indirect branch at the end of the BB. Otherwise if-converter, branch folding
pass may incorrectly update its successor info if it consider BB as fallthrough
to the next BB.
rdar://13782395
llvm-svn: 181161
Now even the small structures could be passed within byval (small enough
to be stored in GPRs).
In regression tests next function prototypes are checked:
PR15293:
%artz = type { i32 }
define void @foo(%artz* byval %s)
define void @foo2(%artz* byval %s, i32 %p, %artz* byval %s2)
foo: "s" stored in R0
foo2: "s" stored in R0, "s2" stored in R2.
Next AAPCS rules are checked:
5.5 Parameters Passing, C.4 and C.5,
"ParamSize" is parameter size in 32bit words:
-- NSAA != 0, NCRN < R4 and NCRN+ParamSize > R4.
Parameter should be sent to the stack; NCRN := R4.
-- NSAA != 0, and NCRN < R4, NCRN+ParamSize < R4.
Parameter stored in GPRs; NCRN += ParamSize.
llvm-svn: 181148
X86ISelLowering has support to treat:
(icmp ne (and (xor %flags, -1), (shl 1, flag)), 0)
as if it were actually:
(icmp eq (and %flags, (shl 1, flag)), 0)
However, r179386 has code at the InstCombine level to handle this.
llvm-svn: 181145
Add support for min/max reductions when "no-nans-float-math" is enabled. This
allows us to assume we have ordered floating point math and treat ordered and
unordered predicates equally.
radar://13723044
llvm-svn: 181144
Add support for matching 'ordered' and 'unordered' floating point min/max
constructs.
In LLVM we can express min/max functions as a combination of compare and select.
We have support for matching such constructs for integers but not for floating
point. In floating point math there is no total order because of the presence of
'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics
when interpreting floating point compare select combinations as a minimum or
maximum function. The resulting 'ordered/unordered' floating point maximum
function has to select the same value as the select/fcmp combination it is based
on.
ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise
unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise
ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise
unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise
This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R)
construct.
Any code using this predicate has to preserve this semantics.
A follow-up patch will use this to implement floating point min/max reductions
in the vectorizer.
radar://13723044
llvm-svn: 181143
This is about the simplest relocation, but surprisingly rare in actual
code.
It occurs in (for example) the MCJIT test test-ptr-reloc.ll.
llvm-svn: 181134
As with global accesses, external functions could exist anywhere in
memory. Therefore the stub must create a complete 64-bit address. This
patch implements the fragment as (roughly):
movz x16, #:abs_g3:somefunc
movk x16, #:abs_g2_nc:somefunc
movk x16, #:abs_g1_nc:somefunc
movk x16, #:abs_g0_nc:somefunc
br x16
In principle we could save 4 bytes by using a literal-load instead,
but it is unclear that would be more efficient and can only be tested
when real hardware is readily available.
This allows (for example) the MCJIT test 2003-05-07-ArgumentTest to
pass on AArch64.
llvm-svn: 181133
The large memory model (default and main viable for JIT) emits
addresses in need of relocation as
movz x0, #:abs_g3:somewhere
movk x0, #:abs_g2_nc:somewhere
movk x0, #:abs_g1_nc:somewhere
movk x0, #:abs_g0_nc:somewhere
To support this we must implement those four relocations in the
dynamic loader.
This allows (for example) the test-global.ll MCJIT test to pass on
AArch64.
llvm-svn: 181132
R_AARCH64_PCREL32 is present in even trivial .eh_frame sections and so
is required to compile any function without the "nounwind" attribute.
This change implements very basic infrastructure in the RuntimeDyldELF
file and allows (for example) the test-shift.ll MCJIT test to pass
on AArch64.
llvm-svn: 181131
AArch64 is going to need some kind of cache-invalidation in order to
successfully JIT since it has a weak memory-model. This is provided by
a __clear_cache builtin in libgcc, which acts very much like the
32-bit ARM equivalent (on platforms where it exists).
llvm-svn: 181129
This let us to remove some custom code that matched constant offsets
from globals at instruction selection time as a special addressing mode.
No intended functionality change.
llvm-svn: 181126
The code now makes use of ComputeMaskedBits,
SelectionDAG::isBaseWithConstantOffset and TargetLowering::isGAPlusOffset
where appropriate reducing the amount of logic needed in XCoreISelLowering.
No intended functionality change.
llvm-svn: 181125
Thread local storage is not supported by the XMOS linker so we handle
thread local variables by lowering the variable to an array of n elements
(where n is the number of hardware threads per core, currently 8
for all XMOS devices) indexed by the the current thread ID.
Previously this lowering was spread across the XCoreISelLowering and the
XCoreAsmPrinter classes. Moving this to a separate pass should be much
cleaner.
llvm-svn: 181124
The MOVZ/MOVK instruction sequence may not be the most efficient (a
literal-pool load could be better) but adding that would require
reinstating the ConstantIslands pass.
For now the sequence is correct, and that's enough. Beware, as of
commit GNU ld does not appear to support the relocations needed for
this. Its primary purpose (for now) will be to support JITed code,
since in that case there is no guarantee of where your code will end
up in memory relative to external symbols it references.
llvm-svn: 181117
Evan Cheng