Get rid of the "; mem:" suffix and use the one we use in MIR: ":: (load 2)".
rdar://38163529
Differential Revision: https://reviews.llvm.org/D42377
llvm-svn: 327580
Optionally allow the order of restoring the callee-saved registers in the
epilogue to be reversed.
The flag -reverse-csr-restore-seq generates the following code:
```
stp x26, x25, [sp, #-64]!
stp x24, x23, [sp, #16]
stp x22, x21, [sp, #32]
stp x20, x19, [sp, #48]
; [..]
ldp x24, x23, [sp, #16]
ldp x22, x21, [sp, #32]
ldp x20, x19, [sp, #48]
ldp x26, x25, [sp], #64
ret
```
Note how the CSRs are restored in the same order as they are saved.
One exception to this rule is the last `ldp`, which allows us to merge
the stack adjustment and the ldp into a post-index ldp. This is done by
first generating:
ldp x26, x27, [sp]
add sp, sp, #64
which gets merged by the arm64 load store optimizer into
ldp x26, x25, [sp], #64
The flag is disabled by default.
llvm-svn: 327569
I removed this in r316797 because the coverage report showed no coverage and I thought it should have been handled by the auto generated table. I now see that there is code that bypasses the table if the shift amount is out of bounds.
This adds back the code. We'll codegen out of bounds i8 shifts to effectively (amount & 0x1f). The 0x1f is a strange quirk of x86 that shift amounts are always masked to 5-bits(except 64-bits). So if the masked value is still out bounds the result will be 0.
Fixes PR36731.
llvm-svn: 327540
I had to modify the bswap recognition to allow unshrunk masks to make this work.
Fixes PR36689.
Differential Revision: https://reviews.llvm.org/D44442
llvm-svn: 327530
swifterror llvm values model the swifterror register as memory at the
LLVM IR level. ISel will perform adhoc mem-to-reg on them. swifterror
values are constraint in how they can be used. Spilling them to memory
is not allowed.
SjLjEHPrepare tried to lower swifterror values to memory which is
unecessary since the back-end will spill and reload the register as
neccessary (as long as clobbering calls are marked as such which is the
case here) and further leads to invalid IR because swifterror values
can't be stored to memory.
rdar://38164004
llvm-svn: 327521
Support G_LSHR/G_ASHR/G_SHL. We have 3 variance for
shift instructions : shift gpr, shift imm, shift 1.
Currently GlobalIsel TableGen generate patterns for
shift imm and shift 1, but with shiftCount i8.
In G_LSHR/G_ASHR/G_SHL like LLVM-IR both arguments
has the same type, so for now only shift i8 can use
auto generated TableGen patterns.
The support of G_SHL/G_ASHR enables tryCombineSExt
from LegalizationArtifactCombiner.h to hit, which
results in different legalization for the following tests:
LLVM :: CodeGen/X86/GlobalISel/ext-x86-64.ll
LLVM :: CodeGen/X86/GlobalISel/gep.ll
LLVM :: CodeGen/X86/GlobalISel/legalize-ext-x86-64.mir
-; X64-NEXT: movsbl %dil, %eax
+; X64-NEXT: movl $24, %ecx
+; X64-NEXT: # kill: def $cl killed $ecx
+; X64-NEXT: shll %cl, %edi
+; X64-NEXT: movl $24, %ecx
+; X64-NEXT: # kill: def $cl killed $ecx
+; X64-NEXT: sarl %cl, %edi
+; X64-NEXT: movl %edi, %eax
..which is not optimal and should be addressed later.
Rework of the patch by igorb
Reviewed By: igorb
Differential Revision: https://reviews.llvm.org/D44395
llvm-svn: 327499
This better able to detect undef and zeros pieces in the concat. Or cases when only one subvector is non-zero. This allows us to avoid silly things like double inserts into progressively larger undefs.
This still builds 512 bit concats of 128 bits by building up through 256 bits first. But I don't know if that's best.
We probably want to merge this with the vXi1 concat code since they are very similar.
llvm-svn: 327454
BUILD_VECTORs aren't themselves legalized until LegalizeDAG so we should still be able to create an "illegal" one before that. This helps combine with BUILD_VECTORS that are introduced during LegalizeVectorOps due to unrolling.
llvm-svn: 327446
Nothing prevents us from having both frame-setup and frame-destroy on
the same instruction.
When merging:
* frame-setup OPCODE1
* frame-destroy OPCODE2
into
* frame-setup frame-destroy OPCODE3
we want to be able to print and parse both flags.
llvm-svn: 327442
Nops should have zero latency because there is no result.
Idioms like 'xorps xmm0, xmm0' may have zero latency because
they are handled without using an execution unit.
llvm-svn: 327435
Under some circumstances the divrems won't have been combined together before getting to this code.
So replace the assertion with a if() guard to not expand to X-((X/C)*C) to give the other combine chance to happen.
Reduced from OSS-Fuzz #6883https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=6883
llvm-svn: 327424
For the MIPS O32 ABI, the current call lowering logic naively lowers each
call, creating the reserved argument area to hold the argument spill areas for
$a0..$a3 and the outgoing parameter area if one is required at each call site.
In the case of a sufficently large byval argument, a call to memcpy is used
to write the start+16..end of the argument into the outgoing parameter area.
This is done within the CALLSEQ_START..CALLSEQ_END of the callee. The CALLSEQ
nodes are responsible for performing the necessary stack adjustments.
Since the O32/N32/N64 MIPS ABIs do not have a red-zone and writing below the
stack pointer and reading the values back is unpredictable, the call to memcpy
cannot be hoisted out of the callee's CALLSEQ nodes.
However, for the O32 ABI requires the reserved argument area for functions
which have parameters. The naive lowering of calls will then create nested
CALLSEQ sequences. For N32 and N64 these nodes are also created, but with
zero stack adjustments as those ABIs do not have a reserved argument area.
This patch addresses the correctness issue by recognizing the special case
of lowering a byval argument that uses memcpy. By recognizing that the
incoming chain already has a CALLSEQ_START node on it when calling memcpy,
the CALLSEQ nodes are not created. For the N32 and N64 ABIs, this is not an
issue, as no stack adjustment has to be performed.
For the O32 ABI, the correctness reasoning is different. In the case of a
sufficently large byval argument, registers a0..a3 are going to be used for
the callee's arguments, mandating the creation of the reserved argument area.
The call to memcpy in the naive case will also create its own reserved
argument area. However, since the reserved argument area consists of undefined
values, both calls can use the same reserved argument area.
Reviewers: abeserminji, atanasyan
Differential Revision: https://reviews.llvm.org/D44296
llvm-svn: 327388
splitMergedValStore will split a store into two if target prefers this, or if
-force-split-store is passed.
This patch adds the missing handling for endianness in this function along
with a test case.
Review: Eli Friedman
https://reviews.llvm.org/D44396
llvm-svn: 327375
The current zero extension elimination was restricted to operands of
comparison. It actually could be extended to more cases.
For example:
int *inc_p (int *p, unsigned a)
{
return p + a;
}
'a' will be promoted to i64 during addition, and the zero extension could
be eliminated as well.
For the elimination optimization, it should be much better to start
recognizing the candidate sequence from the SRL instruction instead of J*
instructions.
This patch makes it an generic zero extension elimination pass instead of
one restricted with comparison.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 327367
There is a mistake in current code that we "break" out the optimization
when the first operand of J*_RR doesn't qualify the elimination. This
caused some elimination opportunities missed, for example the one in the
testcase.
The code should just fall through to handle the second operand.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 327366
The current subregister definition check stops after the MOV_32_64
instruction.
This means we are thinking all the following instruction sequences
are safe to be eliminated:
MOV_32_64 rB, wA
SLL_ri rB, rB, 32
SRL_ri rB, rB, 32
However, this is *not* true. The source subregister wA of MOV_32_64 could
come from a implicit truncation of 64-bit register in which case the high
bits of the 64-bit register is not zeroed, therefore we can't eliminate
above sequence.
For example, for i32_val, we shouldn't do the elimination:
long long bar ();
int foo (int b, int c)
{
unsigned int i32_val = (unsigned int) bar();
if (i32_val < 10)
return b;
else
return c;
}
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 327365
Improve the test accuracy by adding more check directives.
Shifts are expected to be eliminated for zero extension but not for signed
extension.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 327364
This adds two features: "packets", and "nvj".
Enabling "packets" allows the compiler to generate instruction packets,
while disabling it will prevent it and disable all optimizations that
generate them. This feature is enabled by default on all subtargets.
The feature "nvj" allows the compiler to generate new-value jumps and it
implies "packets". It is enabled on all subtargets.
The exception is made for packets with endloop instructions, since they
require a certain minimum number of instructions in the packets to which
they apply. Disabling "packets" will not prevent hardware loops from
being generated.
llvm-svn: 327302
Since the enqueued kernels have internal linkage, their names may be dropped.
In this case, give them unique names __amdgpu_enqueued_kernel or
__amdgpu_enqueued_kernel.n where n is a sequential number starting from 1.
Differential Revision: https://reviews.llvm.org/D44322
llvm-svn: 327291
64-bit MMX vector generation usually ends up lowering into SSE instructions before being spilled/reloaded as a MMX type.
This patch creates a MMX vector from MMX source values, taking the lowest element from each source and constructing broadcasts/build_vectors with direct calls to the MMX PUNPCKL/PSHUFW intrinsics.
We're missing a few consecutive load combines that could be handled in a future patch if that would be useful - my main interest here is just avoiding a lot of the MMX/SSE crossover.
Differential Revision: https://reviews.llvm.org/D43618
llvm-svn: 327247
Same as the VPERMILPS/VPERMILPD approach for v8f32/v4f64 cases, rely on PSHUFB using bits[3:0] for indexing - we can ignore the sign bit (zero element) as those index vector values are considered undefined. The select between the lo/hi permute results based on the index size.
llvm-svn: 327242
As VPERMILPS/VPERMILPD only selects elements based on the bits[1:0]/bit[1] then we can permute both the (repeated) lo/hi 128-bit vectors in each case and then select between these results based on whether the index was for for lo/hi.
For v4i64/v4f64 this avoids some rather nasty v4i64 multiples on the AVX2 implementation, which seems to be worse than the extra port5 pressure from the additional shuffles/blends.
llvm-svn: 327239
Francis Visoiu Mistrih