This reverts commit r300732. This breaks a few tests.
I think the problem is related to adding more uses of
the condition that don't yet exist at this point.
llvm-svn: 301242
In call sequence setups, there may not be a frame index base
and the pointer is a constant offset from the frame
pointer / scratch wave offset register.
llvm-svn: 301230
Merges equivalent initializations of M0 and hoists them into a common
dominator block. Technically the same code can be used with any
register, physical or virtual.
Differential Revision: https://reviews.llvm.org/D32279
llvm-svn: 301228
Currently the operand type for ATOMIC_FENCE assumes value type of a pointer in address space 0.
This is fine for most targets. However for amdgcn target, the size of pointer in address space 0
depends on triple environment. For amdgiz environment, it is 64 bit but for other environment it is
32 bit. On the other hand, amdgcn target expects 32 bit fence operands independent of the target
triple environment. Therefore a hook is need in target lowering for getting the fence operand type.
This patch has no effect on targets other than amdgcn.
Differential Revision: https://reviews.llvm.org/D32186
llvm-svn: 301215
Fixes traps in any block besides the entry block,
and fixes depending on a live-in physical register
by using a virtual register copy.
Also happens to stop emitting a nop in the case
debug trap is not supported.
llvm-svn: 301206
Summary:
Fix a compiler bug when the lane select happens to end up in a VGPR.
Clarify the semantic of the corresponding intrinsic to be that of
the corresponding GLSL: the lane select must be uniform across a
wave front, otherwise results are undefined.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D32343
llvm-svn: 301197
SI_MASKED_UNREACHABLE does not have machine instruction encoding.
It needs special handling in AMDGPUAsmPrinter::EmitInstruction like some
other pseudo instructions.
This patch fixes compilation failure of RadeonRays.
Differential Revision: https://reviews.llvm.org/D32364
llvm-svn: 301025
Recently alloca address space has been added to data layout. Due to this
change, pointer returned by alloca may have different size as pointer in
address space 0.
However, currently the value type of frame index is assumed to be of the
same size as pointer in address space 0.
This patch fixes that.
Most targets assume alloca returning pointer in address space 0, which
is the default alloca address space. Therefore it is NFC for them.
AMDGCN target with amdgiz environment requires this change since it
assumes alloca returning pointer to addr space 5 and its size is 32,
which is different from the size of pointer in addr space 0 which is 64.
Differential Revision: https://reviews.llvm.org/D32021
llvm-svn: 300864
This is inserted directly in the text section. The relocation
for the function ends up resolving to the beginning of the
amd_kernel_code_t header rather than the actual function
entry point.
Also skip some of the comments for initialization
that only makes sense for kernels.
llvm-svn: 300736
The most common case for a branch condition is
a single use compare. Directly invert the branch
predicate rather than adding a lot of xor i1 true
which the DAG will have to fold later.
This produces nicer to read structurizer output.
This produces some random changes in codegen
due to the DAG swapping branch conditions itself,
and then does a poor job of dealing with those
inverts.
llvm-svn: 300732
Avoid looping through program to determine register counts.
This avoids needing to look at regmask operands.
Also fixes some counting errors with flat_scr when there
are no stack objects.
llvm-svn: 300482
If a kernel's pointer argument is known to be readonly
set access qualifier accordingly. This allows RT not to
flush caches before dispatches.
Differential Revision: https://reviews.llvm.org/D32091
llvm-svn: 300362
In many cases ds operations can be combined even if offsets do not
fit into 8 bit encoding. What it takes is to adjust base address.
Differential Revision: https://reviews.llvm.org/D31993
llvm-svn: 300227
If workgroup size is known inform llvm about range returned by local
id and local size queries.
Differential Revision: https://reviews.llvm.org/D31804
llvm-svn: 300102
Increase threshold to unroll a loop which contains an "if" statement
whose condition defined by a PHI belonging to the loop. This may help
to eliminate if region and potentially even PHI itself, saving on
both divergence and registers used for the PHI.
Add a small bonus for each of such "if" statements.
Differential Revision: https://reviews.llvm.org/D31693
llvm-svn: 299779
Summary:
Difference beetween PreRegAlloc() and MachineSSAOptimization() are that the former is run despite of -O0 optimization level. In my undestanding SiShrinkInstructions and SDWAPeephole shouldn't run when optimizations are disabled.
With this change order of passes will not change.
Reviewers: arsenm, vpykhtin, rampitec
Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye
Differential Revision: https://reviews.llvm.org/D31705
llvm-svn: 299757
The new codepath has been in the tree for years, and there isn't any
reason to use two codepaths here.
Differential Revision: https://reviews.llvm.org/D30596
llvm-svn: 299723
This is possible in ways that are not compiler bugs,
so stop asserting on them.
This emits an extra error when emitting objects when it
can't encode the new pseudo, but I'm not sure that matters.
llvm-svn: 299712
If a workgroup size is known to be not greater than wavefront size
the s_barrier instruction is not needed since all threads are guarantied
to come to the same point at the same time.
Differential Revision: https://reviews.llvm.org/D31731
llvm-svn: 299659
Add a new node to act as a fancy bitcast from f16 operations to
i32 that implicitly zero the high 16-bits of the result.
Alternatively could try making v2f16 legal and canonicalizing
on build_vectors.
llvm-svn: 299246
Previously compiler often extracted common immediates into specific register, e.g.:
```
%vreg0 = S_MOV_B32 0xff;
%vreg2 = V_AND_B32_e32 %vreg0, %vreg1
%vreg4 = V_AND_B32_e32 %vreg0, %vreg3
```
Because of this SDWA peephole failed to find SDWA convertible pattern. E.g. in previous example this could be converted into 2 SDWA src operands:
```
SDWA src: %vreg2 src_sel:BYTE_0
SDWA src: %vreg4 src_sel:BYTE_0
```
With this change peephole check if operand is either immediate or register that is copy of immediate.
llvm-svn: 299202
If set to false it does not remove global aliases. With this parameter
set to false it should be safe to run the pass before link.
Differential Revision: https://reviews.llvm.org/D31489
llvm-svn: 299108
This is less important than increase threshold for private memory,
but still brings performance improvements in a wide range of tests.
Unrolling more for local memory serves three purposes: it allows
to combine ds operations if offset becomes static, saves registers
used for offsets in case of static offsets, and allows better lds
latency hiding.
Differential Revision: https://reviews.llvm.org/D31412
llvm-svn: 298948
Previously it was covered by the internalization. It turns out we cannot
run internalizer in FE, it break separate compilation tests. Thus early
inliner gets its own option.
Differential Revision: https://reviews.llvm.org/D31429
llvm-svn: 298935
Switch data layout by target triple environment amdgiz and amdgizcl indicating using of an address space mapping in which generic address space is 0.
amdgiz is for non-OpenCL environment where generic address space is 0.
amdgizcl is for OpenCL environment where generic address space is 0.
Differential Revision: https://reviews.llvm.org/D31211
llvm-svn: 298758
If the branch condition for a loop was a phi which itself
was fed from a phi from a loop, it isn't safe to try
to delete the phi until after the loop is handled.
llvm-svn: 298737
StructurizeCFG can't handle cases with multiple
returns creating regions with multiple exits.
Create a copy of UnifyFunctionExitNodes that only
unifies exit nodes that skips exit nodes
with uniform branch sources.
llvm-svn: 298729
- Rename runtime metadata -> code object metadata
- Make metadata not flow
- Switch enums to use ScalarEnumerationTraits
- Cleanup and move AMDGPUCodeObjectMetadata.h to AMDGPU/MCTargetDesc
- Introduce in-memory representation for attributes
- Code object metadata streamer
- Create metadata for isa and printf during EmitStartOfAsmFile
- Create metadata for kernel during EmitFunctionBodyStart
- Finalize and emit metadata to .note during EmitEndOfAsmFile
- Other minor improvements/bug fixes
Differential Revision: https://reviews.llvm.org/D29948
llvm-svn: 298552
This is used for a specific type of return to a shader part's
epilog code. Rename to try avoiding confusion from a true
call's return.
llvm-svn: 298452
Fix two problems related to r298025:
- SplitKit would create duplicate VNIs in some cases leading to crashs
when hoisting copies.
- VirtRegMap could fail expanding copies at the beginning of a basic
block.
This fixes http://llvm.org/PR32353
llvm-svn: 298448
Currently the default C calling convention functions are treated
the same as compute kernels. Make this explicit so the default
calling convention can be changed to a non-kernel.
Converted with perl -pi -e 's/define void/define amdgpu_kernel void/'
on the relevant test directories (and undoing in one place that actually
wanted a non-kernel).
llvm-svn: 298444
This adds a parameter to @llvm.objectsize that makes it return
conservative values if it's given null.
This fixes PR23277.
Differential Revision: https://reviews.llvm.org/D28494
llvm-svn: 298430
Summary:
First iteration of SDWA peephole.
This pass tries to combine several instruction into one SDWA instruction. E.g. it converts:
'''
V_LSHRREV_B32_e32 %vreg0, 16, %vreg1
V_ADD_I32_e32 %vreg2, %vreg0, %vreg3
V_LSHLREV_B32_e32 %vreg4, 16, %vreg2
'''
Into:
'''
V_ADD_I32_sdwa %vreg4, %vreg1, %vreg3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
'''
Pass structure:
1. Iterate over machine instruction in basic block and try to apply "SDWA patterns" to each of them. SDWA patterns match machine instruction into either source or destination SDWA operand. E.g. ''' V_LSHRREV_B32_e32 %vreg0, 16, %vreg1''' is matched to source SDWA operand '''%vreg1 src_sel:WORD_1'''.
2. Iterate over found SDWA operands and find instruction that could be potentially coverted into SDWA. E.g. for source SDWA operand potential instruction are all instruction in this basic block that uses '''%vreg0'''
3. Iterate over all potential instructions and check if they can be converted into SDWA.
4. Convert instructions to SDWA.
This review contains basic implementation of SDWA peephole pass. This pass requires additional testing fot both correctness and performance (no performance testing done).
There are several ways this pass can be improved:
1. Make this pass work on whole function not only basic block. As I can see this can be done right now without changes to pass.
2. Introduce more SDWA patterns
3. Introduce mnemonics to limit when SDWA patterns should apply
Reviewers: vpykhtin, alex-t, arsenm, rampitec
Subscribers: wdng, nhaehnle, mgorny
Differential Revision: https://reviews.llvm.org/D30038
llvm-svn: 298365
Folding instructions when selecting can cause them to become dead.
Don't select these dead instructions (if they don't have other side
effects, and don't define physical registers).
Preserve existing tests by adding COPYs.
In some tests, the G_CONSTANT vregs never get constrained to a class:
the only use of the vreg was folded into another instruction, so the
G_CONSTANT, now dead, never gets selected.
llvm-svn: 298224
This is direct port of HSAILAliasAnalysis pass, just cleaned for
style and renamed.
Differential Revision: https://reviews.llvm.org/D31103
llvm-svn: 298172
If the loop condition was an i1 phi with a constantexpr input, this
would add a loop intrinsic fed by a phi dependent on a call to
if.break in the same block. Insert the call in the loop header.
llvm-svn: 298121
- This fixes a bug where subregister incompatible with the vregs register
class where used.
- Implement the case where multiple copies are necessary to cover a
given lanemask.
Differential Revision: https://reviews.llvm.org/D30438
llvm-svn: 298025
We can mark functions to always inline early in the opt. Since we do not have
call support this early inlining creates opportunities for inter-procedural
optimizations which would not occur otherwise.
Differential Revision: https://reviews.llvm.org/D31016
llvm-svn: 297958
computeKnownBits didn't handle fp_to_fp16 to report
the high bits as 0. ARM maps the generic node to an instruction
that does not modify the high bits of the register, so introduce
a target node where the high bits are known 0.
llvm-svn: 297873
Recommiting with compiler time improvements
Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner.
* Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search and chain alias analysis which only
checks for parallel stores through the chain subgraph. This is cleaner
as the separation of non-interfering loads/stores from the
store-merging logic.
When merging stores search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited.
This improves the quality of the output SelectionDAG and the output
Codegen (save perhaps for some ARM cases where we correctly constructs
wider loads, but then promotes them to float operations which appear
but requires more expensive constant generation).
Some minor peephole optimizations to deal with improved SubDAG shapes (listed below)
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the chain aggregation in the merged stores across code
paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seems sufficient to not cause regressions in
tests.
5. Remove Chain dependencies of Memory operations on CopyfromReg
nodes as these are captured by data dependence
6. Forward loads-store values through tokenfactors containing
{CopyToReg,CopyFromReg} Values.
7. Peephole to convert buildvector of extract_vector_elt to
extract_subvector if possible (see
CodeGen/AArch64/store-merge.ll)
8. Store merging for the ARM target is restricted to 32-bit as
some in some contexts invalid 64-bit operations are being
generated. This can be removed once appropriate checks are
added.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable, improving load-store forwarding. One test in
particular is worth noting:
CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store
forwarding converts a load-store pair into a parallel store and
a memory-realized bitcast of the same value. However, because we
lose the sharing of the explicit and implicit store values we
must create another local store. A similar transformation
happens before SelectionDAG as well.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
llvm-svn: 297695
Since v_max_f32_e64/v_max_f16_e64 can be folded if the target
instruction supports the clamp bit, we also need to maintain
modifiers when converting v_mac to v_mad.
This fixes a rendering issue with Dirt Rally because a v_mac
instruction with the clamp bit set was converted to a v_mad
but that bit was lost during the conversion.
Fixes: e184e01dd79 ("AMDGPU: Fold FP clamp as modifier bit")
Patch by Samuel Pitoiset <samuel.pitoiset@gmail.com>
llvm-svn: 297556
This method inverts the Reason field of a scheduling candidate.
It does right comparison between RegCritical and RegExcess, but
everything else is broken. In fact it can prefer less strong reason
such as Weak over RegCritical because Weak > -RegCritical.
The CandReason enum is properly sorted, so just remove artificial
ranking.
Differential Revision: https://reviews.llvm.org/D30557
llvm-svn: 297536
If there is only one successor, and that successor only
has one predecessor the wait can obviously be delayed until
uses or the end of the next block. This avoids code quality
regressions when there are trivial fallthrough blocks inserted
for structurization.
llvm-svn: 297251
This patch causes compile times for some patterns to explode. I have
a (large, unreduced) test case that slows down by more than 20x and
several test cases slow down by 2x. I'm sending some of the test cases
directly to Nirav and following up with more details in the review log,
but this should unblock anyone else hitting this.
llvm-svn: 296862
This commit also relied on r296812, which I just reverted. We should probably
apply it again, after the r296812 has been discussed and been reapplied in some
variant.
llvm-svn: 296820
Surprisingly, one of the three interference checks in LiveRegMatrix was
using the main live range instead of the apropriate subregister range
resulting in unnecessarily conservative results.
llvm-svn: 296722
Resubmit r295336 after the bug with non-zero offset patterns on BE targets is fixed (r296336).
Support {a|s}ext, {a|z|s}ext load nodes as a part of load combine patters.
Reviewed By: filcab
Differential Revision: https://reviews.llvm.org/D29591
llvm-svn: 296651
Modify the test so that it is still testing something
closer to what it was intended to originally.
I think the original intent was to test the situation where
there was a branch on execz and then unconditional branch
required relaxing.With the change in r296539,
there was no longer and execz branch.
Change the test so that there is now an execz branch inserted.
There is no longer an unconditional branch after the execz branch,
so this might need to be tricked in some other way to keep that
there.
llvm-svn: 296574
Recommiting after fixup of 32-bit aliasing sign offset bug in DAGCombiner.
* Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search and chain alias analysis which only
checks for parallel stores through the chain subgraph. This is cleaner
as the separation of non-interfering loads/stores from the
store-merging logic.
When merging stores search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited.
This improves the quality of the output SelectionDAG and the output
Codegen (save perhaps for some ARM cases where we correctly constructs
wider loads, but then promotes them to float operations which appear
but requires more expensive constant generation).
Some minor peephole optimizations to deal with improved SubDAG shapes (listed below)
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the chain aggregation in the merged stores across code
paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seems sufficient to not cause regressions in
tests.
5. Remove Chain dependencies of Memory operations on CopyfromReg
nodes as these are captured by data dependence
6. Forward loads-store values through tokenfactors containing
{CopyToReg,CopyFromReg} Values.
7. Peephole to convert buildvector of extract_vector_elt to
extract_subvector if possible (see
CodeGen/AArch64/store-merge.ll)
8. Store merging for the ARM target is restricted to 32-bit as
some in some contexts invalid 64-bit operations are being
generated. This can be removed once appropriate checks are
added.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable, improving load-store forwarding. One test in
particular is worth noting:
CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store
forwarding converts a load-store pair into a parallel store and
a memory-realized bitcast of the same value. However, because we
lose the sharing of the explicit and implicit store values we
must create another local store. A similar transformation
happens before SelectionDAG as well.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
llvm-svn: 296476
Recommiting after fixup of 32-bit aliasing sign offset bug in DAGCombiner.
* Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search and chain alias analysis which only
checks for parallel stores through the chain subgraph. This is cleaner
as the separation of non-interfering loads/stores from the
store-merging logic.
When merging stores search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited.
This improves the quality of the output SelectionDAG and the output
Codegen (save perhaps for some ARM cases where we correctly constructs
wider loads, but then promotes them to float operations which appear
but requires more expensive constant generation).
Some minor peephole optimizations to deal with improved SubDAG shapes (listed below)
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the chain aggregation in the merged stores across code
paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seems sufficient to not cause regressions in
tests.
5. Remove Chain dependencies of Memory operations on CopyfromReg
nodes as these are captured by data dependence
6. Forward loads-store values through tokenfactors containing
{CopyToReg,CopyFromReg} Values.
7. Peephole to convert buildvector of extract_vector_elt to
extract_subvector if possible (see
CodeGen/AArch64/store-merge.ll)
8. Store merging for the ARM target is restricted to 32-bit as
some in some contexts invalid 64-bit operations are being
generated. This can be removed once appropriate checks are
added.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable, improving load-store forwarding. One test in
particular is worth noting:
CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store
forwarding converts a load-store pair into a parallel store and
a memory-realized bitcast of the same value. However, because we
lose the sharing of the explicit and implicit store values we
must create another local store. A similar transformation
happens before SelectionDAG as well.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
llvm-svn: 296252
This reverts commit r296009. It broke one out of tree target and also
does not account for all partial lines added or removed when calculating
PressureDiff.
llvm-svn: 296182
The motivation for filling out these select-of-constants cases goes back to D24480,
where we discussed removing an IR fold from add(zext) --> select. And that goes back to:
https://reviews.llvm.org/rL75531https://reviews.llvm.org/rL159230
The idea is that we should always canonicalize patterns like this to a select-of-constants
in IR because that's the smallest IR and the best for value tracking. Note that we currently
do the opposite in some cases (like the cases in *this* patch). Ie, the proposed folds in
this patch already exist in InstCombine today:
https://github.com/llvm-mirror/llvm/blob/master/lib/Transforms/InstCombine/InstCombineSelect.cpp#L1151
As this patch shows, most targets generate better machine code for simple ext/add/not ops
rather than a select of constants. So the follow-up steps to make this less of a patchwork
of special-case folds and missing IR canonicalization:
1. Have DAGCombiner convert any select of constants into ext/add/not ops.
2 Have InstCombine canonicalize in the other direction (create more selects).
Differential Revision: https://reviews.llvm.org/D30180
llvm-svn: 296137
If a subreg is used in an instruction it counts as a whole superreg
for the purpose of register pressure calculation. This patch corrects
improper register pressure calculation by examining operand's lane mask.
Differential Revision: https://reviews.llvm.org/D29835
llvm-svn: 296009
The manual is unclear on the details of this. It's not
clear to me if denormals are not allowed with clamp,
or if that is only omod. Not allowing denorms for
fp16 or fp64 isn't useful so I also question if that
is really a restriction. Same with whether this is valid
without IEEE mode enabled.
llvm-svn: 295905
This should avoid reporting any stack needs to be allocated in the
case where no stack is truly used. An unused stack slot is still
left around in other cases where there are real stack objects
but no spilling occurs.
llvm-svn: 295891
This allows us to ensure that 0 is never a valid pointer
to a user object, and ensures that the offset is always legal
without needing a register to access it. This comes at the cost
of usable offsets and wasted stack space.
llvm-svn: 295877
Change implementation to use max instead of add.
min/max/med3 do not flush denormals regardless of the mode,
so it is OK to use it whether or not they are enabled.
Also allow using clamp with f16, and use knowledge
of dx10_clamp.
llvm-svn: 295788
Before frame offsets are calculated, try to eliminate the
frame indexes used by SGPR spills. Then we can delete them
after.
I think for now we can be sure that no other instruction
will be re-using the same frame indexes. It should be easy
to notice if this assumption ever breaks since everything
asserts if it tries to use a dead frame index later.
The unused emergency stack slot seems to still be left behind,
so an additional 4 bytes is still wasted.
llvm-svn: 295753
A line number doesn't make much sense if you don't say where it's
from. Add a verifier check for this and update some tests that had
bogus debug info.
llvm-svn: 295516
Resubmit -r295314 with PowerPC and AMDGPU tests updated.
Support {a|s}ext, {a|z|s}ext load nodes as a part of load combine patters.
Reviewed By: filcab
Differential Revision: https://reviews.llvm.org/D29591
llvm-svn: 295336
Marek Olsak