FP_REG_KILL instructions at the end of blocks involved with critical edges.
Fix a bug where FP_REG_KILL instructions weren't inserted in fall through
unconditional branches. Perhaps this will fix some linscan problems?
llvm-svn: 11019
instruction selector by adding a new pseudo-instruction
FP_REG_KILL. This instruction implicitly defines all x86 fp registers
and is a terminator so that passes which add machine code at the end
of basic blocks (like phi elimination) do not add instructions between
it and the branch or return instruction.
llvm-svn: 10562
getelementptr code path for use by other code paths (like malloc and alloca).
* Optimize comparisons with zero
* Generate neg, not, inc, and dec instructions, when possible.
This gives some code size wins, which might translate into performance. We'll
see tommorow in the nightly tester.
llvm-svn: 9267
until we implement unwinding.
Add support for the invoke instruction, which codegens just like a call with
a branch after it.
The end effect of this change is that programs using the invoke instruction,
but never unwinding, will work fine. Programs that unwind will abort until
we get unwind support.
llvm-svn: 8187
Avoid a fall-through in the (stubby) treatment of the longjmp intrinsic
call which causes llc & lli to core-dump.
Add a sort-of treatment of cast double to ulong. I am not really sure
what a user should expect to see upon casting a negative FP value to
unsigned long long. But with what is given here, I was able to write
a program that could cast -123.456 to ulong and back and get -123.0,
which seems like a step in the right direction. GCC seems to give you
0. I don't know if I'd consider that useful.
These cases were coming up in GNU coreutils-5.0.
llvm-svn: 7205
This avoid generating a register to hold C, which in turn speeds up the
register allocator by a lot: ~9% on 164.gzip and ~17% on 256.bzip2. This
also speeds up other passes. This also speeds up execution of the program
marginally, and makes the asm much easier to read. :)
llvm-svn: 6626
This improves the performance of the power benchmark by a few percent.
This will be neccesary for SSE code, which requires 16 byte alignment of
the stack.
llvm-svn: 5320
* Support arbitrary FP constants
* Fix bugs in frame layout for function calls and incoming arguments
* Insert copies for constant arguments to PHI nodes into the BOTTOM of
predecessor blocks, not the top.
* Implement _floating point_ support: setcc, return, load, store, cast
* Fix several bugs in the cast instruction
* Abstract out call emission and load/store for FP
* Implement malloc/free without previous lowering pass.
* Make use of new forms of MachineOperand
* Implement _long_ support!
* Fix many bugs in FP support
* Change branch from je/jne to je/jmp
* Simplify code generated for GEP instructions
llvm-svn: 5223
* Only load incoming arguments into virtual registers once at the
beginning of the function
* Assign different virtual registers to each reference to constants/globals
llvm-svn: 5088
Fix some bugs in use of MBB vs. BB and iterators that are invalidated before
we use them.
Reference targetClass by enum name, not by number.
llvm-svn: 5069
The MachineBasicBlock variable name patrol hereby fines Chris Lattner
one bag of nachos, for shadowing global names while his license to do so
was under suspension.
llvm-svn: 5014
because arguments start two stack slots off of EBP. Break out of the
for loop once the argument is found. Increment the counter at the end
of the loop instead of the beginning. Use addRegOffset and compute
the scale * index part at compile time instead of using the fancy
load instruction. Just because an instruction set has wacky addressing
modes doesn't mean we ought to use them (at least, if you believe Dave
Patterson).
lib/Target/X86/X86InstrBuilder.h: Add some comments.
test/Regression/Jello/test-loadstore.ll: Let main return int 0.
llvm-svn: 4999
implementation of getReg() for arguments.
MachineCodeEmitter.cpp: Fix using EBP with index, scale and no
displacement (whew!) due to Chris.
Printer.cpp: Fix printing out index and scale in memory references.
llvm-svn: 4998
InstSelectSimple.cpp: Add stub implementation of visitFreeInst.
Add comments that mention how we are failing to implement malloc/free.
Add initial implementation of visitAllocaInst.
X86TargetMachine.cpp: Include llvm/Transforms/Scalar.h.
Add LowerAllocations pass before instruction selection.
jello/Makefile: Add scalaropts.a.
llvm-svn: 4994
(lib/Target/X86) InstSelectSimple.cpp:
Include llvm/DerivedTypes.h and iostream.
Refactor visitMul out into a wrapper around doMultiply(), so that we
can do multiplications on temporary values when we are doing
getelementptrs.
Refactor part of getReg out into makeAnotherReg, so that we can create
registers willy-nilly to hold temporary values, when we are doing
getelementptrs.
Add stub implementations of visitMallocInst and visitAllocaInst.
Add initial implementation of visitGetElementPtrInst.
In copyConstantToRegister:
We throw a *lot* of our asserts here. So, when we want to throw an
assert, print out to stderr whatever expr or whatever constant made
us barf.
Support copying ConstantPointerNull to register, using a move immediate
of zero.
Rename FLDr4 and FLDr8 to FLDr32 and FLDr64, so that they match the meanings
of the numbers in the other instruction names. All uses modified.
Teach visitCallInst to extract byte- and short-class return values
from subregs of EAX. Add a FIXME note about how we would do it for
float-class return values.
Add a FIXME note about how we would cast float to int and back.
X86InstrInfo.def:
Rename FLDr4 and FLDr8 to FLDr32 and FLDr64, so that they match the meanings
of the numbers in the other instruction names. All uses modified.
(tools/jello) GlobalVars.cpp:
Include iostream.
If we have to emit a floating-point constant to memory, gamble and use
the same method as for ints.
If we have to emit a ConstantPointerNull to memory, try using a "void *"
and "NULL".
Otherwise, if we are going to throw an assert, print out whatever constant
made us barf, first.
llvm-svn: 4973
InstSelectSimple.cpp: Refactor out conversion of byte, short -> int
from visitReturnInst() to new method, promote32().
Use it in both visitReturnInst() and visitCallInst().
llvm-svn: 4839
InstSelectSimple.cpp: First draft of visitCallInst method, handling
int/float args.
X86InstrInfo.def: Add entries for CALL with 32-bit pc relative arg, and
PUSH with 32-bit reg arg.
llvm-svn: 4838
lib/Target/X86/InstSelectSimple.cpp: Add a little something to
visitBranchInst which supports conditional branches.
lib/Target/X86/X86InstrInfo.def: Add defs of JNE, JE, CMPri8
llvm-svn: 4755
convenience method. Fix typo in comment.
lib/Target/X86/InstSelectSimple.cpp: Explicitly specify some implicit uses.
Use MOVZX/MOVSX instead of MOV instructions with sign extend instructions.
Take out LEAVE instructions.
32-bit IDIV and DIV use CDQ, not CWQ (CWQ is a typo).
Fix typo in comment and remove some FIXME comments.
lib/Target/X86/Printer.cpp: Include X86InstrInfo.h and llvm/Function.h.
Add some simple code to Printer::runOnFunction to iterate over
MachineBasicBlocks and call X86InstrInfo::print().
lib/Target/X86/X86InstrInfo.def: Make some more instructions with
implicit defs "Void". Add more sign/zero extending "move" insns
(movsx, movzx).
lib/Target/X86/X86RegisterInfo.def: Add EFLAGS as a register.
llvm-svn: 4707
Add ISel::visitShiftInst() to instruction select shift instructions.
Add a comment in visitAdd about how to do 64 bit adds.
X86InstrInfo.def: Add register-to-register move opcodes and shift opcodes.
llvm-svn: 4477
