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
Lay out trellis-shaped CFGs optimally.
A trellis of the shape below:
A B
|\ /|
| \ / |
| X |
| / \ |
|/ \|
C D
would be laid out A; B->C ; D by the current layout algorithm. Now we identify
trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an
increasing number of predecessors. A trellis is a a group of 2 or more
predecessor blocks that all have the same successors.
because of this we can tail duplicate to extend existing trellises.
As an example consider the following CFG:
B D F H
/ \ / \ / \ / \
A---C---E---G---Ret
Where A,C,E,G are all small (Currently 2 instructions).
The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret.
The current code will copy C into B, E into D and G into F and yield the layout
A,C,B(C),E,D(E),F(G),G,H,ret
define void @straight_test(i32 %tag) {
entry:
br label %test1
test1: ; A
%tagbit1 = and i32 %tag, 1
%tagbit1eq0 = icmp eq i32 %tagbit1, 0
br i1 %tagbit1eq0, label %test2, label %optional1
optional1: ; B
call void @a()
br label %test2
test2: ; C
%tagbit2 = and i32 %tag, 2
%tagbit2eq0 = icmp eq i32 %tagbit2, 0
br i1 %tagbit2eq0, label %test3, label %optional2
optional2: ; D
call void @b()
br label %test3
test3: ; E
%tagbit3 = and i32 %tag, 4
%tagbit3eq0 = icmp eq i32 %tagbit3, 0
br i1 %tagbit3eq0, label %test4, label %optional3
optional3: ; F
call void @c()
br label %test4
test4: ; G
%tagbit4 = and i32 %tag, 8
%tagbit4eq0 = icmp eq i32 %tagbit4, 0
br i1 %tagbit4eq0, label %exit, label %optional4
optional4: ; H
call void @d()
br label %exit
exit:
ret void
}
here is the layout after D27742:
straight_test: # @straight_test
; ... Prologue elided
; BB#0: # %entry ; A (merged with test1)
; ... More prologue elided
mr 30, 3
andi. 3, 30, 1
bc 12, 1, .LBB0_2
; BB#1: # %test2 ; C
rlwinm. 3, 30, 0, 30, 30
beq 0, .LBB0_3
b .LBB0_4
.LBB0_2: # %optional1 ; B (copy of C)
bl a
nop
rlwinm. 3, 30, 0, 30, 30
bne 0, .LBB0_4
.LBB0_3: # %test3 ; E
rlwinm. 3, 30, 0, 29, 29
beq 0, .LBB0_5
b .LBB0_6
.LBB0_4: # %optional2 ; D (copy of E)
bl b
nop
rlwinm. 3, 30, 0, 29, 29
bne 0, .LBB0_6
.LBB0_5: # %test4 ; G
rlwinm. 3, 30, 0, 28, 28
beq 0, .LBB0_8
b .LBB0_7
.LBB0_6: # %optional3 ; F (copy of G)
bl c
nop
rlwinm. 3, 30, 0, 28, 28
beq 0, .LBB0_8
.LBB0_7: # %optional4 ; H
bl d
nop
.LBB0_8: # %exit ; Ret
ld 30, 96(1) # 8-byte Folded Reload
addi 1, 1, 112
ld 0, 16(1)
mtlr 0
blr
The tail-duplication has produced some benefit, but it has also produced a
trellis which is not laid out optimally. With this patch, we improve the layouts
of such trellises, and decrease the cost calculation for tail-duplication
accordingly.
This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have
back edges, which is a negative, but it has a bigger compensating
positive, which is that it handles the case where there are long strings
of skipped blocks much better than the original layout. Both layouts
handle runs of executed blocks equally well. Branch prediction also
improves if there is any correlation between subsequent optional blocks.
Here is the resulting concrete layout:
straight_test: # @straight_test
; BB#0: # %entry ; A (merged with test1)
mr 30, 3
andi. 3, 30, 1
bc 12, 1, .LBB0_4
; BB#1: # %test2 ; C
rlwinm. 3, 30, 0, 30, 30
bne 0, .LBB0_5
.LBB0_2: # %test3 ; E
rlwinm. 3, 30, 0, 29, 29
bne 0, .LBB0_6
.LBB0_3: # %test4 ; G
rlwinm. 3, 30, 0, 28, 28
bne 0, .LBB0_7
b .LBB0_8
.LBB0_4: # %optional1 ; B (Copy of C)
bl a
nop
rlwinm. 3, 30, 0, 30, 30
beq 0, .LBB0_2
.LBB0_5: # %optional2 ; D (Copy of E)
bl b
nop
rlwinm. 3, 30, 0, 29, 29
beq 0, .LBB0_3
.LBB0_6: # %optional3 ; F (Copy of G)
bl c
nop
rlwinm. 3, 30, 0, 28, 28
beq 0, .LBB0_8
.LBB0_7: # %optional4 ; H
bl d
nop
.LBB0_8: # %exit
Differential Revision: https://reviews.llvm.org/D28522
llvm-svn: 295223
This patch reverts region's scheduling to the original untouched state
in case if we have have decreased occupancy.
In addition it switches to use TargetRegisterInfo occupancy callback
for pressure limits instead of gradually increasing limits which were
just passed by. We are going to stay with the best schedule so we do
not need to tolerate worsened scheduling anymore.
Differential Revision: https://reviews.llvm.org/D29971
llvm-svn: 295206
This patch corrects the maximum workgroups per CU if we have big
workgroups (more than 128). This calculation contributes to the
occupancy calculation in respect to LDS size.
Differential Revision: https://reviews.llvm.org/D29974
llvm-svn: 295134
This change returns empty PSet list for M0 register. Otherwise its
PSet as defined by tablegen is SReg_32. This results in incorrect
register pressure calculation every time an instruction uses M0.
Such uses count as SReg_32 PSet and inadequately increase pressure
on SGPRs.
Differential Revision: https://reviews.llvm.org/D29798
llvm-svn: 294691
For amdgcn target Clang generates addrspacecast to represent null pointers in private and local address spaces.
In LLVM codegen, the static variable initializer is lowered by virtual function AsmPrinter::lowerConstant which is target generic. Since addrspacecast is target specific, AsmPrinter::lowerConst
This patch overrides AsmPrinter::lowerConstant with AMDGPUAsmPrinter::lowerConstant, which is able to lower the target-specific addrspacecast in the null pointer representation so that -1 is co
Differential Revision: https://reviews.llvm.org/D29284
llvm-svn: 294265
An assert occurs when calling SlotIndexes::getInstructionIndex with
a DBG_VALUE instruction because the function expects an instruction
with a slot index. However, there is no slot index for a DBG_VALUE
instruction.
Differential Revision: https://reviews.llvm.org/D29048
llvm-svn: 294070
Recommiting after fixing X86 inc/dec chain bug.
* 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: 293893
The operand types were defined to fit the fp16_to_fp node, which
has the half as an integer type. v_cvt_f32_f16 does support
source modifiers, so change this to have an FP type and modifiers.
For targets without legal f16, this requires recognizing the
bit operations and trying to produce them.
llvm-svn: 293857
Functions matching LDS use to occupancy return results for a workgroup
of 64 workitems. The numbers has to be adjusted for bigger workgroups.
For example a workgroup of size 256 already occupies 4 waves just by
itself. Given that all numbers of LDS use in the compiler are per
workgroup, occupancy shall be multiplied by 4 in this case. Each 64
workitems still limited by the same number, but 4 subrgoups 64 workitems
each can afford 4 times more LDS to get the same occupancy.
In addition change initializes LDS size in the subtarget to a real value
for SI+ targets. This is required since LDS size is a variable in these
calculations.
Differential Revision: https://reviews.llvm.org/D29423
llvm-svn: 293837
These were simply preserving the flags of the original operation,
which was too conservative in most cases and incorrect for mul.
nsw/nuw may be needed for some combines to cleanup messes when
intermediate sext_inregs are introduced later.
Tested valid combinations with alive.
llvm-svn: 293776
When choosing the best successor for a block, ordinarily we would have preferred
a block that preserves the CFG unless there is a strong probability the other
direction. For small blocks that can be duplicated we now skip that requirement
as well, subject to some simple frequency calculations.
Differential Revision: https://reviews.llvm.org/D28583
llvm-svn: 293716
Summary:
For some reason instructions are being inserted in the wrong order with some
builds. I'm not sure why this is happening.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, tony-tye, tpr, llvm-commits
Differential Revision: https://reviews.llvm.org/D29325
llvm-svn: 293639
Summary:
The affected transforms all implicitly use associativity of addition,
for which we usually require unsafe math to be enabled.
The "Aggressive" flag is only meant to convey information about the
performance of the fused ops relative to a fmul+fadd sequence.
Fixes Bug 31626.
Reviewers: spatel, hfinkel, mehdi_amini, arsenm, tstellarAMD
Subscribers: jholewinski, nemanjai, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D28675
llvm-svn: 293635
I think this is safe as long as no inputs are known to ever
be nans.
Also add an intrinsic for fmed3 to be able to handle all safe
math cases.
llvm-svn: 293598
Since we have no call support and late linking we can produce code
only for used symbols. This saves compilation time, size of the final
executable, and size of any intermediate dumps.
Run Internalize pass early in the opt pipeline followed by global
DCE pass. To enable it RT can pass -amdgpu-internalize-symbols option.
Differential Revision: https://reviews.llvm.org/D29214
llvm-svn: 293549
This is worse if the original constant is an inline immediate.
This should also be done for 64-bit adds, but requires fixing
operand folding bugs first.
llvm-svn: 293540
Accomplishes what r292982 was supposed to, which ended up
only really making the necessary test changes.
This should be applied to the 4.0 branch.
Patch by Vedran Miletić <vedran@miletic.net>
llvm-svn: 293310
With the adjustPassManager interface that is now possible to use
custom early module passes.
Differential Revision: https://reviews.llvm.org/D29189
llvm-svn: 293300
* 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: 293184
clang already emits this with -cl-no-signed-zeros, but codegen
doesn't do anything with it. Treat it like the other fast math
attributes, and change one place to use it.
llvm-svn: 293024
Leave early ifcvt disabled for now since there are some
shader-db regressions.
This causes some immediate improvements, but could be better.
The cost checking that the pass does is based on critical path
length for out of order CPUs which we do not want so it skips out
on many cases we want.
llvm-svn: 293016
The sequence like this:
v_cmpx_le_f32_e32 vcc, 0, v0
s_branch BB0_30
s_cbranch_execnz BB0_30
; BB#29:
exp null off, off, off, off done vm
s_endpgm
BB0_30:
; %endif110
is likely wrong. The s_branch instruction will unconditionally jump
to BB0_30 and the skip block (exp done + endpgm) inserted for
performing the kill instruction will never be executed. This results
in a GPU hang with Star Ruler 2.
The s_branch instruction is added during the "Control Flow Optimizer"
pass which seems to re-organize the basic blocks, and we assume
that SI_KILL_TERMINATOR is always the last instruction inside a
basic block. Thus, after inserting a skip block we just go to the
next BB without looking at the subsequent instructions after the
kill, and the s_branch op is never removed.
Instead, we should remove the unconditional out branches and let
skip the two instructions if the exec mask is non-zero.
This patch fixes the GPU hang and doesn't introduce any regressions
with "make check".
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=99019
Patch by Samuel Pitoiset <samuel.pitoiset@gmail.com>
llvm-svn: 292985
This switches to the workaround that HSA defaults to
for the mesa path.
This should be applied to the 4.0 branch.
Patch by Vedran Miletić <vedran@miletic.net>
llvm-svn: 292982
Regalloc creates COPY instructions which do not formally use VALU.
That results in v_mov instructions displaced after exec mask modification.
One pass which do it is SIOptimizeExecMasking, but potentially it can be
done by other passes too.
This patch adds a pass immediately after regalloc to add implicit exec
use operand to all VGPR copy instructions.
Differential Revision: https://reviews.llvm.org/D28874
llvm-svn: 292956
Inline spiller can decide to move a spill as early as possible in the basic block.
It will skip phis and label, but we also need to make sure it skips instructions
in the basic block prologue which restore exec mask.
Added isPositionLike callback in TargetInstrInfo to detect instructions which
shall be skipped in addition to common phis, labels etc.
Differential Revision: https://reviews.llvm.org/D27997
llvm-svn: 292554
For -(x + y) -> (-x) + (-y), if x == -y, this would
change the result from -0.0 to 0.0. Since the fma/fmad
combine is an extension of this problem it also
applies there.
fmul should be fine, and I don't think any of the unary
operators or conversions should be a problem either.
llvm-svn: 292473
Limit register coalescer by not allowing it to artificially increase
size of registers beyond dword. Such super-registers are in fact
register sequences and not distinct HW registers.
With more super-regs we would need to allocate adjacent registers
and constraint regalloc more than needed. Moreover, our super
registers are overlapping. For instance we have VGPR0_VGPR1_VGPR2,
VGPR1_VGPR2_VGPR3, VGPR2_VGPR3_VGPR4 etc, which complicates registers
allocation even more, resulting in excessive spilling.
Differential Revision: https://reviews.llvm.org/D28782
llvm-svn: 292413
This produces worse code when i16 is legal, mostly
due to combines getting confused by conversions inserted
for uniform 16-bit operations.
llvm-svn: 291717
This reverts commit ada6595a526d71df04988eb0a4b4fe84df398ded.
This needs a simple probability check because there are some cases where it is
not profitable.
llvm-svn: 291695
Even with aggressive fusion enabled, this requires duplicating
the fmul, or increases an fadd to another fma which is not an
improvement.
llvm-svn: 291642
When choosing the best successor for a block, ordinarily we would have preferred
a block that preserves the CFG unless there is a strong probability the other
direction. For small blocks that can be duplicated we now skip that requirement
as well.
Differential revision: https://reviews.llvm.org/D27742
llvm-svn: 291609
If a vector index is out of bounds, the result is supposed to be
undefined but is not undefined behavior. Change the legalization
for indexing the vector on the stack so that an out of bounds
index does not create an out of bounds memory access.
llvm-svn: 291604
For i16 zeroext arguments when i16 was a legal type, the
known bits information from the truncate was lost. Insert
a zeroext so the known bits optimizations work with the 32-bit
loads.
Fixes code quality regressions vs. SI in min.ll test.
llvm-svn: 291461
Summary:
Originally
i64 = umax t8, Constant:i64<4>
was expanded into
i32,i32 = umax Constant:i32<0>, Constant:i32<0>
i32,i32 = umax t7, Constant:i32<4>
Now instead the two produced umax:es return i32 instead of i32, i32.
Thanks to Jan Vesely for help with the test case.
Patch by mikael.holmen at ericsson.com
Reviewers: bogner, jvesely, tstellarAMD, arsenm
Subscribers: test, wdng, RKSimon, arsenm, nhaehnle, llvm-commits
Differential Revision: https://reviews.llvm.org/D28135
llvm-svn: 291441
Canonicalize a select with a constant to the false side. This
enables more instruction shrinking opportunities since an
inline immediate can be used for the false side of v_cndmask_b32_e32.
This seems to usually be better but causes some code size regressions
in some tests.
llvm-svn: 290372
When the instruction is processed the first time, it may be
deleted resulting in crashes. While the new test adds the same
user to the worklist twice, this particular case doesn't crash
but I'm not sure why.
llvm-svn: 290191
Summary:
Without a MachineMemOperand, the scheduler was assuming MIMG instructions
were ordered memory references, so no loads or stores could be reordered
across them.
Reviewers: arsenm
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, tony-tye
Differential Revision: https://reviews.llvm.org/D27536
llvm-svn: 290179
Retrying after fixing after removing load-store factoring through
token factors in favor of improved token factor operand pruning
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search 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.
Whem 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 generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the 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 seemed sufficient to not cause regressions in
tests.
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. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and
merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, dsanders, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 289659
Summary:
This change adds some verification in the IR verifier around struct path
TBAA metadata.
Other than some basic sanity checks (e.g. we get constant integers where
we expect constant integers), this checks:
- That by the time an struct access tuple `(base-type, offset)` is
"reduced" to a scalar base type, the offset is `0`. For instance, in
C++ you can't start from, say `("struct-a", 16)`, and end up with
`("int", 4)` -- by the time the base type is `"int"`, the offset
better be zero. In particular, a variant of this invariant is needed
for `llvm::getMostGenericTBAA` to be correct.
- That there are no cycles in a struct path.
- That struct type nodes have their offsets listed in an ascending
order.
- That when generating the struct access path, you eventually reach the
access type listed in the tbaa tag node.
Reviewers: dexonsmith, chandlerc, reames, mehdi_amini, manmanren
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D26438
llvm-svn: 289402
Since 32-bit instructions with 32-bit input immediate behavior
are used to materialize 16-bit constants in 32-bit registers
for 16-bit instructions, determining the legality based
on the size is incorrect. Change operands to have the size
specified in the type.
Also adds a workaround for a disassembler bug that
produces an immediate MCOperand for an operand that
is supposed to be OPERAND_REGISTER.
The assembler appears to accept out of bounds immediates and
truncates them, but this seems to be an issue for 32-bit
already.
llvm-svn: 289306
Summary:
There is no point in setting SGPRS=104, because VI allocates SGPRs
in multiples of 16, so 104 -> 112. That enables us to use all 102 SGPRs
for general purposes.
Reviewers: tstellarAMD
Subscribers: qcolombet, arsenm, kzhuravl, wdng, nhaehnle, yaxunl, tony-tye
Differential Revision: https://reviews.llvm.org/D27149
llvm-svn: 289260
test/CodeGen/MIR should contain tests that intent to test the MIR
printing or parsing. Tests that test something else should be in
test/CodeGen/TargetName even when they are written in .mir.
As a rule of thumb, only tests using "llc -run-pass none" should be in
test/CodeGen/MIR.
llvm-svn: 289254
Retrying after fixing overly aggressive load-store forwarding optimization.
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search 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.
Whem 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 generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the 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 seemed sufficient to not cause regressions in
tests.
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. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and
merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, dsanders, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 289221
These were selecting directly to the VOP2 form instead
of VOP3 like the i32 instructions. Fixes regressions in
future commits where an immediate isn't folded because it was
initially used for the second operand.
Because uniform 16-bit operations are promoted to i32, it's
difficult to get a simple testcase where this matters. Fold
failures in SIFoldOperands here tend to be hidden by commute
and fold in SIShrinkInstructions.
llvm-svn: 289189
Not having this legal led to combine failures, resulting
in dumb things like bitcasts of constants not being folded
away.
The only reason I'm leaving the v_mov_b32 hack that f32
already uses is to avoid madak formation test regressions.
PeepholeOptimizer has an ordering issue where the immediate
fold attempt is into the sgpr->vgpr copy instead of the actual
use. Running it twice avoids that problem.
llvm-svn: 289096
The correct commutable opcode was set to itself, so this
was simply swapping the operands to commute instead of also
changing the opcode to v_subrev_u16.
llvm-svn: 289093
Multiple metadata values for records such as opencl.ocl.version, llvm.ident
and similar are created after linking several modules. For some of them, notably
opencl.ocl.version, this creates semantic problem because we cannot tell which
version of OpenCL the composite module conforms.
Moreover, such repetitions of identical values often create a huge list of
unneeded metadata, which grows bitcode size both in memory and stored on disk.
It can go up to several Mb when linked against our OpenCL library. Lastly, such
long lists obscure reading of dumped IR.
The pass unifies metadata after linking.
Differential Revision: https://reviews.llvm.org/D25381
llvm-svn: 289092
Summary:
LC can currently select scalar load for uniform memory access
basing on readonly memory address space only. This restriction
originated from the fact that in HW prior to VI vector and scalar caches
are not coherent. With MemoryDependenceAnalysis we can check that the
memory location corresponding to the memory operand of the LOAD is not
clobbered along the all paths from the function entry.
Reviewers: rampitec, tstellarAMD, arsenm
Subscribers: wdng, arsenm, nhaehnle
Differential Revision: https://reviews.llvm.org/D26917
llvm-svn: 289076
Summary:
Without the fix to isFrameOffsetLegal to consider the instruction's
immediate offset, the new test case hits the corresponding assertion in
resolveFrameIndex, because the LocalStackSlotAllocation pass re-uses a
different base register.
With only the fix to isFrameOffsetLegal, code quality reduces in a bunch of
places because frame base registers are added where they're not needed.
This is addressed by properly implementing needsFrameBaseReg, which also
helps to avoid unnecessary zero frame indices in a bunch of other places.
Fixes piglit glsl-1.50/execution/variable-indexing/gs-output-array-vec4-index-wr.shader_test
Reviewers: arsenm, tstellarAMD
Subscribers: qcolombet, kzhuravl, wdng, yaxunl, tony-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D27344
llvm-svn: 289048
Summary:
If we write an immediate to a VGPR and then copy the VGPR to an
SGPR, we can replace the copy with a S_MOV_B32 sgpr, imm, rather than
moving the copy to the SALU.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, llvm-commits, tony-tye
Differential Revision: https://reviews.llvm.org/D27272
llvm-svn: 288849
The structured CFG is just an aid to inserting exec
mask modification instructions, once that is done
we don't really need it anymore. We also
do not analyze blocks with terminators that
modify exec, so this should only be impacting
true branches.
llvm-svn: 288744
Structure the definitions a bit more like the other classes.
The main change here is to split EXP with the done bit set
to a separate opcode, so we can set mayLoad = 1 so that it won't
be reordered before the other exp stores, since this has the special
constraint that if the done bit is set then this should be the last
exp in she shader.
Previously all exp instructions were inferred to have unmodeled
side effects.
llvm-svn: 288695
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
Since the spill is for the whole wave, these
don't have the swizzling problems that vector stores do
and a single 4-byte allocation is enough to spill a 64 element
register. This should reduce the number of spill instructions and
put all the spills for a register in the same cacheline.
This should save allocated private size, but for now it doesn't.
The extra slots are allocated for each component, but never used
because the frame layout is essentially finalized before frame
indices are replaced. For always using the scalar store path,
this should probably be moved into processFunctionBeforeFrameFinalized.
llvm-svn: 288445
The coalescer eliminates copies from reserved registers of the form:
%vregX = COPY %rY
in the case where %rY is a reserved register. However this turns out to
be invalid if only some of the subregisters are reserved (see also
https://reviews.llvm.org/D26648).
Differential Revision: https://reviews.llvm.org/D26687
llvm-svn: 288428
This is not in the list of valid inputs for the encoding.
When spilling, copies from exec can be folded directly
into the spill instruction which results in broken
stores.
This only fixes the operand constraints, more codegen
work is required to avoid emitting the invalid
spills.
This sort of breaks the dbg.value test. Because the
register class of the s_load_dwordx2 changes, there
is a copy to SReg_64, and the copy is the operand
of dbg_value. The copy is later dead, and removed
from the dbg_value.
llvm-svn: 288191
It isn't generally safe to fold the frame index
directly into the operand since it will possibly
not be an inline immediate after it is expanded.
This surprisingly seems to produce better code, since
the FI doesn't prevent folding other immediate operands.
llvm-svn: 288185
Codegen prepare sinks comparisons close to a user is we have only one register
for conditions. For AMDGPU we have many SGPRs capable to hold vector conditions.
Changed BE to report we have many condition registers. That way IR LICM pass
would hoist an invariant comparison out of a loop and codegen prepare will not
sink it.
With that done a condition is calculated in one block and used in another.
Current behavior is to store workitem's condition in a VGPR using v_cndmask_b32
and then restore it with yet another v_cmp instruction from that v_cndmask's
result. To mitigate the issue a propagation of source SGPR pair in place of v_cmp
is implemented. Additional side effect of this is that we may consume less VGPRs
at a cost of more SGPRs in case if holding of multiple conditions is needed, and
that is a clear win in most cases.
Differential Revision: https://reviews.llvm.org/D26114
llvm-svn: 288053
m0 may need to be written for spill code, so
we don't want general code uses relying on the
value stored in it.
This introduces a few code quality regressions where copies
from m0 are not coalesced into copies of a copy of m0.
llvm-svn: 287841
The size and offset were wrong. The size of the object was
being used for the size of the access, when here it is really
being split into 4-byte accesses. The underlying object size
is set in the MachinePointerInfo, which also didn't have the
offset set.
llvm-svn: 287806
A target intrinsic may be defined as possibly reading memory,
but the call site may have additional knowledge that it doesn't read
memory. The intrinsic lowering will expect the pessimistic
assumption of the intrinsic definition, so the chain should
still be used.
llvm-svn: 287593
Summary:
The 32-bit instructions don't zero the high 16-bits like the 16-bit
instructions do.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, llvm-commits, tony-tye
Differential Revision: https://reviews.llvm.org/D26828
llvm-svn: 287342
Summary:
The addr64-based legalization is incorrect for MUBUF instructions with idxen
set as well as for BUFFER_LOAD/STORE_FORMAT_* instructions. This affects
e.g. shaders that access buffer textures.
Since we never actually need the addr64-legalization in shaders, this patch
takes the easy route and keys off the calling convention. If this ever
affects (non-OpenGL) compute, the type of legalization needs to be chosen
based on some TSFlag.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98664
Reviewers: arsenm, tstellarAMD
Subscribers: kzhuravl, wdng, yaxunl, tony-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D26747
llvm-svn: 287339
Summary:
1. Don't try to copy values to and from the same register class.
2. Replace copies with of registers with immediate values with v_mov/s_mov
instructions.
The main purpose of this change is to make MachineSink do a better job of
determining when it is beneficial to split a critical edge, since the pass
assumes that copies will become move instructions.
This prevents a regression in uniform-cfg.ll if we enable critical edge
splitting for AMDGPU.
Reviewers: arsenm
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: https://reviews.llvm.org/D23408
llvm-svn: 287131
wbinvl.* are vector instruction that do not sue vector registers.
v2: check only M?BUF instructions
Differential Revision: https://reviews.llvm.org/D26633
llvm-svn: 287056
The wave barrier represents the discardable barrier. Its main purpose is to
carry convergent attribute, thus preventing illegal CFG optimizations. All lanes
in a wave come to convergence point simultaneously with SIMT, thus no special
instruction is needed in the ISA. The barrier is discarded during code generation.
Differential Revision: https://reviews.llvm.org/D26585
llvm-svn: 287007
Summary:
Extend image intrinsics to support data types of V1F32 and V2F32.
TODO: we should define a mapping table to change the opcode for data type of V2F32 but just one channel is active,
even though such case should be very rare.
Reviewers:
tstellarAMD
Differential Revision:
http://reviews.llvm.org/D26472
llvm-svn: 286860
nThis avoids the nasty problems caused by using
memory instructions that read the exec mask while
spilling / restoring registers used for control flow
masking, but only for VI when these were added.
This always uses the scalar stores when enabled currently,
but it may be better to still try to spill to a VGPR
and use this on the fallback memory path.
The cache also needs to be flushed before wave termination
if a scalar store is used.
llvm-svn: 286766
Summary:
This pass was assuming that when a PHI instruction defined a register
used by another PHI instruction that the defining insstruction would
be legalized before the using instruction.
This assumption was causing the pass to not legalize some PHI nodes
within divergent flow-control.
This fixes a bug that was uncovered by r285762.
Reviewers: nhaehnle, arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, tony-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D26303
llvm-svn: 286676
Matt Arsenault