Adding support for intrinsics of AMX-BF16.
This patch alse fix a bug that AMX-INT8 instructions will be selected with wrong
predicate.
Differential Revision: https://reviews.llvm.org/D97358
This is an optimized approach for D94155.
Previous code build the model that tile config register is the user of
each AMX instruction. There is a problem for the tile config register
spill. When across function, the ldtilecfg instruction may be inserted
on each AMX instruction which use tile config register. This cause all
tile data register clobber.
To fix this issue, we remove the model of tile config register. Instead,
we analyze the AMX instructions between one call to another. We will
insert ldtilecfg after the first call if we find any AMX instructions.
Reviewed By: LuoYuanke
Differential Revision: https://reviews.llvm.org/D95136
Revert the change to use APInt::isSignedIntN from
5ff5cf8e05.
Its clear that the games we were playing to avoid the topological
sort aren't working. So just fix it once and for all.
Fixes PR48888.
Previous code build the model that tile config register is the user of
each AMX instruction. There is a problem for the tile config register
spill. When across function, the ldtilecfg instruction may be inserted
on each AMX instruction which use tile config register. This cause all
tile data register clobber.
To fix this issue, we remove the model of tile config register. We
analyze the regmask of call instruction and insert ldtilecfg if there is
any tile data register live across the call. Inserting the sttilecfg
before the call is unneccessary, because the tile config doesn't change
and we can just reload the config.
Besides we also need check tile config register interference. Since we
don't model the config register we should check interference from the
ldtilecfg to each tile data register def.
ldtilecfg
/ \
BB1 BB2
/ \
call BB3
/ \
%1=tileload %2=tilezero
We can start from the instruction of each tile def, and backward to
ldtilecfg. If there is any call instruction, and tile data register is
not preserved, we should insert ldtilecfg after the call instruction.
Differential Revision: https://reviews.llvm.org/D94155
The x86_amx is used for AMX intrisics. <256 x i32> is bitcast to x86_amx when
it is used by AMX intrinsics, and x86_amx is bitcast to <256 x i32> when it
is used by load/store instruction. So amx intrinsics only operate on type x86_amx.
It can help to separate amx intrinsics from llvm IR instructions (+-*/).
Thank Craig for the idea. This patch depend on https://reviews.llvm.org/D87981.
Differential Revision: https://reviews.llvm.org/D91927
The ABI explains that %fs:(%eax) zero-extends %eax to 64 bits, and adds
that the TLS base address, but that the TLS base address need not be
at the start of the TLS block, TLS references may use negative offsets.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D93158
This patch implements amx programming model that discussed in llvm-dev
(http://lists.llvm.org/pipermail/llvm-dev/2020-August/144302.html).
Thank Hal for the good suggestion in the RA. The fast RA is not in the patch yet.
This patch implemeted 7 components.
1. The c interface to end user.
2. The AMX intrinsics in LLVM IR.
3. Transform load/store <256 x i32> to AMX intrinsics or split the
type into two <128 x i32>.
4. The Lowering from AMX intrinsics to AMX pseudo instruction.
5. Insert psuedo ldtilecfg and build the def-use between ldtilecfg to amx
intruction.
6. The register allocation for tile register.
7. Morph AMX pseudo instruction to AMX real instruction.
Change-Id: I935e1080916ffcb72af54c2c83faa8b2e97d5cb0
Differential Revision: https://reviews.llvm.org/D87981
Pretty sure we meant to be checking signed 32 immediates here
rather than unsigned 32 bit. I suspect I messed this up because
in MathExtras.h we have isIntN and isUIntN so isIntN differs in
signedness depending on whether you're using APInt or plain integers.
This fixes a case where we didn't fold a constant created
by shrinkAndImmediate. Since shrinkAndImmediate doesn't topologically
sort constants it creates, we can fail to convert the Constant
to a TargetConstant. This leads to very strange behavior later.
Fixes PR48458.
Since x32 supports PC-relative address, it shouldn't use EBX for TLS
address. Instead of checking N.getValueType(), we should check
Subtarget->is32Bit(). This fixes PR 22676.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D16474
This uses PreprocessISelDAG to replace the constant before
instruction selection instead of matching opcodes after.
Reviewed By: pengfei
Differential Revision: https://reviews.llvm.org/D89178
In small code model, program and its symbols are linked in the lower 2 GB of
the address space. Try encoding global address even when the range is unknown
in such case.
Differential Revision: https://reviews.llvm.org/D89341
We should avoid emitting MachineSDNodes from lowering.
We can use the the implicit def handling in InstrEmitter to avoid
manually copying from each xmm result register. We only need to
manually emit the copies for the implicit uses.
Instead of emitting MachineSDNodes during lowering, emit X86ISD
opcodes. These opcodes will either be selected by tablegen
patterns or custom selection code.
Emitting MachineSDNodes during lowering is uncommon so this makes
things more consistent. It also allows selectAddr to be called to
perform address matching during instruction selection.
I had trouble getting tablegen to accept XMM0-XMM7 as results in
an isel pattern for the WIDE instructions so I had to use custom
instruction selection.
This is the type declared in X86InstrFragmentsSIMD.td. ISel pattern
matching doesn't check so it doesn't matter in practice. Maybe for
SelectionDAG CSE it would matter.
By factoring out the end of tryVPTERNLOG, we can use the same code
to directly match X86ISD::VPTERNLOG. This allows us to remove
around 3-4K worth of X86GenDAGISel.inc.
When we use mask compare intrinsics under strict FP option, the masked
elements shouldn't raise any exception. So, we cann't replace the
intrinsic with a full compare + "and" operation.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D85385
Now we try to load and broadcast together for operand 1. Followed
by load and broadcast for operand 1. Previously we tried load
operand 1, load operand 1, broadcast operand 0, broadcast operand 1.
Now we have a single helper that tries load and broadcast for
one operand that we can just call twice.
Rather than hardcoding immediate values for 12 different combinations
in a nested pair of switches, we can perform the matched logic
operation on 3 magic constants to calculate the immediate.
Special thanks to this tweet https://twitter.com/rygorous/status/1187034321992871936
for making me realize I could do this.
Previously we just matched the logic ops and replaced with an
X86ISD::VPTERNLOG node that we would send through the normal
pattern match. But that approach couldn't handle a bitcast
between the logic ops. Extending that approach would require us
to peek through the bitcasts and emit new bitcasts to match
the types. Those new bitcasts would then have to be properly
topologically sorted.
This patch instead switches to directly emitting the
MachineSDNode and skips the normal tablegen pattern matching.
We do have to handle load folding and broadcast load folding
ourselves now. Which also means commuting the immediate control.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D83630
I think this mostly looks ok. The only weird thing I noticed was
a couple rotate vXi8 tests picked up an extra logic op where we have
(and (or (and), (andn)), X). Previously we matched the (or (and), (andn))
to vpternlog, but now we match the (and (or), X) and leave the and/andn
unmatched.
This function picks X86 opcode name based on type, masking,
and whether not a load or broadcast has been folded using multiple
switch statements. The contents of the switches mostly just vary in
a few characters in the instruction name. So use some macros to
build the instruction names to reduce the repetiveness.
> relocImm was a complexPattern that handled both ConstantSDNode
> and X86Wrapper. But it was only applied selectively because using
> it would cause patterns to be not importable into FastISel or
> GlobalISel. So it only got applied to flag setting instructions,
> stores, RMW arithmetic instructions, and rotates.
>
> Most of the test changes are a result of making patterns available
> to GlobalISel or FastISel. The absolute-cmp.ll change is due to
> this fixing a pattern ordering issue to make an absolute symbol
> match to an 8-bit immediate before trying a 32-bit immediate.
>
> I tried to use PatFrags to reduce the repetition, but I was getting
> errors from TableGen.
This caused "Invalid EmitNode" assertions, see the llvm-commits thread for
discussion.
relocImm was a complexPattern that handled both ConstantSDNode
and X86Wrapper. But it was only applied selectively because using
it would cause patterns to be not importable into FastISel or
GlobalISel. So it only got applied to flag setting instructions,
stores, RMW arithmetic instructions, and rotates.
Most of the test changes are a result of making patterns available
to GlobalISel or FastISel. The absolute-cmp.ll change is due to
this fixing a pattern ordering issue to make an absolute symbol
match to an 8-bit immediate before trying a 32-bit immediate.
I tried to use PatFrags to reduce the repetition, but I was getting
errors from TableGen.
As shown in PR46237:
https://bugs.llvm.org/show_bug.cgi?id=46237
The size-savings win for hoisting an 8-bit ALU immediate (intentionally
excluding store constants) requires extreme conditions; it may not even
be possible when including REX prefix bytes on x86-64.
I did draft a version of this patch that included use counts after the
loop, but I suspect that accounting is not working as expected. I think
that is because the number of constant uses are changing as we select
instructions (for example as we transform shl/add into LEA).
Differential Revision: https://reviews.llvm.org/D81468
The instruction is defined to only produce high result if both
destinations are the same. We can exploit this to avoid
unnecessarily clobbering a register.
In order to hide this from register allocation we use a pseudo
instruction and expand the result during MCInst creation.
Differential Revision: https://reviews.llvm.org/D80500
Liu, Chen3