path: root/lib/Target/X86/README-SSE.txt

//===---------------------------------------------------------------------===//
// Random ideas for the X86 backend: SSE-specific stuff.
//===---------------------------------------------------------------------===//

- Consider eliminating the unaligned SSE load intrinsics, replacing them with
  unaligned LLVM load instructions.

//===---------------------------------------------------------------------===//

Expand libm rounding functions inline:  Significant speedups possible.
http://gcc.gnu.org/ml/gcc-patches/2006-10/msg00909.html

//===---------------------------------------------------------------------===//

When compiled with unsafemath enabled, "main" should enable SSE DAZ mode and
other fast SSE modes.

//===---------------------------------------------------------------------===//

Think about doing i64 math in SSE regs.

//===---------------------------------------------------------------------===//

This testcase should have no SSE instructions in it, and only one load from
a constant pool:

double %test3(bool %B) {
        %C = select bool %B, double 123.412, double 523.01123123
        ret double %C
}

Currently, the select is being lowered, which prevents the dag combiner from
turning 'select (load CPI1), (load CPI2)' -> 'load (select CPI1, CPI2)'

The pattern isel got this one right.

//===---------------------------------------------------------------------===//

SSE doesn't have [mem] op= reg instructions.  If we have an SSE instruction
like this:

  X += y

and the register allocator decides to spill X, it is cheaper to emit this as:

Y += [xslot]
store Y -> [xslot]

than as:

tmp = [xslot]
tmp += y
store tmp -> [xslot]

..and this uses one fewer register (so this should be done at load folding
time, not at spiller time).  *Note* however that this can only be done
if Y is dead.  Here's a testcase:

@.str_3 = external global [15 x i8]		; <[15 x i8]*> [#uses=0]
declare void @printf(i32, ...)
define void @main() {
build_tree.exit:
	br label %no_exit.i7

no_exit.i7:		; preds = %no_exit.i7, %build_tree.exit
	%tmp.0.1.0.i9 = phi double [ 0.000000e+00, %build_tree.exit ], [ %tmp.34.i18, %no_exit.i7 ]		; <double> [#uses=1]
	%tmp.0.0.0.i10 = phi double [ 0.000000e+00, %build_tree.exit ], [ %tmp.28.i16, %no_exit.i7 ]		; <double> [#uses=1]
	%tmp.28.i16 = add double %tmp.0.0.0.i10, 0.000000e+00		; <double> [#uses=1]
	%tmp.34.i18 = add double %tmp.0.1.0.i9, 0.000000e+00		; <double> [#uses=2]
	br i1 false, label %Compute_Tree.exit23, label %no_exit.i7

Compute_Tree.exit23:		; preds = %no_exit.i7
	tail call void (i32, ...)* @printf( i32 0 )
	store double %tmp.34.i18, double* null
	ret void
}

We currently emit:

.BBmain_1:
        xorpd %XMM1, %XMM1
        addsd %XMM0, %XMM1
***     movsd %XMM2, QWORD PTR [%ESP + 8]
***     addsd %XMM2, %XMM1
***     movsd QWORD PTR [%ESP + 8], %XMM2
        jmp .BBmain_1   # no_exit.i7

This is a bugpoint reduced testcase, which is why the testcase doesn't make
much sense (e.g. its an infinite loop). :)

//===---------------------------------------------------------------------===//

SSE should implement 'select_cc' using 'emulated conditional moves' that use
pcmp/pand/pandn/por to do a selection instead of a conditional branch:

double %X(double %Y, double %Z, double %A, double %B) {
        %C = setlt double %A, %B
        %z = add double %Z, 0.0    ;; select operand is not a load
        %D = select bool %C, double %Y, double %z
        ret double %D
}

We currently emit:

_X:
        subl $12, %esp
        xorpd %xmm0, %xmm0
        addsd 24(%esp), %xmm0
        movsd 32(%esp), %xmm1
        movsd 16(%esp), %xmm2
        ucomisd 40(%esp), %xmm1
        jb LBB_X_2
LBB_X_1:
        movsd %xmm0, %xmm2
LBB_X_2:
        movsd %xmm2, (%esp)
        fldl (%esp)
        addl $12, %esp
        ret

//===---------------------------------------------------------------------===//

It's not clear whether we should use pxor or xorps / xorpd to clear XMM
registers. The choice may depend on subtarget information. We should do some
more experiments on different x86 machines.

//===---------------------------------------------------------------------===//

Lower memcpy / memset to a series of SSE 128 bit move instructions when it's
feasible.

//===---------------------------------------------------------------------===//

Codegen:
  if (copysign(1.0, x) == copysign(1.0, y))
into:
  if (x^y & mask)
when using SSE.

//===---------------------------------------------------------------------===//

Use movhps to update upper 64-bits of a v4sf value. Also movlps on lower half
of a v4sf value.

//===---------------------------------------------------------------------===//

Better codegen for vector_shuffles like this { x, 0, 0, 0 } or { x, 0, x, 0}.
Perhaps use pxor / xorp* to clear a XMM register first?

//===---------------------------------------------------------------------===//

How to decide when to use the "floating point version" of logical ops? Here are
some code fragments:

	movaps LCPI5_5, %xmm2
	divps %xmm1, %xmm2
	mulps %xmm2, %xmm3
	mulps 8656(%ecx), %xmm3
	addps 8672(%ecx), %xmm3
	andps LCPI5_6, %xmm2
	andps LCPI5_1, %xmm3
	por %xmm2, %xmm3
	movdqa %xmm3, (%edi)

	movaps LCPI5_5, %xmm1
	divps %xmm0, %xmm1
	mulps %xmm1, %xmm3
	mulps 8656(%ecx), %xmm3
	addps 8672(%ecx), %xmm3
	andps LCPI5_6, %xmm1
	andps LCPI5_1, %xmm3
	orps %xmm1, %xmm3
	movaps %xmm3, 112(%esp)
	movaps %xmm3, (%ebx)

Due to some minor source change, the later case ended up using orps and movaps
instead of por and movdqa. Does it matter?

//===---------------------------------------------------------------------===//

X86RegisterInfo::copyRegToReg() returns X86::MOVAPSrr for VR128. Is it possible
to choose between movaps, movapd, and movdqa based on types of source and
destination?

How about andps, andpd, and pand? Do we really care about the type of the packed
elements? If not, why not always use the "ps" variants which are likely to be
shorter.

//===---------------------------------------------------------------------===//

External test Nurbs exposed some problems. Look for
__ZN15Nurbs_SSE_Cubic17TessellateSurfaceE, bb cond_next140. This is what icc
emits:

        movaps    (%edx), %xmm2                                 #59.21
        movaps    (%edx), %xmm5                                 #60.21
        movaps    (%edx), %xmm4                                 #61.21
        movaps    (%edx), %xmm3                                 #62.21
        movl      40(%ecx), %ebp                                #69.49
        shufps    $0, %xmm2, %xmm5                              #60.21
        movl      100(%esp), %ebx                               #69.20
        movl      (%ebx), %edi                                  #69.20
        imull     %ebp, %edi                                    #69.49
        addl      (%eax), %edi                                  #70.33
        shufps    $85, %xmm2, %xmm4                             #61.21
        shufps    $170, %xmm2, %xmm3                            #62.21
        shufps    $255, %xmm2, %xmm2                            #63.21
        lea       (%ebp,%ebp,2), %ebx                           #69.49
        negl      %ebx                                          #69.49
        lea       -3(%edi,%ebx), %ebx                           #70.33
        shll      $4, %ebx                                      #68.37
        addl      32(%ecx), %ebx                                #68.37
        testb     $15, %bl                                      #91.13
        jne       L_B1.24       # Prob 5%                       #91.13

This is the llvm code after instruction scheduling:

cond_next140 (0xa910740, LLVM BB @0xa90beb0):
	%reg1078 = MOV32ri -3
	%reg1079 = ADD32rm %reg1078, %reg1068, 1, %NOREG, 0
	%reg1037 = MOV32rm %reg1024, 1, %NOREG, 40
	%reg1080 = IMUL32rr %reg1079, %reg1037
	%reg1081 = MOV32rm %reg1058, 1, %NOREG, 0
	%reg1038 = LEA32r %reg1081, 1, %reg1080, -3
	%reg1036 = MOV32rm %reg1024, 1, %NOREG, 32
	%reg1082 = SHL32ri %reg1038, 4
	%reg1039 = ADD32rr %reg1036, %reg1082
	%reg1083 = MOVAPSrm %reg1059, 1, %NOREG, 0
	%reg1034 = SHUFPSrr %reg1083, %reg1083, 170
	%reg1032 = SHUFPSrr %reg1083, %reg1083, 0
	%reg1035 = SHUFPSrr %reg1083, %reg1083, 255
	%reg1033 = SHUFPSrr %reg1083, %reg1083, 85
	%reg1040 = MOV32rr %reg1039
	%reg1084 = AND32ri8 %reg1039, 15
	CMP32ri8 %reg1084, 0
	JE mbb<cond_next204,0xa914d30>

Still ok. After register allocation:

cond_next140 (0xa910740, LLVM BB @0xa90beb0):
	%EAX = MOV32ri -3
	%EDX = MOV32rm <fi#3>, 1, %NOREG, 0
	ADD32rm %EAX<def&use>, %EDX, 1, %NOREG, 0
	%EDX = MOV32rm <fi#7>, 1, %NOREG, 0
	%EDX = MOV32rm %EDX, 1, %NOREG, 40
	IMUL32rr %EAX<def&use>, %EDX
	%ESI = MOV32rm <fi#5>, 1, %NOREG, 0
	%ESI = MOV32rm %ESI, 1, %NOREG, 0
	MOV32mr <fi#4>, 1, %NOREG, 0, %ESI
	%EAX = LEA32r %ESI, 1, %EAX, -3
	%ESI = MOV32rm <fi#7>, 1, %NOREG, 0
	%ESI = MOV32rm %ESI, 1, %NOREG, 32
	%EDI = MOV32rr %EAX
	SHL32ri %EDI<def&use>, 4
	ADD32rr %EDI<def&use>, %ESI
	%XMM0 = MOVAPSrm %ECX, 1, %NOREG, 0
	%XMM1 = MOVAPSrr %XMM0
	SHUFPSrr %XMM1<def&use>, %XMM1, 170
	%XMM2 = MOVAPSrr %XMM0
	SHUFPSrr %XMM2<def&use>, %XMM2, 0
	%XMM3 = MOVAPSrr %XMM0
	SHUFPSrr %XMM3<def&use>, %XMM3, 255
	SHUFPSrr %XMM0<def&use>, %XMM0, 85
	%EBX = MOV32rr %EDI
	AND32ri8 %EBX<def&use>, 15
	CMP32ri8 %EBX, 0
	JE mbb<cond_next204,0xa914d30>

This looks really bad. The problem is shufps is a destructive opcode. Since it
appears as operand two in more than one shufps ops. It resulted in a number of
copies. Note icc also suffers from the same problem. Either the instruction
selector should select pshufd or The register allocator can made the two-address
to three-address transformation.

It also exposes some other problems. See MOV32ri -3 and the spills.

//===---------------------------------------------------------------------===//

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25500

LLVM is producing bad code.

LBB_main_4:	# cond_true44
	addps %xmm1, %xmm2
	subps %xmm3, %xmm2
	movaps (%ecx), %xmm4
	movaps %xmm2, %xmm1
	addps %xmm4, %xmm1
	addl $16, %ecx
	incl %edx
	cmpl $262144, %edx
	movaps %xmm3, %xmm2
	movaps %xmm4, %xmm3
	jne LBB_main_4	# cond_true44

There are two problems. 1) No need to two loop induction variables. We can
compare against 262144 * 16. 2) Known register coalescer issue. We should
be able eliminate one of the movaps:

	addps %xmm2, %xmm1    <=== Commute!
	subps %xmm3, %xmm1
	movaps (%ecx), %xmm4
	movaps %xmm1, %xmm1   <=== Eliminate!
	addps %xmm4, %xmm1
	addl $16, %ecx
	incl %edx
	cmpl $262144, %edx
	movaps %xmm3, %xmm2
	movaps %xmm4, %xmm3
	jne LBB_main_4	# cond_true44

//===---------------------------------------------------------------------===//

Consider:

__m128 test(float a) {
  return _mm_set_ps(0.0, 0.0, 0.0, a*a);
}

This compiles into:

movss 4(%esp), %xmm1
mulss %xmm1, %xmm1
xorps %xmm0, %xmm0
movss %xmm1, %xmm0
ret

Because mulss doesn't modify the top 3 elements, the top elements of 
xmm1 are already zero'd.  We could compile this to:

movss 4(%esp), %xmm0
mulss %xmm0, %xmm0
ret

//===---------------------------------------------------------------------===//

Here's a sick and twisted idea.  Consider code like this:

__m128 test(__m128 a) {
  float b = *(float*)&A;
  ...
  return _mm_set_ps(0.0, 0.0, 0.0, b);
}

This might compile to this code:

movaps c(%esp), %xmm1
xorps %xmm0, %xmm0
movss %xmm1, %xmm0
ret

Now consider if the ... code caused xmm1 to get spilled.  This might produce
this code:

movaps c(%esp), %xmm1
movaps %xmm1, c2(%esp)
...

xorps %xmm0, %xmm0
movaps c2(%esp), %xmm1
movss %xmm1, %xmm0
ret

However, since the reload is only used by these instructions, we could 
"fold" it into the uses, producing something like this:

movaps c(%esp), %xmm1
movaps %xmm1, c2(%esp)
...

movss c2(%esp), %xmm0
ret

... saving two instructions.

The basic idea is that a reload from a spill slot, can, if only one 4-byte 
chunk is used, bring in 3 zeros the the one element instead of 4 elements.
This can be used to simplify a variety of shuffle operations, where the
elements are fixed zeros.

//===---------------------------------------------------------------------===//

For this:

#include <emmintrin.h>
void test(__m128d *r, __m128d *A, double B) {
  *r = _mm_loadl_pd(*A, &B);
}

We generates:

	subl $12, %esp
	movsd 24(%esp), %xmm0
	movsd %xmm0, (%esp)
	movl 20(%esp), %eax
	movapd (%eax), %xmm0
	movlpd (%esp), %xmm0
	movl 16(%esp), %eax
	movapd %xmm0, (%eax)
	addl $12, %esp
	ret

icc generates:

        movl      4(%esp), %edx                                 #3.6
        movl      8(%esp), %eax                                 #3.6
        movapd    (%eax), %xmm0                                 #4.22
        movlpd    12(%esp), %xmm0                               #4.8
        movapd    %xmm0, (%edx)                                 #4.3
        ret                                                     #5.1

So icc is smart enough to know that B is in memory so it doesn't load it and
store it back to stack.

This should be fixed by eliminating the llvm.x86.sse2.loadl.pd intrinsic, 
lowering it to a load+insertelement instead.  Already match the load+shuffle 
as movlpd, so this should be easy.  We already get optimal code for:

define void @test2(<2 x double>* %r, <2 x double>* %A, double %B) {
entry:
	%tmp2 = load <2 x double>* %A, align 16
	%tmp8 = insertelement <2 x double> %tmp2, double %B, i32 0
	store <2 x double> %tmp8, <2 x double>* %r, align 16
	ret void
}

//===---------------------------------------------------------------------===//

__m128d test1( __m128d A, __m128d B) {
  return _mm_shuffle_pd(A, B, 0x3);
}

compiles to

shufpd $3, %xmm1, %xmm0

Perhaps it's better to use unpckhpd instead?

unpckhpd %xmm1, %xmm0

Don't know if unpckhpd is faster. But it is shorter.

//===---------------------------------------------------------------------===//

This code generates ugly code, probably due to costs being off or something:

define void @test(float* %P, <4 x float>* %P2 ) {
        %xFloat0.688 = load float* %P
        %tmp = load <4 x float>* %P2
        %inFloat3.713 = insertelement <4 x float> %tmp, float 0.0, i32 3
        store <4 x float> %inFloat3.713, <4 x float>* %P2
        ret void
}

Generates:

_test:
	movl	8(%esp), %eax
	movaps	(%eax), %xmm0
	pxor	%xmm1, %xmm1
	movaps	%xmm0, %xmm2
	shufps	$50, %xmm1, %xmm2
	shufps	$132, %xmm2, %xmm0
	movaps	%xmm0, (%eax)
	ret

Would it be better to generate:

_test:
        movl 8(%esp), %ecx
        movaps (%ecx), %xmm0
	xor %eax, %eax
        pinsrw $6, %eax, %xmm0
        pinsrw $7, %eax, %xmm0
        movaps %xmm0, (%ecx)
        ret

?

//===---------------------------------------------------------------------===//

Some useful information in the Apple Altivec / SSE Migration Guide:

http://developer.apple.com/documentation/Performance/Conceptual/
Accelerate_sse_migration/index.html

e.g. SSE select using and, andnot, or. Various SSE compare translations.

//===---------------------------------------------------------------------===//

Add hooks to commute some CMPP operations.

//===---------------------------------------------------------------------===//

Apply the same transformation that merged four float into a single 128-bit load
to loads from constant pool.

//===---------------------------------------------------------------------===//

Floating point max / min are commutable when -enable-unsafe-fp-path is
specified. We should turn int_x86_sse_max_ss and X86ISD::FMIN etc. into other
nodes which are selected to max / min instructions that are marked commutable.

//===---------------------------------------------------------------------===//

We should compile this:
#include <xmmintrin.h>
typedef union {
  int i[4];
  float f[4];
  __m128 v;
} vector4_t;
void swizzle (const void *a, vector4_t * b, vector4_t * c) {
  b->v = _mm_loadl_pi (b->v, (__m64 *) a);
  c->v = _mm_loadl_pi (c->v, ((__m64 *) a) + 1);
}

to:

_swizzle:
        movl    4(%esp), %eax
        movl    8(%esp), %edx
        movl    12(%esp), %ecx
        movlps  (%eax), %xmm0
        movlps  %xmm0, (%edx)
        movlps  8(%eax), %xmm0
        movlps  %xmm0, (%ecx)
        ret

not:

swizzle:
        movl 8(%esp), %eax
        movaps (%eax), %xmm0
        movl 4(%esp), %ecx
        movlps (%ecx), %xmm0
        movaps %xmm0, (%eax)
        movl 12(%esp), %eax
        movaps (%eax), %xmm0
        movlps 8(%ecx), %xmm0
        movaps %xmm0, (%eax)
        ret

//===---------------------------------------------------------------------===//

These functions should produce the same code:

#include <emmintrin.h>

typedef long long __m128i __attribute__ ((__vector_size__ (16)));

int foo(__m128i* val) {
  return __builtin_ia32_vec_ext_v4si(*val, 1);
}
int bar(__m128i* val) {
  union vs {
    __m128i *_v;
    int* _s;
  } v = {val};
  return v._s[1];
}

We currently produce (with -m64):

_foo:
        pshufd $1, (%rdi), %xmm0
        movd %xmm0, %eax
        ret
_bar:
        movl 4(%rdi), %eax
        ret

//===---------------------------------------------------------------------===//

We should materialize vector constants like "all ones" and "signbit" with 
code like:

     cmpeqps xmm1, xmm1   ; xmm1 = all-ones

and:
     cmpeqps xmm1, xmm1   ; xmm1 = all-ones
     psrlq   xmm1, 31     ; xmm1 = all 100000000000...

instead of using a load from the constant pool.  The later is important for
ABS/NEG/copysign etc.

//===---------------------------------------------------------------------===//

"converting 64-bit constant pool entry to 32-bit not necessarily beneficial"
http://llvm.org/PR1264

For this test case:

define double @foo(double %x) {
        %y = mul double %x, 5.000000e-01
        ret double %y
}

llc -march=x86-64 currently produces a 32-bit constant pool entry and this code:

        cvtss2sd .LCPI1_0(%rip), %xmm1
        mulsd %xmm1, %xmm0

instead of just using a 64-bit constant pool entry with this:

        mulsd .LCPI1_0(%rip), %xmm0

This is due to the code in ExpandConstantFP in LegalizeDAG.cpp. It notices that
x86-64 indeed has an instruction to load a 32-bit float from memory and convert
it into a 64-bit float in a register, however it doesn't notice that this isn't
beneficial because it prevents the load from being folded into the multiply.

//===---------------------------------------------------------------------===//

These functions:

#include <xmmintrin.h>
__m128i a;
void x(unsigned short n) {
  a = _mm_slli_epi32 (a, n);
}
void y(unsigned n) {
  a = _mm_slli_epi32 (a, n);
}

compile to ( -O3 -static -fomit-frame-pointer):
_x:
        movzwl  4(%esp), %eax
        movd    %eax, %xmm0
        movaps  _a, %xmm1
        pslld   %xmm0, %xmm1
        movaps  %xmm1, _a
        ret
_y:
        movd    4(%esp), %xmm0
        movaps  _a, %xmm1
        pslld   %xmm0, %xmm1
        movaps  %xmm1, _a
        ret

"y" looks good, but "x" does silly movzwl stuff around into a GPR.  It seems
like movd would be sufficient in both cases as the value is already zero 
extended in the 32-bit stack slot IIRC.  For signed short, it should also be
save, as a really-signed value would be undefined for pslld.


//===---------------------------------------------------------------------===//

#include <math.h>
int t1(double d) { return signbit(d); }

This currently compiles to:
	subl	$12, %esp
	movsd	16(%esp), %xmm0
	movsd	%xmm0, (%esp)
	movl	4(%esp), %eax
	shrl	$31, %eax
	addl	$12, %esp
	ret

We should use movmskp{s|d} instead.

//===---------------------------------------------------------------------===//

CodeGen/X86/vec_align.ll tests whether we can turn 4 scalar loads into a single
(aligned) vector load.  This functionality has a couple of problems.

1. The code to infer alignment from loads of globals is in the X86 backend,
   not the dag combiner.  This is because dagcombine2 needs to be able to see
   through the X86ISD::Wrapper node, which DAGCombine can't really do.
2. The code for turning 4 x load into a single vector load is target 
   independent and should be moved to the dag combiner.
3. The code for turning 4 x load into a vector load can only handle a direct 
   load from a global or a direct load from the stack.  It should be generalized
   to handle any load from P, P+4, P+8, P+12, where P can be anything.
4. The alignment inference code cannot handle loads from globals in non-static
   mode because it doesn't look through the extra dyld stub load.  If you try
   vec_align.ll without -relocation-model=static, you'll see what I mean.

//===---------------------------------------------------------------------===//

We should lower store(fneg(load p), q) into an integer load+xor+store, which
eliminates a constant pool load.  For example, consider:

define i64 @ccosf(float %z.0, float %z.1) nounwind readonly  {
entry:
	%tmp6 = sub float -0.000000e+00, %z.1		; <float> [#uses=1]
	%tmp20 = tail call i64 @ccoshf( float %tmp6, float %z.0 ) nounwind readonly 		; <i64> [#uses=1]
	ret i64 %tmp20
}

This currently compiles to:

LCPI1_0:					#  <4 x float>
	.long	2147483648	# float -0
	.long	2147483648	# float -0
	.long	2147483648	# float -0
	.long	2147483648	# float -0
_ccosf:
	subl	$12, %esp
	movss	16(%esp), %xmm0
	movss	%xmm0, 4(%esp)
	movss	20(%esp), %xmm0
	xorps	LCPI1_0, %xmm0
	movss	%xmm0, (%esp)
	call	L_ccoshf$stub
	addl	$12, %esp
	ret

Note the load into xmm0, then xor (to negate), then store.  In PIC mode,
this code computes the pic base and does two loads to do the constant pool 
load, so the improvement is much bigger.

The tricky part about this xform is that the argument load/store isn't exposed
until post-legalize, and at that point, the fneg has been custom expanded into 
an X86 fxor.  This means that we need to handle this case in the x86 backend
instead of in target independent code.

//===---------------------------------------------------------------------===//

Non-SSE4 insert into 16 x i8 is atrociously bad.

//===---------------------------------------------------------------------===//

<2 x i64> extract is substantially worse than <2 x f64>, even if the destination
is memory.

//===---------------------------------------------------------------------===//

SSE4 extract-to-mem ops aren't being pattern matched because of the AssertZext
sitting between the truncate and the extract.

//===---------------------------------------------------------------------===//

INSERTPS can match any insert (extract, imm1), imm2 for 4 x float, and insert
any number of 0.0 simultaneously.  Currently we only use it for simple
insertions.

See comments in LowerINSERT_VECTOR_ELT_SSE4.

//===---------------------------------------------------------------------===//

On a random note, SSE2 should declare insert/extract of 2 x f64 as legal, not
Custom.  All combinations of insert/extract reg-reg, reg-mem, and mem-reg are
legal, it'll just take a few extra patterns written in the .td file.

Note: this is not a code quality issue; the custom lowered code happens to be
right, but we shouldn't have to custom lower anything.  This is probably related
to <2 x i64> ops being so bad.