blob: def67cfae72773a4907276f5b491679866fc2a9e [file] [log] [blame]
// Random ideas for the ARM backend.
Reimplement 'select' in terms of 'SEL'.
* We would really like to support UXTAB16, but we need to prove that the
add doesn't need to overflow between the two 16-bit chunks.
* Implement pre/post increment support. (e.g. PR935)
* Implement smarter constant generation for binops with large immediates.
A few ARMv6T2 ops should be pattern matched: BFI, SBFX, and UBFX
Interesting optimization for PIC codegen on arm-linux:
Crazy idea: Consider code that uses lots of 8-bit or 16-bit values. By the
time regalloc happens, these values are now in a 32-bit register, usually with
the top-bits known to be sign or zero extended. If spilled, we should be able
to spill these to a 8-bit or 16-bit stack slot, zero or sign extending as part
of the reload.
Doing this reduces the size of the stack frame (important for thumb etc), and
also increases the likelihood that we will be able to reload multiple values
from the stack with a single load.
The constant island pass is in good shape. Some cleanups might be desirable,
but there is unlikely to be much improvement in the generated code.
1. There may be some advantage to trying to be smarter about the initial
placement, rather than putting everything at the end.
2. There might be some compile-time efficiency to be had by representing
consecutive islands as a single block rather than multiple blocks.
3. Use a priority queue to sort constant pool users in inverse order of
position so we always process the one closed to the end of functions
first. This may simply CreateNewWater.
Eliminate copysign custom expansion. We are still generating crappy code with
default expansion + if-conversion.
Eliminate one instruction from:
define i32 @_Z6slow4bii(i32 %x, i32 %y) {
%tmp = icmp sgt i32 %x, %y
%retval = select i1 %tmp, i32 %x, i32 %y
ret i32 %retval
cmp r0, r1
movgt r1, r0
mov r0, r1
bx lr
cmp r0, r1
movle r0, r1
bx lr
Implement long long "X-3" with instructions that fold the immediate in. These
were disabled due to badness with the ARM carry flag on subtracts.
More load / store optimizations:
1) Better representation for block transfer? This is from Olden/power:
fldd d0, [r4]
fstd d0, [r4, #+32]
fldd d0, [r4, #+8]
fstd d0, [r4, #+40]
fldd d0, [r4, #+16]
fstd d0, [r4, #+48]
fldd d0, [r4, #+24]
fstd d0, [r4, #+56]
If we can spare the registers, it would be better to use fldm and fstm here.
Need major register allocator enhancement though.
2) Can we recognize the relative position of constantpool entries? i.e. Treat
ldr r0, LCPI17_3
ldr r1, LCPI17_4
ldr r2, LCPI17_5
ldr r0, LCPI17
ldr r1, LCPI17+4
ldr r2, LCPI17+8
Then the ldr's can be combined into a single ldm. See Olden/power.
Note for ARM v4 gcc uses ldmia to load a pair of 32-bit values to represent a
double 64-bit FP constant:
adr r0, L6
ldmia r0, {r0-r1}
.align 2
.long -858993459
.long 1074318540
3) struct copies appear to be done field by field
instead of by words, at least sometimes:
struct foo { int x; short s; char c1; char c2; };
void cpy(struct foo*a, struct foo*b) { *a = *b; }
llvm code (-O2)
ldrb r3, [r1, #+6]
ldr r2, [r1]
ldrb r12, [r1, #+7]
ldrh r1, [r1, #+4]
str r2, [r0]
strh r1, [r0, #+4]
strb r3, [r0, #+6]
strb r12, [r0, #+7]
gcc code (-O2)
ldmia r1, {r1-r2}
stmia r0, {r1-r2}
In this benchmark poor handling of aggregate copies has shown up as
having a large effect on size, and possibly speed as well (we don't have
a good way to measure on ARM).
* Consider this silly example:
double bar(double x) {
double r = foo(3.1);
return x+r;
stmfd sp!, {r4, r5, r7, lr}
add r7, sp, #8
mov r4, r0
mov r5, r1
fldd d0, LCPI1_0
fmrrd r0, r1, d0
bl _foo
fmdrr d0, r4, r5
fmsr s2, r0
fsitod d1, s2
faddd d0, d1, d0
fmrrd r0, r1, d0
ldmfd sp!, {r4, r5, r7, pc}
Ignore the prologue and epilogue stuff for a second. Note
mov r4, r0
mov r5, r1
the copys to callee-save registers and the fact they are only being used by the
fmdrr instruction. It would have been better had the fmdrr been scheduled
before the call and place the result in a callee-save DPR register. The two
mov ops would not have been necessary.
Calling convention related stuff:
* gcc's parameter passing implementation is terrible and we suffer as a result:
struct s {
double d1;
int s1;
void foo(struct s S) {
printf("%g, %d\n", S.d1, S.s1);
'S' is passed via registers r0, r1, r2. But gcc stores them to the stack, and
then reload them to r1, r2, and r3 before issuing the call (r0 contains the
address of the format string):
stmfd sp!, {r7, lr}
add r7, sp, #0
sub sp, sp, #12
stmia sp, {r0, r1, r2}
ldmia sp, {r1-r2}
ldr r0, L5
ldr r3, [sp, #8]
add r0, pc, r0
bl L_printf$stub
Instead of a stmia, ldmia, and a ldr, wouldn't it be better to do three moves?
* Return an aggregate type is even worse:
struct s foo(void) {
struct s S = {1.1, 2};
return S;
mov ip, r0
ldr r0, L5
sub sp, sp, #12
add r0, pc, r0
@ lr needed for prologue
ldmia r0, {r0, r1, r2}
stmia sp, {r0, r1, r2}
stmia ip, {r0, r1, r2}
mov r0, ip
add sp, sp, #12
bx lr
r0 (and later ip) is the hidden parameter from caller to store the value in. The
first ldmia loads the constants into r0, r1, r2. The last stmia stores r0, r1,
r2 into the address passed in. However, there is one additional stmia that
stores r0, r1, and r2 to some stack location. The store is dead.
The llvm-gcc generated code looks like this:
csretcc void %foo(%struct.s* %agg.result) {
%S = alloca %struct.s, align 4 ; <%struct.s*> [#uses=1]
%memtmp = alloca %struct.s ; <%struct.s*> [#uses=1]
cast %struct.s* %S to sbyte* ; <sbyte*>:0 [#uses=2]
call void %llvm.memcpy.i32( sbyte* %0, sbyte* cast ({ double, int }* %C.0.904 to sbyte*), uint 12, uint 4 )
cast %struct.s* %agg.result to sbyte* ; <sbyte*>:1 [#uses=2]
call void %llvm.memcpy.i32( sbyte* %1, sbyte* %0, uint 12, uint 0 )
cast %struct.s* %memtmp to sbyte* ; <sbyte*>:2 [#uses=1]
call void %llvm.memcpy.i32( sbyte* %2, sbyte* %1, uint 12, uint 0 )
ret void
llc ends up issuing two memcpy's (the first memcpy becomes 3 loads from
constantpool). Perhaps we should 1) fix llvm-gcc so the memcpy is translated
into a number of load and stores, or 2) custom lower memcpy (of small size) to
be ldmia / stmia. I think option 2 is better but the current register
allocator cannot allocate a chunk of registers at a time.
A feasible temporary solution is to use specific physical registers at the
lowering time for small (<= 4 words?) transfer size.
* ARM CSRet calling convention requires the hidden argument to be returned by
the callee.
We can definitely do a better job on BB placements to eliminate some branches.
It's very common to see llvm generated assembly code that looks like this:
beq LBB3
b LBB2
If BB4 is the only predecessor of BB3, then we can emit BB3 after BB4. We can
then eliminate beq and turn the unconditional branch to LBB2 to a bne.
See McCat/18-imp/ComputeBoundingBoxes for an example.
Pre-/post- indexed load / stores:
1) We should not make the pre/post- indexed load/store transform if the base ptr
is guaranteed to be live beyond the load/store. This can happen if the base
ptr is live out of the block we are performing the optimization. e.g.
mov r1, r2
ldr r3, [r1], #4
ldr r3, [r2]
add r1, r2, #4
In most cases, this is just a wasted optimization. However, sometimes it can
negatively impact the performance because two-address code is more restrictive
when it comes to scheduling.
Unfortunately, liveout information is currently unavailable during DAG combine
2) Consider spliting a indexed load / store into a pair of add/sub + load/store
to solve #1 (in TwoAddressInstructionPass.cpp).
3) Enhance LSR to generate more opportunities for indexed ops.
4) Once we added support for multiple result patterns, write indexed loads
patterns instead of C++ instruction selection code.
5) Use VLDM / VSTM to emulate indexed FP load / store.
Implement support for some more tricky ways to materialize immediates. For
example, to get 0xffff8000, we can use:
mov r9, #&3f8000
sub r9, r9, #&400000
We sometimes generate multiple add / sub instructions to update sp in prologue
and epilogue if the inc / dec value is too large to fit in a single immediate
operand. In some cases, perhaps it might be better to load the value from a
constantpool instead.
GCC generates significantly better code for this function.
int foo(int StackPtr, unsigned char *Line, unsigned char *Stack, int LineLen) {
int i = 0;
if (StackPtr != 0) {
while (StackPtr != 0 && i < (((LineLen) < (32768))? (LineLen) : (32768)))
Line[i++] = Stack[--StackPtr];
if (LineLen > 32768)
while (StackPtr != 0 && i < LineLen)
return StackPtr;
This should compile to the mlas instruction:
int mlas(int x, int y, int z) { return ((x * y + z) < 0) ? 7 : 13; }
At some point, we should triage these to see if they still apply to us:
gcc generates smaller code for this function at -O2 or -Os:
void foo(signed char* p) {
if (*p == 3)
else if (*p == 4)
else if (*p == 5)
llvm decides it's a good idea to turn the repeated if...else into a
binary tree, as if it were a switch; the resulting code requires -1
compare-and-branches when *p<=2 or *p==5, the same number if *p==4
or *p>6, and +1 if *p==3. So it should be a speed win
(on balance). However, the revised code is larger, with 4 conditional
branches instead of 3.
More seriously, there is a byte->word extend before
each comparison, where there should be only one, and the condition codes
are not remembered when the same two values are compared twice.
More LSR enhancements possible:
1. Teach LSR about pre- and post- indexed ops to allow iv increment be merged
in a load / store.
2. Allow iv reuse even when a type conversion is required. For example, i8
and i32 load / store addressing modes are identical.
int foo(int a, int b, int c, int d) {
long long acc = (long long)a * (long long)b;
acc += (long long)c * (long long)d;
return (int)(acc >> 32);
Should compile to use SMLAL (Signed Multiply Accumulate Long) which multiplies
two signed 32-bit values to produce a 64-bit value, and accumulates this with
a 64-bit value.
We currently get this with both v4 and v6:
smull r1, r0, r1, r0
smull r3, r2, r3, r2
adds r3, r3, r1
adc r0, r2, r0
bx lr
#include <algorithm>
std::pair<unsigned, bool> full_add(unsigned a, unsigned b)
{ return std::make_pair(a + b, a + b < a); }
bool no_overflow(unsigned a, unsigned b)
{ return !full_add(a, b).second; }
Should compile to:
adds r2, r1, r2
movcc r1, #0
movcs r1, #1
str r2, [r0, #0]
strb r1, [r0, #4]
mov pc, lr
cmn r0, r1
movcs r0, #0
movcc r0, #1
mov pc, lr
add r3, r2, r1
str r3, [r0]
mov r2, #1
mov r12, #0
cmp r3, r1
movlo r12, r2
str r12, [r0, #+4]
bx lr
add r3, r1, r0
mov r2, #1
mov r1, #0
cmp r3, r0
movhs r1, r2
mov r0, r1
bx lr
Some of the NEON intrinsics may be appropriate for more general use, either
as target-independent intrinsics or perhaps elsewhere in the ARM backend.
Some of them may also be lowered to target-independent SDNodes, and perhaps
some new SDNodes could be added.
For example, maximum, minimum, and absolute value operations are well-defined
and standard operations, both for vector and scalar types.
The current NEON-specific intrinsics for count leading zeros and count one
bits could perhaps be replaced by the target-independent ctlz and ctpop
intrinsics. It may also make sense to add a target-independent "ctls"
intrinsic for "count leading sign bits". Likewise, the backend could use
the target-independent SDNodes for these operations.
ARMv6 has scalar saturating and halving adds and subtracts. The same
intrinsics could possibly be used for both NEON's vector implementations of
those operations and the ARMv6 scalar versions.
Split out LDR (literal) from normal ARM LDR instruction. Also consider spliting
LDR into imm12 and so_reg forms. This allows us to clean up some code. e.g.
ARMLoadStoreOptimizer does not need to look at LDR (literal) and LDR (so_reg)
while ARMConstantIslandPass only need to worry about LDR (literal).
Constant island pass should make use of full range SoImm values for LEApcrel.
Be careful though as the last attempt caused infinite looping on lencod.
Predication issue. This function:
extern unsigned array[ 128 ];
int foo( int x ) {
int y;
y = array[ x & 127 ];
if ( x & 128 )
y = 123456789 & ( y >> 2 );
y = 123456789 & y;
return y;
compiles to:
and r1, r0, #127
ldr r2, LCPI1_0
ldr r2, [r2]
ldr r1, [r2, +r1, lsl #2]
mov r2, r1, lsr #2
tst r0, #128
moveq r2, r1
ldr r0, LCPI1_1
and r0, r2, r0
bx lr
It would be better to do something like this, to fold the shift into the
conditional move:
and r1, r0, #127
ldr r2, LCPI1_0
ldr r2, [r2]
ldr r1, [r2, +r1, lsl #2]
tst r0, #128
movne r1, r1, lsr #2
ldr r0, LCPI1_1
and r0, r1, r0
bx lr
it saves an instruction and a register.
It might be profitable to cse MOVi16 if there are lots of 32-bit immediates
with the same bottom half.
Robert Muth started working on an alternate jump table implementation that
does not put the tables in-line in the text. This is more like the llvm
default jump table implementation. This might be useful sometime. Several
revisions of patches are on the mailing list, beginning at:
Make use of the "rbit" instruction.
Take a look at test/CodeGen/Thumb2/machine-licm.ll. ARM should be taught how
to licm and cse the unnecessary load from cp#1.
The CMN instruction sets the flags like an ADD instruction, while CMP sets
them like a subtract. Therefore to be able to use CMN for comparisons other
than the Z bit, we'll need additional logic to reverse the conditionals
associated with the comparison. Perhaps a pseudo-instruction for the comparison,
with a post-codegen pass to clean up and handle the condition codes?
See PR5694 for testcase.
Given the following on armv5:
int test1(int A, int B) {
return (A&-8388481)|(B&8388480);
We currently generate:
ldr r2, .LCPI0_0
and r0, r0, r2
ldr r2, .LCPI0_1
and r1, r1, r2
orr r0, r1, r0
bx lr
We should be able to replace the second ldr+and with a bic (i.e. reuse the
constant which was already loaded). Not sure what's necessary to do that.
The code generated for bswap on armv4/5 (CPUs without rev) is less than ideal:
int a(int x) { return __builtin_bswap32(x); }
mov r1, #255, 24
mov r2, #255, 16
and r1, r1, r0, lsr #8
and r2, r2, r0, lsl #8
orr r1, r1, r0, lsr #24
orr r0, r2, r0, lsl #24
orr r0, r0, r1
bx lr
Something like the following would be better (fewer instructions/registers):
eor r1, r0, r0, ror #16
bic r1, r1, #0xff0000
mov r1, r1, lsr #8
eor r0, r1, r0, ror #8
bx lr
A custom Thumb version would also be a slight improvement over the generic
Consider the following simple C code:
void foo(unsigned char *a, unsigned char *b, int *c) {
if ((*a | *b) == 0) *c = 0;
currently llvm-gcc generates something like this (nice branchless code I'd say):
ldrb r0, [r0]
ldrb r1, [r1]
orr r0, r1, r0
tst r0, #255
moveq r0, #0
streq r0, [r2]
bx lr
Note that both "tst" and "moveq" are redundant.
When loading immediate constants with movt/movw, if there are multiple
constants needed with the same low 16 bits, and those values are not live at
the same time, it would be possible to use a single movw instruction, followed
by multiple movt instructions to rewrite the high bits to different values.
For example:
volatile store i32 -1, i32* inttoptr (i32 1342210076 to i32*), align 4,
volatile store i32 -1, i32* inttoptr (i32 1342341148 to i32*), align 4,
is compiled and optimized to:
movw r0, #32796
mov.w r1, #-1
movt r0, #20480
str r1, [r0]
movw r0, #32796 @ <= this MOVW is not needed, value is there already
movt r0, #20482
str r1, [r0]
Improve codegen for select's:
if (x != 0) x = 1
if (x == 1) x = 1
ARM codegen used to look like this:
mov r1, r0
cmp r1, #1
mov r0, #0
moveq r0, #1
The naive lowering select between two different values. It should recognize the
test is equality test so it's more a conditional move rather than a select:
cmp r0, #1
movne r0, #0
Currently this is a ARM specific dag combine. We probably should make it into a
target-neutral one.
Optimize unnecessary checks for zero with __builtin_clz/ctz. Those builtins
are specified to be undefined at zero, so portable code must check for zero
and handle it as a special case. That is unnecessary on ARM where those
operations are implemented in a way that is well-defined for zero. For
int f(int x) { return x ? __builtin_clz(x) : sizeof(int)*8; }
should just be implemented with a CLZ instruction. Since there are other
targets, e.g., PPC, that share this behavior, it would be best to implement
this in a target-independent way: we should probably fold that (when using
"undefined at zero" semantics) to set the "defined at zero" bit and have
the code generator expand out the right code.
Clean up the test/MC/ARM files to have more robust register choices.
R0 should not be used as a register operand in the assembler tests as it's then
not possible to distinguish between a correct encoding and a missing operand
encoding, as zero is the default value for the binary encoder.
add r0, r0 // bad
add r3, r5 // good
Register operands should be distinct. That is, when the encoding does not
require two syntactical operands to refer to the same register, two different
registers should be used in the test so as to catch errors where the
operands are swapped in the encoding.
subs.w r1, r1, r1 // bad
subs.w r1, r2, r3 // good