Improved handling of fma, floating point min/max, additional load/store
instructions for floating point types.
Patch by Jyotsna Verma.
llvm-svn: 279239
The pipeliner was generating an invalid Phi name for an operand
in the epilog block, which caused an assert in the live variable
analysis pass. The fix is to the code that generates new Phis
in the epilog block. In this case, there is an existing Phi that
needs to be reused rather than creating a new Phi instruction.
Differential Revision: https://reviews.llvm.org/D23513
llvm-svn: 278805
If a loop is not rotated (for example when optimizing for size), the latch is not the backedge. If we promote an expression to post-inc form, we not only increase register pressure and add a COPY for that IV expression but for all IVs!
Motivating testcase:
void f(float *a, float *b, float *c, int n) {
while (n-- > 0)
*c++ = *a++ + *b++;
}
It's imperative that the pointer increments be located in the latch block and not the header block; if not, we cannot use post-increment loads and stores and we have to keep both the post-inc and pre-inc values around until the end of the latch which bloats register usage.
llvm-svn: 278658
LowerTargetConstantPool is not properly setting the TargetFlag to indicate
desired relocation. Coding error, the offset parameter was omitted, so the
TargetFlag was used as the offset, and the TargetFlag defaulted to zero.
This only affects -fpic compilation, and only those items created in a
Constant Pool, for example a vector of constants. Halide ran into this issue.
llvm-svn: 278614
From the point of view of register assignment, byval parameters are
ignored: a byval parameter is not going to be assigned to a register,
and it will not affect the assignments of subsequent parameters.
When matching registers with parameters in the bit tracker, make sure
to skip byval parameters before advancing the registers.
llvm-svn: 278375
This change makes it possible for tail-duplication and tail-merging to
be disjoint. By being less aggressive when merging during layout, there are no
overlapping cases between tail-duplication and tail-merging, provided the
thresholds are disjoint.
There is a remaining TODO to benchmark the succ_size() test for non-layout tail
merging.
llvm-svn: 278265
Floating point instructions use general purpose registers, so the few
instructions that can put floating point immediates into registers are,
in fact, integer instruction. Use them explicitly instead of having
pseudo-instructions specifically for dealing with floating point values.
Simplify the constant loading instructions (from sdata) to have only two:
one for 32-bit values and one for 64-bit values: CONST32 and CONST64.
llvm-svn: 278244
When the same base address is used to load two different data types, LSR
would assume a memory type of "void". This type is not sized and has no
alignment information. Checking for it causes a crash.
llvm-svn: 277601
Identify patterns where the address is aligned to an 8-byte boundary,
but both the base address and the constant offset are both proper
multiples of 4. In such cases, extract Base+4 into a separate instruc-
tion, and use S2_storerd_io, instead of using S4_storerd_rr.
llvm-svn: 277497
Scavenging slots were only reserved when pseudo-instruction expansion in
frame lowering created new virtual registers. It is possible to still
need a scavenging slot even if no virtual registers were created, in cases
where the stack is large enough to overflow instruction offsets.
llvm-svn: 277355
The DAG combiner will try to merge consecutive stores into a bigger
store, unless the resulting store is not fast. Misaligned vector stores
are allowed on Hexagon, but are not fast. Add a testcase to make sure
this type of merging does not occur.
Patch by Pranav Bhandarkar.
llvm-svn: 277182
The DAG combiner tries to merge stores to adjacent vector wide memory
locations by creating stores which are integral multiples of the vector
width. Discourage this by informing it that this is slow. This should
not affect legalization passes, because all of them ignore the "Fast"
argument.
Patch by Pranav Bhandarkar.
llvm-svn: 277178
Software pipelining is an optimization for improving ILP by
overlapping loop iterations. Swing Modulo Scheduling (SMS) is
an implementation of software pipelining that attempts to
reduce register pressure and generate efficient pipelines with
a low compile-time cost.
This implementaion of SMS is a target-independent back-end pass.
When enabled, the pass should run just prior to the register
allocation pass, while the machine IR is in SSA form. If the pass
is successful, then the original loop is replaced by the optimized
loop. The optimized loop contains one or more prolog blocks, the
pipelined kernel, and one or more epilog blocks.
This pass is enabled for Hexagon only. To enable for other targets,
a couple of target specific hooks must be implemented, and the
pass needs to be called from the target's TargetMachine
implementation.
Differential Review: http://reviews.llvm.org/D16829
llvm-svn: 277169
If the mask of a vector shuffle has alternating odd or even numbers
starting with 1 or 0 respectively up to the largest possible index
for the given type in the given HVX mode (single of double) we can
generate vpacko or vpacke instruction respectively.
E.g.
%42 = shufflevector <32 x i16> %37, <32 x i16> %41,
<32 x i32> <i32 1, i32 3, ..., i32 63>
is %42.h = vpacko(%41.w, %37.w)
Patch by Pranav Bhandarkar.
llvm-svn: 277168
Rebalances address calculation trees and applies Hexagon-specific
optimizations to the trees to improve instruction selection.
Patch by Tobias Edler von Koch.
llvm-svn: 277151
Normally, CFI instructions should be inserted after allocframe, but
if allocframe is in the same packet with a call, the CFI instructions
should be inserted before that packet.
llvm-svn: 277020
Before adding a new preheader block, check if there is a candidate block
where the loop setup could be placed speculatively. This will be off by
default.
llvm-svn: 276919
Consider this case:
vreg1 = A2_zxth vreg0 (1)
...
vreg2 = A2_zxth vreg1 (2)
Redundant instruction elimination could delete the instruction (1)
because the user (2) only cares about the low 16 bits. Then it could
delete (2) because the input is already zero-extended. The problem
is that the properties allowing each individual instruction to be
deleted depend on the existence of the other instruction, so either
one can be deleted, but not both.
The existing check for this situation in RIE was insufficient. The
fix is to update all dependent cells when an instruction is removed
(replaced via COPY) in RIE.
llvm-svn: 276792
Schedule a load and its use in the same packet in MISched. Previously,
isResourceAvailable was returning false for dependences in the same
packet, which prevented MISched from packetizing a load and its use in
the same packet for v60.
Patch by Ikhlas Ajbar.
llvm-svn: 275804
The machine scheduler needs to account for available resources
more accurately in order to avoid scheduling an instruction that
forces a new packet to be created.
This occurs in two ways: First, an instruction without an available
resource may have a large priority due to other metrics and be
scheduled when there are other instructions with available resources.
Second, an instruction with a non-zero latency may become available
prematurely. In both these cases, we attempt change the priority
in order to allow a better instruction to be scheduled.
Patch by Brendon Cahoon.
llvm-svn: 275793
An instruction may have multiple predecessors that are candidates
for using .cur. However, only one of them can use .cur in the
packet. When this case occurs, we need to make sure that only
one of the dependences gets a 0 latency value.
Patch by Brendon Cahoon.
llvm-svn: 275790
- Treat bitwise OR with a frame index as an ADD wherever possible, fold it
into addressing mode.
- Extend patterns for memops to allow memops with frame indexes as address
operands.
llvm-svn: 275569