Quentin Colombet [Tue, 11 Nov 2014 02:23:47 +0000 (02:23 +0000)]
[X86] Custom lower UINT_TO_FP from v4f32 to v4i32, and for v8f32 to v8i32 if
AVX2 is available.
According to IACA, the new lowering has a throughput of 8 cycles instead of 13
with the previous one.
Althought this lowering kicks in some SPECs benchmarks, the performance
improvement was within the noise.
Correctness testing has been done for the whole range of uint32_t with the
following program:
uint4 v = (uint4) {0,1,2,3};
uint32_t i;
//Check correctness over entire range for uint4 -> float4 conversion
for( i = 0; i < 1U << (32-2); i++ )
{
float4 t = test(v);
float4 c = correct(v);
if( 0xf != _mm_movemask_ps( t == c ))
{
printf( "Error @ %vx: %vf vs. %vf\n", v, c, t);
return -1;
}
v += 4;
}
Where "correct" is the old lowering and "test" the new one.
The patch adds a test case for the two custom lowering instruction.
It also modifies the vector cost model, which is why cast.ll and uitofp.ll are
modified.
2009-02-26-MachineLICMBug.ll is also modified because we now hoist 7
instructions instead of 4 (3 more constant loads).
Juergen Ributzka [Mon, 10 Nov 2014 21:05:31 +0000 (21:05 +0000)]
[AArch64][FastISel] Fix kill flags for integer extends.
In the case we optimize an integer extend away and replace it directly with the
source register, we also have to clear all kill flags at all its uses.
This is necessary, because the orignal IR instruction might be trivially dead,
but we replaced it with a nop at MI level.
Juergen Ributzka [Mon, 10 Nov 2014 21:05:27 +0000 (21:05 +0000)]
[SwitchLowering] Fix the "fixPhis" function.
Switch statements may have more than one incoming edge into the same BB if they
all have the same value. When the switch statement is converted these incoming
edges are now coming from multiple BBs. Updating all incoming values to be from
a single BB is incorrect and would generate invalid LLVM IR.
The fix is to only update the first occurrence of an incoming value. Switch
lowering will perform subsequent calls to this helper function for each incoming
edge with a new basic block - updating all edges in the process.
We already use the llvm namespace. Remove the unnecessary prefix. Use the
StringRef::equals method to compare with C strings rather than instantiating
std::strings.
Addresses late review comments from David Majnemer.
Visual Studio 2012 apparently does not support using alias declarations. Use
the more traditional typedef approach. This should let the Windows buildbots
pass. NFC.
This introduces the symbol rewriter. This is an IR->IR transformation that is
implemented as a CodeGenPrepare pass. This allows for the transparent
adjustment of the symbols during compilation.
It provides a clean, simple, elegant solution for symbol inter-positioning. This
technique is often used, such as in the various sanitizers and performance
analysis.
The control of this is via a custom YAML syntax map file that indicates source
to destination mapping, so as to avoid having the compiler to know the exact
details of the source to destination transformations.
Daniel Sanders [Fri, 7 Nov 2014 16:54:21 +0000 (16:54 +0000)]
[mips] Promote i32 arguments to i64 for the N32/N64 ABI and fix <64-bit structs...
Summary:
... and after all that refactoring, it's possible to distinguish softfloat
floating point values from integers so this patch no longer breaks softfloat to
do it.
Remove direct handling of i32's in the N32/N64 ABI by promoting them to
i64. This more closely reflects the ABI documentation and also fixes
problems with stack arguments on big-endian targets.
We now rely on signext/zeroext annotations (already generated by clang) and
the Assert[SZ]ext nodes to avoid the introduction of unnecessary sign/zero
extends.
It was not possible to convert three tests to use signext/zeroext. These tests
are bswap.ll, ctlz-v.ll, ctlz-v.ll. It's not possible to put signext on a
vector type so we just accept the sign extends here for now. These tests don't
pass the vectors the same way clang does (clang puts multiple elements in the
same argument, these map 1 element to 1 argument) so we don't need to worry too
much about it.
With this patch, all known N32/N64 bugs should be fixed and we now pass the
first 10,000 tests generated by ABITest.py.
Daniel Sanders [Fri, 7 Nov 2014 15:33:08 +0000 (15:33 +0000)]
[mips] Removed the remainder of MipsCC. NFC.
Summary:
One of the calls to AllocateStack (the one in LowerCall) doesn't look like
it should be there but it was there before and removing it breaks the
frame size calculation.
Daniel Sanders [Fri, 7 Nov 2014 11:10:48 +0000 (11:10 +0000)]
[mips] Move SpecialCallingConv to MipsCCState and use it from tablegen-erated code. NFC
Summary:
In the long run, it should probably become a calling convention in its own
right but for now just move it out of
MipsISelLowering::analyzeCallOperands() so that we can drop this function
in favour of CCState::AnalyzeCallOperands().
Jay Foad [Fri, 7 Nov 2014 09:08:39 +0000 (09:08 +0000)]
llvm-symbolizer: teach it about PowerPC64 ELF function descriptors
Summary:
Teach llvm-symbolizer about PowerPC64 ELF function descriptors. Symbols in the .opd section point to function descriptors, the first word of which is a pointer to the real function. For the purposes of symbolizing we pretend that the symbol points directly to the function.
This is enough to get decent function names in stack traces for unoptimized binaries, which fixes the sanitizer print-stack-trace test on PowerPC64 Linux.
David Majnemer [Fri, 7 Nov 2014 08:54:19 +0000 (08:54 +0000)]
SCCP: overdefined calls cannot become constant
We would attempt to fold away a call instruction which had been marked
overdefined. However, it's not valid to transition to constant from
overdefined.
These match the `-pic-level' command line argument for clang, and the value of the
preprocessor macro `__PIC__'.
Test Plan:
New flags tests specific for the 'PIC Level' module flag.
Tests to be added as part of a future commit for PowerPC, which will use this new API.
Ahmed Bougacha [Fri, 7 Nov 2014 02:50:00 +0000 (02:50 +0000)]
[AArch64] Keep flags on condition vreg when instantiating a CB branch.
Reversing a CB* instruction used to drop the flags on the condition. On the
included testcase, this lead to a read from an undefined vreg.
Using addOperand keeps the flags, here <undef>.
Fixed an issue with the (v)cvttps2dq and (v)cvttpd2dq instructions being incorrectly put in the 2 source operand folding tables instead of the 1 source operand and added the missing SSE/AVX versions.
Also added missing (v)cvtps2dq and (v)cvtpd2dq instructions to the folding tables.
Base check on the section name, not the variable name.
The variable is private, so the name should not be relied on. Also, the
linker uses the sections, so asan should too when trying to avoid causing
the linker problems.
Frederic Riss [Thu, 6 Nov 2014 17:46:55 +0000 (17:46 +0000)]
Change DIBuilder::createImportedDeclaration from taking a DIScope to a DIDescriptor.
Imported declarations can be DIGlobalVariables which aren't a DIScope. Today
clang (unknowingly I believe) shoehorns these into a DIScope and it all works
just because we never access the fields.
Chad Rosier [Thu, 6 Nov 2014 16:46:37 +0000 (16:46 +0000)]
[Reassociate] Don't reassociate when mixing regular and fast-math FP
instructions. Inlining might cause such cases and it's not valid to
reassociate floating-point instructions without the unsafe algebra flag.
Daniel Sanders [Thu, 6 Nov 2014 16:36:30 +0000 (16:36 +0000)]
[mips] Removed MipsISelLowering::analyzeFormalArguments() in favour of CCState::AnalyzeFormalArguments()
Summary:
As with returns, we must be able to identify f128 arguments despite them
being lowered away. We do this with a pre-analyze step that builds a
vector and then we use this vector from the tablegen-erated code.
Toma Tabacu [Thu, 6 Nov 2014 14:25:42 +0000 (14:25 +0000)]
[mips] Tolerate the use of the %z inline asm operand modifier with non-immediates.
Summary:
Currently, we give an error if %z is used with non-immediates, instead of continuing as if the %z isn't there.
For example, you use the %z operand modifier along with the "Jr" constraints ("r" makes the operand a register, and "J" makes it an immediate, but only if its value is 0).
In this case, you want the compiler to print "$0" if the inline asm input operand turns out to be an immediate zero and you want it to print the register containing the operand, if it's not.
We give an error in the latter case, and we shouldn't (GCC also doesn't).
Sasa Stankovic [Thu, 6 Nov 2014 13:20:12 +0000 (13:20 +0000)]
[mips] Add the following MIPS options that control gp-relative addressing of
small data items: -mgpopt, -mlocal-sdata, -mextern-sdata. Implement gp-relative
addressing for constants.
Toma Tabacu [Thu, 6 Nov 2014 10:02:45 +0000 (10:02 +0000)]
[mips] Improve error/warning messages and testing for the .cpload assembler directive.
Summary:
Improved warning message when using .cpload inside a reorder section and added an error message for using .cpload with Mips16 enabled.
Modified the tests to fit with the changes mentioned above, added a test-case for the N32 ABI in cpload.s and did some reformatting to make the tests easier to read.
Daniel Sanders [Thu, 6 Nov 2014 09:53:05 +0000 (09:53 +0000)]
[JIT] Fix more missing endian conversions (opcodes for AArch64, ARM, and Mips stub functions, and ARM target in general)
Summary:
Fixed all of the missing endian conversions that Lang Hames and I identified in
RuntimeDyldMachOARM.h.
Fixed the opcode emission in RuntimeDyldImpl::createStubFunction() for AArch64,
ARM, Mips when the host endian doesn't match the target endian.
PowerPC will need changing if it's opcodes are affected by endianness but I've
left this for now since I'm unsure if this is the case and it's the only path
that specifies the target endian.
This patch fixes MachO_ARM_PIC_relocations.s on a big-endian Mips host. This
is the last of the known issues on this host.
David Majnemer [Thu, 6 Nov 2014 08:10:41 +0000 (08:10 +0000)]
Object, COFF: Infer symbol sizes from adjacent symbols
Use the position of the subsequent symbol in the object file to infer
the size of it's predecessor. I hope to eventually remove whatever COFF
specific details from this little algorithm so that we can unify this
logic with what Mach-O does.
Justin Bogner [Thu, 6 Nov 2014 06:55:02 +0000 (06:55 +0000)]
GCOV: Make sure that function idents in the .gcda and .gcno match
When generating gcov compatible profiling, we sometimes skip emitting
data for functions for one reason or another. However, this was
emitting different function IDs in the .gcno and .gcda files, because
the .gcno case was using the loop index before skipping functions and
the .gcda the array index after. This resulted in completely invalid
gcov data.
This fixes the problem by making the .gcno loop track the ID
separately from the loop index.
[X86] Lower VSELECT into SHRUNKBLEND when we shrink the bits used into the
condition to match a blend.
This prevents optimizations that work on VSELECT to perform invalid
transformations. Indeed, the optimized condition does not match the vector
boolean content that is expected and bad things may happen.
This patch yields the exact same code on the whole test-suite + specs (-O3 and
-O3 -march=core-avx2), it improves one test case (vector-blend.ll) and fixes a
bug reduced in vselect-avx.ll.
Simon Pilgrim [Wed, 5 Nov 2014 22:28:25 +0000 (22:28 +0000)]
[X86][SSE] Vector integer to float conversion memory folding
Added missing memory folding for the (V)CVTDQ2PS instructions - we can safely fold these (but not the (V)CVTDQ2PD versions which have a register/memory size discrepancy in the source operand). I've added a test case demonstrating that stack folding now works.
projects / llvm / log