This started by noticing that scalar and vector types were producing different results with div ops in PR36305:
https://bugs.llvm.org/show_bug.cgi?id=36305
...but the problem is bigger. I couldn't keep it straight without a table, so I'm attaching that as a PDF to
the review. The x86 tests in undef-ops.ll correspond to that table.
Green means that instsimplify and the DAG agree on the result for all types.
Red means the DAG was returning undef when IR was not.
Yellow means the DAG was returning a non-undef result when IR returned undef.
This patch assumes that we're currently doing the right thing in IR.
Note: I couldn't find any problems with lowering vector constants as the code comments were warning,
but those comments were written long ago in rL36413 .
Differential Revision: https://reviews.llvm.org/D43141
llvm-svn: 324941
We only tagged it with the itinerary class, so completeness checks were erroneously passed (PR35639).
AMD targets can perform these a lot quicker than WriteMicrocoded so will need an override in the models.
llvm-svn: 324897
It asserts building Chromium; see PR36346.
(This also reverts the follow-up r324836.)
> If a load follows a store and reloads data that the store has written to memory, Intel microarchitectures can in many cases forward the data directly from the store to the load, This "store forwarding" saves cycles by enabling the load to directly obtain the data instead of accessing the data from cache or memory.
> A "store forward block" occurs in cases that a store cannot be forwarded to the load. The most typical case of store forward block on Intel Core microarchiticutre that a small store cannot be forwarded to a large load.
> The estimated penalty for a store forward block is ~13 cycles.
>
> This pass tries to recognize and handle cases where "store forward block" is created by the compiler when lowering memcpy calls to a sequence
> of a load and a store.
>
> The pass currently only handles cases where memcpy is lowered to XMM/YMM registers, it tries to break the memcpy into smaller copies.
> breaking the memcpy should be possible since there is no atomicity guarantee for loads and stores to XMM/YMM.
llvm-svn: 324887
Previously we just emitted this as a MOV8rm which would likely get folded during the peephole pass anyway. This just makes it explicit earlier.
The gpr-to-mask.ll test changed because the kaddb instruction has no memory form.
llvm-svn: 324860
Summary:
Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare.
I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior.
Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into.
I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare.
Reviewers: spatel, RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42935
llvm-svn: 324842
This reverses instcombine's demanded bits' transform which always tries to clear bits in constants.
As noted in PR35792 and shown in the test diffs:
https://bugs.llvm.org/show_bug.cgi?id=35792
...we can do better in codegen by trying to form -1. The x86 sub test shows a missed opportunity.
I did investigate changing instcombine's behavior, but it would be more work to change
canonicalization in IR. Clearing bits / shrinking constants can allow killing instructions,
so we'd have to figure out how to not regress those cases.
Differential Revision: https://reviews.llvm.org/D42986
llvm-svn: 324839
This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT.
Differential Revision: https://reviews.llvm.org/D43014
llvm-svn: 324837
If a load follows a store and reloads data that the store has written to memory, Intel microarchitectures can in many cases forward the data directly from the store to the load, This "store forwarding" saves cycles by enabling the load to directly obtain the data instead of accessing the data from cache or memory.
A "store forward block" occurs in cases that a store cannot be forwarded to the load. The most typical case of store forward block on Intel Core microarchiticutre that a small store cannot be forwarded to a large load.
The estimated penalty for a store forward block is ~13 cycles.
This pass tries to recognize and handle cases where "store forward block" is created by the compiler when lowering memcpy calls to a sequence
of a load and a store.
The pass currently only handles cases where memcpy is lowered to XMM/YMM registers, it tries to break the memcpy into smaller copies.
breaking the memcpy should be possible since there is no atomicity guarantee for loads and stores to XMM/YMM.
Change-Id: I620b6dc91583ad9a1444591e3ddc00dd25d81748
llvm-svn: 324835
This patch adds a new function attribute "required-vector-width" that can be set by the frontend to indicate the maximum vector width present in the original source code. The idea is that this would be set based on ABI requirements, intrinsics or explicit vector types being used, maybe simd pragmas, etc. The backend will then use this information to determine if its save to make 512-bit vectors illegal when the preference is for 256-bit vectors.
For code that has no vectors in it originally and only get vectors through the loop and slp vectorizers this allows us to generate code largely similar to our AVX2 only output while still enabling AVX512 features like mask registers and gather/scatter. The loop vectorizer doesn't always obey TTI and will create oversized vectors with the expectation the backend will legalize it. In order to avoid changing the vectorizer and potentially harm our AVX2 codegen this patch tries to make the legalizer behavior similar.
This is restricted to CPUs that support AVX512F and AVX512VL so that we have good fallback options to use 128 and 256-bit vectors and still get masking.
I've qualified every place I could find in X86ISelLowering.cpp and added tests cases for many of them with 2 different values for the attribute to see the codegen differences.
We still need to do frontend work for the attribute and teach the inliner how to merge it, etc. But this gets the codegen layer ready for it.
Differential Revision: https://reviews.llvm.org/D42724
llvm-svn: 324834
Summary:
This patch changes the signature of the avx512 packed fp compare intrinsics to return a vXi1 vector and no longer take a mask as input. The casts to scalar type will now need to be explicit in the IR. The masking node will now be an explicit and in the IR.
This makes the intrinsic look much more similar to an fcmp instruction that we wish we could use for these but can't. We already use icmp instructions for integer compares.
Previously the lowering step of isel would turn the intrinsic into an X86 specific ISD node and a emit the masking nodes as well as some bitcasts. This means DAG combines can't see the vXi1 type until somewhat late, making it more difficult to combine out gpr<->mask transition sequences. By exposing the vXi1 type explicitly in the IR and initial SelectionDAG we give earlier DAG combines and even InstCombine the chance to see it and optimize it.
This should make any issues with gpr<->mask sequences the same between integer and fp. Meaning we only have to fix them once.
Reviewers: spatel, delena, RKSimon, zvi
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D43137
llvm-svn: 324827
Undef VLX, getSetCCResultType returns v2i1/v4i1 for v2f32/v4f32 so default type legalization will end up changing the setcc result type back to vXi1 if it had been extended. The resulting extend gets messed up further by type legalization and is difficult to recombine back to (v4i32 (setcc (v4f32))) after legalization.
I went ahead and enabled this for SSE2 and later since its always the result we want and this helps type legalization get there in less steps.
llvm-svn: 324822
This prevents extends of masks being introduced during lowering where it become difficult to combine them out.
There are a few oddities in here.
We sometimes concatenate two k-registers produced by two compares, sign_extend the combined pair, then extract two halves. This worked better previously because the sign_extend wasn't created until after the fp_to_sint was split which led to a split sign_extend being created.
We probably also need to custom type legalize (v2i32 (sext v2i1)) via widening.
llvm-svn: 324820
The update script sometimes has trouble when there are check-prefixes representing every possible combination of feature flags. I have a patch where the update script was generating something that didn't pass lit.
This patch just removes some check-prefixes and expands out some of the checks to workaround this.
llvm-svn: 324819
This avoids a constant pool load to create 1.
The int->float are showing converts to mask and back. We probably need to widen inputs to sint_to_fp/uint_to_fp before type legalization.
llvm-svn: 324805
* Use uleb128 for code offsets in the LSDA call site table.
* Omit the TTBase offset if the type table is empty.
This change can reduce the size of the DWARF/Itanium LSDA by about half.
Patch by Ryan Prichard!
llvm-svn: 324750
Rely on the assembler to finalize the layout of the DWARF/Itanium
exception-handling LSDA. Rather than calculate the exact size of each
thing in the LSDA, use assembler directives:
To emit the offset to the TTBase label:
.uleb128 .Lttbase0-.Lttbaseref0
.Lttbaseref0:
To emit the size of the call site table:
.uleb128 .Lcst_end0-.Lcst_begin0
.Lcst_begin0:
... call site table entries ...
.Lcst_end0:
To align the type info table:
... action table ...
.balign 4
.long _ZTIi
.long _ZTIl
.Lttbase0:
Using assembler directives simplifies the compiler and allows switching
the encoding of offsets in the call site table from udata4 to uleb128 for
a large code size savings. (This commit does not change the encoding.)
The combination of the uleb128 followed by a balign creates an unfortunate
dependency cycle that the assembler must sometimes resolve either by
padding an LEB or by inserting zero padding before the type table. See
PR35809 or GNU as bug 4029.
Patch by Ryan Prichard!
llvm-svn: 324749
Previously we extracted two subvectors and concatenate. But the concatenate will be lowered to two insert subvectors. Then DAG combine will merge once of the inserts and one of the extracts back into the original vector. We might as well just directly use one extract and one insert.
llvm-svn: 324710
This regresses a couple cases in the shuffle combining test. But those cases use intrinsics that InstCombine knows how to turn into a generic shuffle earlier. This should give opportunities to fold this earlier in InstCombine or DAG combine.
llvm-svn: 324709
The patch essentially makes sure that X86CallLowering adds proper
G_COPY/G_TRUNC and G_ANYEXT/G_COPY when we are doing lowering of
arguments/returns for floating point values passed on registers.
Tests are updated accordingly
Reviewed By: qcolombet
Differential Revision: https://reviews.llvm.org/D42287
llvm-svn: 324665
Most vxi1 constant build vectors have to be implemented in the scalar domain anyway so we'll probably end up with a cast there later. But by then its too late to do the combine to get rid of it.
llvm-svn: 324662
Craig Topper