If an altivec predicate compare is used immediately by a branch, don't
use a (serializing) MFCR instruction to read the CR6 register, which requires
a compare to get it back to CR's.  Instead, just branch on CR6 directly. :)

For example, for:
void foo2(vector float *A, vector float *B) {
  if (!vec_any_eq(*A, *B))
    *B = (vector float){0,0,0,0};
}

We now generate:

_foo2:
        mfspr r2, 256
        oris r5, r2, 12288
        mtspr 256, r5
        lvx v2, 0, r4
        lvx v3, 0, r3
        vcmpeqfp. v2, v3, v2
        bne cr6, LBB1_2 ; UnifiedReturnBlock
LBB1_1: ; cond_true
        vxor v2, v2, v2
        stvx v2, 0, r4
        mtspr 256, r2
        blr
LBB1_2: ; UnifiedReturnBlock
        mtspr 256, r2
        blr

instead of:

_foo2:
        mfspr r2, 256
        oris r5, r2, 12288
        mtspr 256, r5
        lvx v2, 0, r4
        lvx v3, 0, r3
        vcmpeqfp. v2, v3, v2
        mfcr r3, 2
        rlwinm r3, r3, 27, 31, 31
        cmpwi cr0, r3, 0
        beq cr0, LBB1_2 ; UnifiedReturnBlock
LBB1_1: ; cond_true
        vxor v2, v2, v2
        stvx v2, 0, r4
        mtspr 256, r2
        blr
LBB1_2: ; UnifiedReturnBlock
        mtspr 256, r2
        blr

This implements CodeGen/PowerPC/vec_br_cmp.ll.

llvm-svn: 27804

llvm-svn: 27802

even/odd halves.  Thanks to Nate telling me what's what.

llvm-svn: 27793

        vmuloub v5, v3, v2
        vmuleub v2, v3, v2
        vperm v2, v2, v5, v4

This implements CodeGen/PowerPC/vec_mul.ll.  With this, v16i8 multiplies are
6.79x faster than before.

Overall, UnitTests/Vector/multiplies.c is now 2.45x faster with LLVM than with
GCC.

Remove the 'integer multiplies' todo from the README file.

llvm-svn: 27792

        li r5, lo16(LCPI1_0)
        lis r6, ha16(LCPI1_0)
        lvx v4, r6, r5
        vmulouh v5, v3, v2
        vmuleuh v2, v3, v2
        vperm v2, v2, v5, v4

where v4 is:
LCPI1_0:                                        ;  <16 x ubyte>
        .byte   2
        .byte   3
        .byte   18
        .byte   19
        .byte   6
        .byte   7
        .byte   22
        .byte   23
        .byte   10
        .byte   11
        .byte   26
        .byte   27
        .byte   14
        .byte   15
        .byte   30
        .byte   31

This is 5.07x faster on the G5 (measured) than lowering to scalar code +
loads/stores.

llvm-svn: 27789

scalarize the sequence into 4 mullw's and a bunch of load/store traffic.

This speeds up v4i32 multiplies 4.1x (measured) on a G5.  This implements
PowerPC/vec_mul.ll

llvm-svn: 27788

llvm-svn: 27778

allows us to codegen functions as:

_test_rol:
        vspltisw v2, -12
        vrlw v2, v2, v2
        blr

instead of:

_test_rol:
        mfvrsave r2, 256
        mr r3, r2
        mtvrsave r3
        vspltisw v2, -12
        vrlw v2, v2, v2
        mtvrsave r2
        blr

Testcase here: CodeGen/PowerPC/vec_vrsave.ll

llvm-svn: 27777

the vrsave register for the caller.  This allows us to codegen a function as:

_test_rol:
        mfspr r2, 256
        mr r3, r2
        mtspr 256, r3
        vspltisw v2, -12
        vrlw v2, v2, v2
        mtspr 256, r2
        blr

instead of:

_test_rol:
        mfspr r2, 256
        oris r3, r2, 40960
        mtspr 256, r3
        vspltisw v0, -12
        vrlw v2, v0, v0
        mtspr 256, r2
        blr

llvm-svn: 27772

        vspltisw v2, -12
        vrlw v2, v2, v2

instead of:

        vspltisw v0, -12
        vrlw v2, v0, v0

when a function is returning a value.

llvm-svn: 27771

llvm-svn: 27770

llvm-svn: 27769

being a bit more clever, add support for odd splats from -31 to -17.

llvm-svn: 27764

This implements vec_constants.ll:test_vsldoi and test_rol

llvm-svn: 27760

llvm-svn: 27758

new patterns.

llvm-svn: 27754

PowerPC/vec_constants.ll:test_29.

llvm-svn: 27752

Effeciently codegen even splats in the range [-32,30].

This allows us to codegen <30,30,30,30> as:

        vspltisw v0, 15
        vadduwm v2, v0, v0

instead of as a cp load.

llvm-svn: 27750

if it can be implemented in 3 or fewer discrete altivec instructions, codegen
it as such.  This implements Regression/CodeGen/PowerPC/vec_perf_shuffle.ll

llvm-svn: 27748

llvm-svn: 27746

llvm-svn: 27744

llvm-svn: 27742

of various 4-element vectors.

llvm-svn: 27739

one type (v4i32) so that we don't have to write patterns for each type, and
so that more CSE opportunities are exposed.

llvm-svn: 27731

Remove some done items from the todo list.

llvm-svn: 27729

llvm-svn: 27726

go away when I start using evan's binop type canonicalizer

llvm-svn: 27725

llvm-svn: 27714

separate functions, for simplicity and code clarity.

llvm-svn: 27693

functions, which makes the code much cleaner :)

llvm-svn: 27692

tested by CodeGen/Generic/vector.ll

llvm-svn: 27657

llvm-svn: 27654

llvm-svn: 27651

different types.

Codegen spltw(0x7FFFFFFF) and spltw(0x80000000) without a constant pool load,
implementing PowerPC/vec_constants.ll:test1.  This compiles:

typedef float vf __attribute__ ((vector_size (16)));
typedef int vi __attribute__ ((vector_size (16)));
void test(vi *P1, vi *P2, vf *P3) {
  *P1 &= (vi){0x80000000,0x80000000,0x80000000,0x80000000};
  *P2 &= (vi){0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF};
  *P3 = vec_abs((vector float)*P3);
}

to:

_test:
        mfspr r2, 256
        oris r6, r2, 49152
        mtspr 256, r6
        vspltisw v0, -1
        vslw v0, v0, v0
        lvx v1, 0, r3
        vand v1, v1, v0
        stvx v1, 0, r3
        lvx v1, 0, r4
        vandc v1, v1, v0
        stvx v1, 0, r4
        lvx v1, 0, r5
        vandc v0, v1, v0
        stvx v0, 0, r5
        mtspr 256, r2
        blr

instead of (with two constant pool entries):

_test:
        mfspr r2, 256
        oris r6, r2, 49152
        mtspr 256, r6
        li r6, lo16(LCPI1_0)
        lis r7, ha16(LCPI1_0)
        li r8, lo16(LCPI1_1)
        lis r9, ha16(LCPI1_1)
        lvx v0, r7, r6
        lvx v1, 0, r3
        vand v0, v1, v0
        stvx v0, 0, r3
        lvx v0, r9, r8
        lvx v1, 0, r4
        vand v1, v1, v0
        stvx v1, 0, r4
        lvx v1, 0, r5
        vand v0, v1, v0
        stvx v0, 0, r5
        mtspr 256, r2
        blr

GCC produces (with 2 cp entries):

_test:
        mfspr r0,256
        stw r0,-4(r1)
        oris r0,r0,0xc00c
        mtspr 256,r0
        lis r2,ha16(LC0)
        lis r9,ha16(LC1)
        la r2,lo16(LC0)(r2)
        lvx v0,0,r3
        lvx v1,0,r5
        la r9,lo16(LC1)(r9)
        lwz r12,-4(r1)
        lvx v12,0,r2
        lvx v13,0,r9
        vand v0,v0,v12
        stvx v0,0,r3
        vspltisw v0,-1
        vslw v12,v0,v0
        vandc v1,v1,v12
        stvx v1,0,r5
        lvx v0,0,r4
        vand v0,v0,v13
        stvx v0,0,r4
        mtspr 256,r12
        blr

llvm-svn: 27624

Canonicalize BUILD_VECTOR's that match VSPLTI's into a single type for each
form, eliminating a bunch of Pat patterns in the .td file and allowing us to
CSE stuff more aggressively.  This implements
PowerPC/buildvec_canonicalize.ll:VSPLTI

llvm-svn: 27614

each other.  This implements CodeGen/PowerPC/vxor-canonicalize.ll

llvm-svn: 27609

llvm-svn: 27594

llvm-svn: 27593

llvm-svn: 27592

llvm-svn: 27588