on 256.bzip2 and gzip to improve a lot. On
bzip2 for example, we go from this:
10665 asm-printer - Number of machine instrs printed
40 ra-local - Number of loads/stores folded into instructions
1708 ra-local - Number of loads added
1532 ra-local - Number of stores added
1354 twoaddressinstruction - Number of instructions added
1354 twoaddressinstruction - Number of two-address instructions
2794 x86-peephole - Number of peephole optimization performed
to this:
9873 asm-printer - Number of machine instrs printed
41 ra-local - Number of loads/stores folded into instructions
1710 ra-local - Number of loads added
1521 ra-local - Number of stores added
789 twoaddressinstruction - Number of instructions added
789 twoaddressinstruction - Number of two-address instructions
2142 x86-peephole - Number of peephole optimization performed
... and these types of instructions are often in tight loops.
Linear scan is also helped, but not as much. It goes from:
8787 asm-printer - Number of machine instrs printed
2389 liveintervals - Number of identity moves eliminated after coalescing
2288 liveintervals - Number of interval joins performed
3522 liveintervals - Number of intervals after coalescing
5810 liveintervals - Number of original intervals
700 spiller - Number of loads added
487 spiller - Number of stores added
303 spiller - Number of register spills
1354 twoaddressinstruction - Number of instructions added
1354 twoaddressinstruction - Number of two-address instructions
363 x86-peephole - Number of peephole optimization performed
to:
7982 asm-printer - Number of machine instrs printed
1759 liveintervals - Number of identity moves eliminated after coalescing
1658 liveintervals - Number of interval joins performed
3282 liveintervals - Number of intervals after coalescing
4940 liveintervals - Number of original intervals
635 spiller - Number of loads added
452 spiller - Number of stores added
288 spiller - Number of register spills
789 twoaddressinstruction - Number of instructions added
789 twoaddressinstruction - Number of two-address instructions
258 x86-peephole - Number of peephole optimization performed
Though I'm not complaining about the drop in the number of intervals. :)
llvm-svn: 11820
to do analysis.
*** FOLD getelementptr instructions into loads and stores when possible,
making use of some of the crazy X86 addressing modes.
For example, the following C++ program fragment:
struct complex {
double re, im;
complex(double r, double i) : re(r), im(i) {}
};
inline complex operator+(const complex& a, const complex& b) {
return complex(a.re+b.re, a.im+b.im);
}
complex addone(const complex& arg) {
return arg + complex(1,0);
}
Used to be compiled to:
_Z6addoneRK7complex:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, DWORD PTR [%ESP + 8]
*** mov %EDX, %ECX
fld QWORD PTR [%EDX]
fld1
faddp %ST(1)
*** add %ECX, 8
fld QWORD PTR [%ECX]
fldz
faddp %ST(1)
*** mov %ECX, %EAX
fxch %ST(1)
fstp QWORD PTR [%ECX]
*** add %EAX, 8
fstp QWORD PTR [%EAX]
ret
Now it is compiled to:
_Z6addoneRK7complex:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, DWORD PTR [%ESP + 8]
fld QWORD PTR [%ECX]
fld1
faddp %ST(1)
fld QWORD PTR [%ECX + 8]
fldz
faddp %ST(1)
fxch %ST(1)
fstp QWORD PTR [%EAX]
fstp QWORD PTR [%EAX + 8]
ret
Other programs should see similar improvements, across the board. Note that
in addition to reducing instruction count, this also reduces register pressure
a lot, always a good thing on X86. :)
llvm-svn: 11819
into a single LEA instruction. This should improve the code generated for
things like X->A.B.C[12].D.
The bigger benefit is still coming though. Note that this uses an LEA instruction
instead of an add, giving the register allocator more freedom. We should probably
never generate ADDri32's.
llvm-svn: 11817
Also fix problem where we didn't check to see if a node pointer was null.
Though fclose(null) doesn't make a lot of sense, 300.twolf does it.
llvm-svn: 11810
longer was getting this #include, it always fell back on the less precise
floating point initializer values, causing some testsuite failures.
llvm-svn: 11803
allocator.
The implementation is completely rewritten and now employs several
optimizations not exercised before. For example for 164.gzip we have
997 loads and 699 stores vs the 1221 loads and 880 stores we have
before.
llvm-svn: 11798
This case occurs many times in various benchmarks, especially when combined
with the previous patch. This allows it to get stuff like:
if (X == 4 || X == 3)
if (X == 5 || X == 8)
and
switch (X) {
case 4: case 5: case 6:
if (X == 4 || X == 5)
llvm-svn: 11797
