In order to set breakpoints on labels and list source code around
labels, we need collect debug information for labels, i.e., label
name, the function label belong, line number in the file, and the
address label located. In order to keep these information in LLVM
IR and to allow backend to generate debug information correctly.
We create a new kind of metadata for labels, DILabel. The format
of DILabel is
!DILabel(scope: !1, name: "foo", file: !2, line: 3)
We hope to keep debug information as much as possible even the
code is optimized. So, we create a new kind of intrinsic for label
metadata to avoid the metadata is eliminated with basic block.
The intrinsic will keep existing if we keep it from optimized out.
The format of the intrinsic is
llvm.dbg.label(metadata !1)
It has only one argument, that is the DILabel metadata. The
intrinsic will follow the label immediately. Backend could get the
label metadata through the intrinsic's parameter.
We also create DIBuilder API for labels to be used by Frontend.
Frontend could use createLabel() to allocate DILabel objects, and use
insertLabel() to insert llvm.dbg.label intrinsic in LLVM IR.
Differential Revision: https://reviews.llvm.org/D45024
Patch by Hsiangkai Wang.
llvm-svn: 331841
Commit 37962a331c77 ("bpf: Improve expanding logic in LowerSELECT_CC")
intended to improve code quality for certain jmp conditions. The
commit, however, has a couple of issues:
(1). In code, just swap is not enough, ConditionalCode CC
should also be swapped, otherwise incorrect code will
be generated.
(2). The ConditionalCode swap should be subject to
getHasJmpExt(). If getHasJmpExt() is False, certain
conditional codes will not be supported and swap
may generate incorrect code.
The original goal for this patch is to optimize jmp operations
which does not have JmpExt turned on. If JmpExt is on,
better code could be generated. For example, the test
select_ri.ll is introduced to demonstrate the optimization.
The same result can be achieved with -mcpu=v2 flag.
Signed-off-by: Yonghong Song <yhs@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
llvm-svn: 329043
The current zero extension elimination was restricted to operands of
comparison. It actually could be extended to more cases.
For example:
int *inc_p (int *p, unsigned a)
{
return p + a;
}
'a' will be promoted to i64 during addition, and the zero extension could
be eliminated as well.
For the elimination optimization, it should be much better to start
recognizing the candidate sequence from the SRL instruction instead of J*
instructions.
This patch makes it an generic zero extension elimination pass instead of
one restricted with comparison.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 327367
There is a mistake in current code that we "break" out the optimization
when the first operand of J*_RR doesn't qualify the elimination. This
caused some elimination opportunities missed, for example the one in the
testcase.
The code should just fall through to handle the second operand.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 327366
The current subregister definition check stops after the MOV_32_64
instruction.
This means we are thinking all the following instruction sequences
are safe to be eliminated:
MOV_32_64 rB, wA
SLL_ri rB, rB, 32
SRL_ri rB, rB, 32
However, this is *not* true. The source subregister wA of MOV_32_64 could
come from a implicit truncation of 64-bit register in which case the high
bits of the 64-bit register is not zeroed, therefore we can't eliminate
above sequence.
For example, for i32_val, we shouldn't do the elimination:
long long bar ();
int foo (int b, int c)
{
unsigned int i32_val = (unsigned int) bar();
if (i32_val < 10)
return b;
else
return c;
}
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 327365
Improve the test accuracy by adding more check directives.
Shifts are expected to be eliminated for zero extension but not for signed
extension.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 327364
This patch adds some unit tests for 32-bit subregister support.
We want to make sure ALU32, subregister load/store and new peephole
optimization are truely enabled once -mattr=+alu32 specified.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Yonghong Song <yhs@fb.com>
llvm-svn: 325992
In DWARF v5 the Line Number Program Header is extensible, allowing values with
new content types. In this extension a content type is added,
DW_LNCT_LLVM_source, which contains the embedded source code of the file.
Add new optional attribute for !DIFile IR metadata called source which contains
source text. Use this to output the source to the DWARF line table of code
objects. Analogously extend METADATA_FILE in Bitcode and .file directive in ASM
to support optional source.
Teach llvm-dwarfdump and llvm-objdump about the new values. Update the output
format of llvm-dwarfdump to make room for the new attribute on file_names
entries, and support embedded sources for the -source option in llvm-objdump.
Differential Revision: https://reviews.llvm.org/D42765
llvm-svn: 325970
Enable multiple COPY hints to eliminate more COPYs during register allocation.
Note that this is something all targets should do, see
https://reviews.llvm.org/D38128.
Review: Yonghong Song
llvm-svn: 325457
The reference '&' is missing in the function parameter. If there are
back-to-back optimizations in terms of dag node list like below:
t29: i64,ch = load<LD4[bitcast (%struct.test_t* @test.t to i8*)+12](dereferenceable), zext from i32> t3, t43, undef:i64
t34: i64,ch = load<LD4[bitcast (%struct.test_t* @test.t to i8*)](dereferenceable), zext from i32> t3, t41, undef:i64
The bug will trigger a segfault for the added test case remove_truncate_5.ll:
LLVMSymbolizer: error reading file: No such file or directory
#0 0x000000000241c4d9 (llc+0x241c4d9)
#1 0x000000000241c56a (llc+0x241c56a)
#2 0x000000000241aa50 (llc+0x241aa50)
...
#22 0x0000000000fd5edf (llc+0xfd5edf)
#23 0x00007f0fe03bec05 __libc_start_main (/lib64/libc.so.6+0x21c05)
#24 0x0000000000fd3e69 (llc+0xfd3e69)
...
Segmentation fault
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 325267
LowerSELECT_CC is not generating optimal Select_Ri pattern at the moment. It
is not guaranteed to place ConstantNode at RHS which would miss matching
Select_Ri.
A new testcase added into the existing select_ri.ll, also there is an
existing case in cmp.ll which would be improved to use Select_Ri after this
patch, it is adjusted accordingly.
Reported-by: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Reviewed-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
llvm-svn: 324560
The major visible difference here is that in line-table dumps,
directory and file names are wrapped in double-quotes; previously,
directory names got single quotes and file names were not quoted at
all.
The improvement in this patch is that when a DWARF v5 line table
header has indirect strings, in a verbose dump these will all have
their section[offset] printed as well as the name itself. This
matches the format used for dumping strings in the .debug_info
section.
Differential Revision: https://reviews.llvm.org/D42802
llvm-svn: 324270
Discussed here:
http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html
In preparation for adding support for named vregs we are changing the sigil for
physical registers in MIR to '$' from '%'. This will prevent name clashes of
named physical register with named vregs.
llvm-svn: 323922
As commented on the existing code:
// The Reg operand should be a virtual register, which is defined
// outside the current basic block. DAG combiner has done a pretty
// good job in removing truncating inside a single basic block.
However, when the Reg operand comes from bpf_load_[byte | half | word]
intrinsics, the generic optimizer doesn't understand their results are
zero extended, so these single basic block elimination opportunities were
missed.
Acked-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
llvm-svn: 322534
Add support for 'objdump -print-imm-hex' for imm64, operand imm
and branch target. If user programs encode immediate values
as hex numbers, such an option will make it easy to correlate
asm insns with source code. This option also makes it easy
to correlate imm values with insn encoding.
There is one changed behavior in this patch. In old way, we
print the 64bit imm as u64:
O << (uint64_t)Op.getImm();
and the new way is:
O << formatImm(Op.getImm());
The formatImm is defined in llvm/MC/MCInstPrinter.h as
format_object<int64_t> formatImm(int64_t Value)
So the new way to print 64bit imm is i64 type.
If a 64bit value has the highest bit set, the old way
will print the value as a positive value and the
new way will print as a negative value. The new way
is consistent with x86_64.
For the code (see the test program):
...
if (a == 0xABCDABCDabcdabcdULL)
...
x86_64 objdump, with and without -print-imm-hex, looks like:
48 b8 cd ab cd ab cd ab cd ab movabsq $-6067004223159161907, %rax
48 b8 cd ab cd ab cd ab cd ab movabsq $-0x5432543254325433, %rax
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 321215
As part of the unification of the debug format and the MIR format,
always print registers as lowercase.
* Only debug printing is affected. It now follows MIR.
Differential Revision: https://reviews.llvm.org/D40417
llvm-svn: 319187
Summary:
Now that store-merge is only generates type-safe stores, do a second
pass just before instruction selection to allow lowered intrinsics to
be merged as well.
Reviewers: jyknight, hfinkel, RKSimon, efriedma, rnk, jmolloy
Subscribers: javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D33675
llvm-svn: 319036
Commit b5cbc7760ab8 ("[bpf] allow direct and indirect calls")
allowed more than one function in the bpf program, and
commit 114353884415 ("bpf: fix a bug in trunc-op optimization")
fixed a bug in trunc-op optimization which only showed up
with more than one function in the bpf program.
This patch added a test case for trunc-op optimization
for bpf programs with two functions. Reverting commit
"bpf: fix a bug in trunc-op optimization" will cause
failure for this test case.
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 318695
kernel verifier is becoming smarter and soon will support
direct and indirect function calls.
Remove obsolete error from BPF backend.
Make call to use PCRel_4 fixup.
'bpf to bpf' calls are distinguished from 'bpf to kernel' calls
by insn->src_reg == BPF_PSEUDO_CALL == 1 which is used as relocation
indicator similar to ld_imm64->src_reg == BPF_PSEUDO_MAP_FD == 1
The actual 'call' instruction remains the same for both
'bpf to kernel' and 'bpf to bpf' calls.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
llvm-svn: 318614
Add hook in BPF backend so that llvm-objdump can print out
the jmp target with label names, e.g.,
...
if r1 != 2 goto 6 <LBB0_2>
...
goto 7 <LBB0_4>
...
LBB0_2:
...
LBB0_4:
...
Signed-off-by: Yonghong Song <yhs@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
llvm-svn: 318358
We came across an llvm bug when compiling some testcases that 64-bit
immediates are silently truncated into 32-bit and then packed into
BPF_JMP | BPF_K encoding. This caused comparison with wrong value.
This bug looks to be introduced by r308080. The Select_Ri pattern is
supposed to be lowered into J*_Ri while the latter only support 32-bit
immediate encoding, therefore Select_Ri should have similar immediate
predicate check as what J*_Ri are doing.
Reported-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Yonghong Song <yhs@fb.com>
llvm-svn: 315889
This patch adds new insn, "reg = be16/be32/be64 reg",
for bswap to little endian for big-endian target (bpfeb).
It also adds new insn for negation "reg = -reg".
Currently, for source code, e.g.,
b = -a
LLVM still prefers to generate:
b = 0 - a
But "reg = -reg" format can be used in assembly code.
Signed-off-by: Yonghong Song <yhs@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
llvm-svn: 314376
This partially revert previous fix in commit f5858045aa0b
("bpf: proper print imm64 expression in inst printer").
In that commit, the original suffix "ll" is removed from
LD_IMM64 asmstring. In the customer print method, the "ll"
suffix is printed if the rhs is an immediate. For example,
"r2 = 5ll" => "r2 = 5ll", and "r3 = varll" => "r3 = var".
This has an issue though for assembler. Since assembler
relies on asmstring to do pattern matching, it will not
be able to distiguish between "mov r2, 5" and
"ld_imm64 r2, 5" since both asmstring is "r2 = 5".
In such cases, the assembler uses 64bit load for all
"r = <val>" asm insts.
This patch adds back " ll" suffix for ld_imm64 with one
additional space for "#reg = #global_var" case.
Signed-off-by: Yonghong Song <yhs@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
llvm-svn: 312978
As discussed on llvm-dev in
http://lists.llvm.org/pipermail/llvm-dev/2017-September/117301.html
this changes the command line interface of llvm-dwarfdump to match the
one used by the dwarfdump utility shipping on macOS. In addition to
being shorter to type this format also has the advantage of allowing
more than one section to be specified at the same time.
In a nutshell, with this change
$ llvm-dwarfdump --debug-dump=info
$ llvm-dwarfdump --debug-dump=apple-objc
becomes
$ dwarfdump --debug-info --apple-objc
Differential Revision: https://reviews.llvm.org/D37714
llvm-svn: 312970
This change simplifies code that has to deal with
DIGlobalVariableExpression and mirrors how we treat DIExpressions in
debug info intrinsics. Before this change there were two ways of
representing empty expressions on globals, a nullptr and an empty
!DIExpression().
If someone needs to upgrade out-of-tree testcases:
perl -pi -e 's/(!DIGlobalVariableExpression\(var: ![0-9]*)\)/\1, expr: !DIExpression())/g' <MYTEST.ll>
will catch 95%.
llvm-svn: 312144
-mcpu=# will support:
. generic: the default insn set
. v1: insn set version 1, the same as generic
. v2: insn set version 2, version 1 + additional jmp insns
. probe: the compiler will probe the underlying kernel to
decide proper version of insn set.
We did not not use -mcpu=native since llc/llvm will interpret -mcpu=native
as the underlying hardware architecture regardless of -march value.
Currently, only x86_64 supports -mcpu=probe. Other architecture will
silently revert to "generic".
Also added -mcpu=help to print available cpu parameters.
llvm will print out the information only if there are at least one
cpu and at least one feature. Add an unused dummy feature to
enable the printout.
Examples for usage:
$ llc -march=bpf -mcpu=v1 -filetype=asm t.ll
$ llc -march=bpf -mcpu=v2 -filetype=asm t.ll
$ llc -march=bpf -mcpu=generic -filetype=asm t.ll
$ llc -march=bpf -mcpu=probe -filetype=asm t.ll
$ llc -march=bpf -mcpu=v3 -filetype=asm t.ll
'v3' is not a recognized processor for this target (ignoring processor)
...
$ llc -march=bpf -mcpu=help -filetype=asm t.ll
Available CPUs for this target:
generic - Select the generic processor.
probe - Select the probe processor.
v1 - Select the v1 processor.
v2 - Select the v2 processor.
Available features for this target:
dummy - unused feature.
Use +feature to enable a feature, or -feature to disable it.
For example, llc -mcpu=mycpu -mattr=+feature1,-feature2
...
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Yonghong Song <yhs@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
llvm-svn: 311522
This pattern shows up when lowering byval copies on AMDGPU.
The byval object access is split into 4-byte chunks, adding a
constant offset to the FixedStack base. When some of the offsets
turn into ors, this prevents combining the constant offsets.
This makes it not apparent that the object is there when matching
addressing modes, so it ends up using a scratch wave offset
relative access and the lengthy frame index expansion for that.
llvm-svn: 309775
Check the actual memory type stored and not the extended value size
when considering if truncated store merge is worthwhile.
Reviewers: efriedma, RKSimon, spatel, jyknight
Reviewed By: efriedma
Subscribers: llvm-commits, nhaehnle
Differential Revision: https://reviews.llvm.org/D35623
llvm-svn: 308833
Currently, for code like below,
===
inner_map = bpf_map_lookup_elem(outer_map, &port_key);
if (!inner_map) {
inner_map = &fallback_map;
}
===
the compiler generates (pseudo) code like the below:
===
I1: r1 = bpf_map_lookup_elem(outer_map, &port_key);
I2: r2 = 0
I3: if (r1 == r2)
I4: r6 = &fallback_map
I5: ...
===
During kernel verification process, After I1, r1 holds a state
map_ptr_or_null. If I3 condition is not taken
(path [I1, I2, I3, I5]), supposedly r1 should become map_ptr.
Unfortunately, kernel does not recognize this pattern
and r1 remains map_ptr_or_null at insn I5. This will cause
verificaiton failure later on.
Kernel, however, is able to recognize pattern "if (r1 == 0)"
properly and give a map_ptr state to r1 in the above case.
LLVM here generates suboptimal code which causes kernel verification
failure. This patch fixes the issue by changing BPF insn pattern
matching and lowering to generate proper codes if the righthand
parameter of the above condition is a constant. A test case
is also added.
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 308080
Relanding after rewriting undef.ll test to avoid host-dependant
endianness.
As discussed in D34087, rewrite areNonVolatileConsecutiveLoads using
generic checks. Also, propagate missing local handling from there to
BaseIndexOffset checks.
Tests of note:
* test/CodeGen/X86/build-vector* - Improved.
* test/CodeGen/BPF/undef.ll - Improved store alignment allows an
additional store merge
* test/CodeGen/X86/clear_upper_vector_element_bits.ll - This is a
case we already do not handle well. Here, the DAG is improved, but
scheduling causes a code size degradation.
Reviewers: RKSimon, craig.topper, spatel, andreadb, filcab
Subscribers: nemanjai, llvm-commits
Differential Revision: https://reviews.llvm.org/D34472
llvm-svn: 307114
As discussed in D34087, rewrite areNonVolatileConsecutiveLoads using
generic checks. Also, propagate missing local handling from there to
BaseIndexOffset checks.
Tests of note:
* test/CodeGen/X86/build-vector* - Improved.
* test/CodeGen/BPF/undef.ll - Improved store alignment allows an
additional store merge
* test/CodeGen/X86/clear_upper_vector_element_bits.ll - This is a
case we already do not handle well. Here, the DAG is improved, but
scheduling causes a code size degradation.
Reviewers: RKSimon, craig.topper, spatel, andreadb, filcab
Subscribers: nemanjai, llvm-commits
Differential Revision: https://reviews.llvm.org/D34472
llvm-svn: 306819
For networking-type bpf program, it often needs to access
packet data. A context data structure is provided to the bpf
programs with two fields:
u32 data;
u32 data_end;
User can access these two fields with ctx->data and ctx->data_end.
During program verification process, the kernel verifier modifies
the bpf program with loading of actual pointer value from kernel
data structure.
r = ctx->data ===> r = actual data start ptr
r = ctx->data_end ===> r = actual data end ptr
A typical program accessing ctx->data like
char *data_ptr = (char *)(long)ctx->data
will result in a 32-bit load followed by a zero extension.
Such an operation is combined into a single LDW in DAG combiner
as bpf LDW does zero extension automatically.
In cases like the below (which can be a result of global value numbering
and partial redundancy elimination before insn selection):
B1:
u32 a = load-32-bit &ctx->data
u64 pa = zext a
...
B2:
u32 b = load-32-bit &ctx->data
u64 pb = zext b
...
B3:
u32 m = PHI(a, b)
u64 pm = zext m
In B3, "pm = zext m" cannot be removed, which although is legal
from compiler perspective, will generate incorrect code after
kernel verification.
This patch recognizes this pattern and traces through PHI node
to see whether the operand of "zext m" is defined with LDWs or not.
If it is, the "zext m" itself can be removed.
The patch also recognizes the pattern where the load and use of
the load value not in the same basic block, where truncate operation
may be removed as well.
The patch handles 1-byte, 2-byte and 4-byte truncation.
Two test cases are added to verify the transformation happens properly
for the above code pattern.
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 306685
If users tried to have a structure decl/init code like below
struct test_t t = { .memeber1 = 45 };
It is very likely that compiler will generate a readonly section
to hold up the init values for variable t. Later load of t members,
e.g., t.member1 will result in a read from readonly section.
BPF program cannot handle relocation. This will force users to
write:
struct test_t t = {};
t.member1 = 45;
This is just inconvenient and unintuitive.
This patch addresses this issue by implementing BPF PreprocessISelDAG.
For any load from a global constant structure or an global array of
constant struct, it attempts to
translate it into a constant directly. The traversal of the
constant struct and other constant data structures are similar
to where the assembler emits read-only sections.
Four different unit test cases are also added to cover
different scenarios.
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 305560
o Add bpfeb support in BPF dwarfdump unit test case
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Alexei Starovoitov <ast@fb.com>
llvm-svn: 302265
. there should be no runtime relocation inside the bpf function.
. relocation supported here mostly for debugging.
. a test case is added.
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
llvm-svn: 302055
. swap 4-bit register encoding, 16-bit offset and 32-bit imm to support big endian archs
. add a test
Reported-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
llvm-svn: 301653
If the offset cannot fit into the instruction, an addition to the
pointer is emitted before the actual access. However, BPF offsets are
16-bit but LLVM considers them to be, for the matter of this check,
to be 32-bit long.
This causes the following program:
int bpf_prog1(void *ign)
{
volatile unsigned long t = 0x8983984739ull;
return *(unsigned long *)((0xffffffff8fff0002ull) + t);
}
To generate the following (wrong) code:
0: 18 01 00 00 39 47 98 83 00 00 00 00 89 00 00 00
r1 = 590618314553ll
2: 7b 1a f8 ff 00 00 00 00 *(u64 *)(r10 - 8) = r1
3: 79 a1 f8 ff 00 00 00 00 r1 = *(u64 *)(r10 - 8)
4: 79 10 02 00 00 00 00 00 r0 = *(u64 *)(r1 + 2)
5: 95 00 00 00 00 00 00 00 exit
Fix it by changing the offset check to 16-bit.
Patch by Nadav Amit <nadav.amit@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Differential Revision: https://reviews.llvm.org/D32055
llvm-svn: 300269
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
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
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
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
