- Still O(N^2), just a faster form, and now its the MCAsmLayout's fault.
On the .s I am tuning against (combine.s from 403.gcc):
--
ddunbar@lordcrumb:MC$ diff stats-before.txt stats-after.txt
5,10c5,10
< 1728 assembler - Number of assembler layout and relaxation steps
< 7707 assembler - Number of emitted assembler fragments
< 120588 assembler - Number of emitted object file bytes
< 2233448 assembler - Number of evaluated fixups
< 1727 assembler - Number of relaxed instructions
< 6723845 mcexpr - Number of MCExpr evaluations
---
> 3 assembler - Number of assembler layout and relaxation steps
> 7707 assembler - Number of emitted assembler fragments
> 120588 assembler - Number of emitted object file bytes
> 14796 assembler - Number of evaluated fixups
> 1727 assembler - Number of relaxed instructions
> 67889 mcexpr - Number of MCExpr evaluations
--
Feel free to LOL at the -before numbers, if you like.
I am a little surprised we make more than 2 relaxation passes. It's pretty
trivial for us to do relaxation out-of-order if that would give a speedup.
llvm-svn: 99543
the custom insertion hook deletes the instruction, then we try to set dead
flags on it. Neither the code that I added nor the code that was there
before was safe.
llvm-svn: 99538
On Nehalem and newer CPUs there is a 2 cycle latency penalty on using a register
in a different domain than where it was defined. Some instructions have
equvivalents for different domains, like por/orps/orpd.
The SSEDomainFix pass tries to minimize the number of domain crossings by
changing between equvivalent opcodes where possible.
This is a work in progress, in particular the pass doesn't do anything yet. SSE
instructions are tagged with their execution domain in TableGen using the last
two bits of TSFlags. Note that not all instructions are tagged correctly. Life
just isn't that simple.
The SSE execution domain issue is very similar to the ARM NEON/VFP pipeline
issue handled by NEONMoveFixPass. This pass may become target independent to
handle both.
llvm-svn: 99524
now configures prerequisite projects individually but also ignores them in the
big project switch statement to avoid the incorrect warning.
llvm-svn: 99506
bytes instead of one byte. This is important because
we're running up to too many opcodes to fit in a byte
and it is aggrevated by FIRST_TARGET_MEMORY_OPCODE
making the numbering sparse. This just bites the
bullet and bloats out the table. In practice, this
increases the size of the x86 isel table from 74.5K
to 76K. I think we'll cope :)
This fixes rdar://7791648
llvm-svn: 99494
If a TableGen class has an initializer expression containing an X.Y subexpression,
AND X depends on template parameters,
AND those template parameters have defaults,
AND some parameters with defaults are beyond position 1,
THEN parts of the initializer expression are evaluated prematurely with the default values when the first explicit template parameter is substituted, before the remaining explicit template parameters have been substituted.
llvm-svn: 99492
happening.
Enhance scheduling to set the DEAD flag on implicit defs
more aggressively. Before, we'd set an implicit def operand
to dead if it were present in the SDNode corresponding to
the machineinstr but had no use. Now we do it in this case
AND if the implicit def does not exist in the SDNode at all.
This exposes a couple of problems: one is the FIXME, which
causes a live intervals crash on CodeGen/X86/sibcall.ll.
The second is that it makes machinecse and licm more
aggressive (which is a good thing) but also exposes a case
where licm hoists a set0 and then it doesn't get resunk.
Talking to codegen folks about both these issues, but I need
this patch in in the meantime.
llvm-svn: 99485
Here is a theoretical example that illustrates why the placement is important.
tmp1 =
store tmp1 -> x
...
tmp2 = add ...
...
call
...
store tmp2 -> x
Now mem2reg comes along:
tmp1 =
dbg_value (tmp1 -> x)
...
tmp2 = add ...
...
call
...
dbg_value (tmp2 -> x)
When the debugger examine the value of x after the add instruction but before the call, it should have the value of tmp1.
Furthermore, for dbg_value's that reference constants, they should not be emitted at the beginning of the block (since they do not have "producers").
This patch also cleans up how SDISel manages DbgValue nodes. It allow a SDNode to be referenced by multiple SDDbgValue nodes. When a SDNode is deleted, it uses the information to find the SDDbgValues and invalidate them. They are not deleted until the corresponding SelectionDAG is destroyed.
llvm-svn: 99469
addl $12, %esp
popl %esi
popl %edi
popl %ebx
popl %ebp
jmpl *__Block_deallocator-L1$pb(%esi) # TAILCALL
The problem is the global base register is assigned GR32 register class. TCRETURNmi needs the registers making up the address mode to have the GR32_TC register class.
The *proper* fix is for X86DAGToDAGISel::getGlobalBaseReg() to return a copy from the global base register of the machine function rather than returning the register itself. But that has the potential of causing it to be coalesced to a more restrictive register class: GR32_TC. It can introduce additional copies and spills. For something as important the PIC base, it's not worth it especially since this is not an issue on 64-bit.
llvm-svn: 99455
--- Reverse-merging r99440 into '.':
U test/MC/AsmParser/X86/x86_32-bit_cat.s
U test/MC/AsmParser/X86/x86_32-encoding.s
U include/llvm/IntrinsicsX86.td
U include/llvm/CodeGen/SelectionDAGNodes.h
U lib/Target/X86/X86InstrSSE.td
U lib/Target/X86/X86ISelLowering.h
llvm-svn: 99450
Evan Cheng