7e439874cb
Lower v8i16 multiply into this code:
...
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
2006-04-18 03:43:48 +00:00
a2cae1bb10
Custom lower v4i32 multiplies into a cute sequence, instead of having legalize
...
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
2006-04-18 03:24:30 +00:00
e54133cfba
Make sure to check splats of every constant we can, handle splat(31) by
...
being a bit more clever, add support for odd splats from -31 to -17.
llvm-svn: 27764
2006-04-17 18:09:22 +00:00
264c908e3a
Teach the ppc backend to use rol and vsldoi to generate splatted constants.
...
This implements vec_constants.ll:test_vsldoi and test_rol
llvm-svn: 27760
2006-04-17 17:55:10 +00:00
1b3806ace5
Make some code more general, adding support for constant formation of several
...
new patterns.
llvm-svn: 27754
2006-04-17 06:58:41 +00:00
f8dd76df5b
Learn how to make odd splatted constants in range [17,29]. This implements
...
PowerPC/vec_constants.ll:test_29.
llvm-svn: 27752
2006-04-17 06:07:44 +00:00
2a099c04c1
Pull some code out into a helper function.
...
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
2006-04-17 06:00:21 +00:00
071ad01ceb
Implement a TODO: for any shuffle that can be viewed as a v4[if]32 shuffle,
...
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
2006-04-17 05:28:54 +00:00
06a21ba96b
Implement a TODO: have the legalizer canonicalize a bunch of operations to
...
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
2006-04-16 01:37:57 +00:00
fa5aa396c2
Make the BUILD_VECTOR lowering code much more aggressive w.r.t constant vectors.
...
Remove some done items from the todo list.
llvm-svn: 27729
2006-04-16 01:01:29 +00:00
24acbe46c0
Fix a crash when faced with a shuffle vector that has an undef in its mask.
...
llvm-svn: 27726
2006-04-15 23:48:05 +00:00
559c8ba466
Allow undef in a shuffle mask
...
llvm-svn: 27714
2006-04-14 23:19:08 +00:00
4211ca9108
Move the rest of the PPCTargetLowering::LowerOperation cases out into
...
separate functions, for simplicity and code clarity.
llvm-svn: 27693
2006-04-14 06:01:58 +00:00
19e9055eb5
Pull the VECTOR_SHUFFLE and BUILD_VECTOR lowering code out into separate
...
functions, which makes the code much cleaner :)
llvm-svn: 27692
2006-04-14 05:19:18 +00:00
883fb053bd
Force non-darwin targets to use a static relo model. This fixes PR734,
...
tested by CodeGen/Generic/vector.ll
llvm-svn: 27657
2006-04-13 17:10:48 +00:00
147e50e1c5
Add a new way to match vector constants, which make it easier to bang bits of
...
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
2006-04-12 19:07:14 +00:00
74cf9ff761
Rename get_VSPLI_elt -> get_VSPLTI_elt
...
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
2006-04-12 17:37:20 +00:00
e318a7574e
Ensure that zero vectors are always v4i32, which forces them to CSE with
...
each other. This implements CodeGen/PowerPC/vxor-canonicalize.ll
llvm-svn: 27609
2006-04-12 16:53:28 +00:00
e4db08a2f1
Vector function results go into V2 according to GCC. The darwin ABI doc
...
doesn't say where they go :-/
llvm-svn: 27579
2006-04-11 01:38:39 +00:00
92533cfb4a
Move some return-handling code from lowerarguments to the ISD::RET handling stuff.
...
No functionality change.
llvm-svn: 27577
2006-04-11 01:21:43 +00:00
3a68f3c3ca
properly mark vector selects as expanded to select_cc
...
llvm-svn: 27544
2006-04-08 22:59:15 +00:00
0a3d1bbca4
Add VRRC select support
...
llvm-svn: 27543
2006-04-08 22:45:08 +00:00
d9e80f4516
Implement PowerPC/CodeGen/vec_splat.ll:spltish to use vsplish instead of a
...
constant pool load.
llvm-svn: 27538
2006-04-08 07:14:26 +00:00
d71a1f946d
Change the interface to the predicate that determines if vsplti* can be used.
...
No functionality changes.
llvm-svn: 27536
2006-04-08 06:46:53 +00:00
466841ddc7
Make sure to return the result in the right type.
...
llvm-svn: 27469
2006-04-06 23:12:19 +00:00
a4bbfaed5c
Match vpku[hw]um(x,x).
...
Convert vsldoi(x,x) to work the same way other (x,x) cases work.
llvm-svn: 27467
2006-04-06 22:28:36 +00:00
f38e033270
Add support for matching vmrg(x,x) patterns
...
llvm-svn: 27463
2006-04-06 22:02:42 +00:00
d1dcb52093
Pattern match vmrg* instructions, which are now lowered by the CFE into shuffles.
...
llvm-svn: 27457
2006-04-06 21:11:54 +00:00
1d33819194
Support pattern matching vsldoi(x,y) and vsldoi(x,x), which allows the f.e. to
...
lower it and LLVM to have one fewer intrinsic. This implements
CodeGen/PowerPC/vec_shuffle.ll
llvm-svn: 27450
2006-04-06 18:26:28 +00:00
e8b83b4206
Compile the vpkuhum/vpkuwum intrinsics into vpkuhum/vpkuwum instead of into
...
vperm with a perm mask lvx'd from the constant pool.
llvm-svn: 27448
2006-04-06 17:23:16 +00:00
39cc717c65
Fix CodeGen/PowerPC/2006-04-05-splat-ish.ll
...
llvm-svn: 27439
2006-04-05 17:39:25 +00:00
2cf4232ced
Fallthrough to expand if a VECTOR_SHUFFLE cannot be custom lowered.
...
llvm-svn: 27433
2006-04-05 06:09:26 +00:00
4a744e5c9d
Fix some broken logic that would cause us to codegen {2147483647,2147483647,2147483647,2147483647} as 'vspltisb v0, -1'.
...
llvm-svn: 27413
2006-04-04 22:28:35 +00:00
95c7adc7cb
Ask legalize to promote all vector shuffles to be v16i8 instead of having to
...
handle all 4 PPC vector types. This simplifies the matching code and allows
us to eliminate a bunch of patterns. This also adds cases we were missing,
such as CodeGen/PowerPC/vec_splat.ll:splat_h.
llvm-svn: 27400
2006-04-04 17:25:31 +00:00
447a7968af
Revert accidentally committed hunks.
...
llvm-svn: 27386
2006-04-03 23:58:04 +00:00
533aed9a35
Make sure to mark unsupported SCALAR_TO_VECTOR operations as expand.
...
llvm-svn: 27385
2006-04-03 23:55:43 +00:00
c5287c0ece
Inform the dag combiner that the predicate compares only return a low bit.
...
llvm-svn: 27359
2006-04-02 06:26:07 +00:00
9b2d6e7886
Custom lower all BUILD_VECTOR's so that we can compile vec_splat_u8(8) into
...
"vspltisb v0, 8" instead of a constant pool load.
llvm-svn: 27335
2006-04-02 00:43:36 +00:00
baa73e0d91
Rearrange code a bit
...
llvm-svn: 27306
2006-03-31 19:52:36 +00:00
754b41c84b
Add, sub and shuffle are legal for all vector types
...
llvm-svn: 27305
2006-03-31 19:48:58 +00:00
829a061abf
note to self: *save* file, then check it in
...
llvm-svn: 27291
2006-03-31 06:04:53 +00:00
d4058a59d4
Implement an item from the readme, folding vcmp/vcmp. instructions with
...
identical instructions into a single instruction. For example, for:
void test(vector float *x, vector float *y, int *P) {
int v = vec_any_out(*x, *y);
*x = (vector float)vec_cmpb(*x, *y);
*P = v;
}
we now generate:
_test:
mfspr r2, 256
oris r6, r2, 49152
mtspr 256, r6
lvx v0, 0, r4
lvx v1, 0, r3
vcmpbfp. v0, v1, v0
mfcr r4, 2
stvx v0, 0, r3
rlwinm r3, r4, 27, 31, 31
xori r3, r3, 1
stw r3, 0(r5)
mtspr 256, r2
blr
instead of:
_test:
mfspr r2, 256
oris r6, r2, 57344
mtspr 256, r6
lvx v0, 0, r4
lvx v1, 0, r3
vcmpbfp. v2, v1, v0
mfcr r4, 2
*** vcmpbfp v0, v1, v0
rlwinm r4, r4, 27, 31, 31
stvx v0, 0, r3
xori r3, r4, 1
stw r3, 0(r5)
mtspr 256, r2
blr
Testcase here: CodeGen/PowerPC/vcmp-fold.ll
llvm-svn: 27290
2006-03-31 06:02:07 +00:00
d7495ae7e9
Lower vector compares to VCMP nodes, just like we lower vector comparison
...
predicates to VCMPo nodes.
llvm-svn: 27285
2006-03-31 05:13:27 +00:00
bca5fbe914
Mark INSERT_VECTOR_ELT as expand
...
llvm-svn: 27276
2006-03-31 01:48:55 +00:00
1b3928765d
Add a few more altivec intrinsics
...
llvm-svn: 27215
2006-03-28 04:15:58 +00:00
cb5ec07cc3
Use normal lvx for scalar_to_vector instead of lve*x. They do the exact
...
same thing and we have a dag node for the former.
llvm-svn: 27205
2006-03-28 01:43:22 +00:00
e55d171ccd
Tblgen doesn't like multiple SDNode<> definitions that map to the sameenum value. Split them into separate enums.
...
llvm-svn: 27201
2006-03-28 00:40:33 +00:00
ed728c1291
SelectionDAGISel can now natively handle Switch instructions, in the same
...
manner that the LowerSwitch LLVM to LLVM pass does: emitting a binary
search tree of basic blocks. The new approach has several advantages:
it is faster, it generates significantly smaller code in many cases, and
it paves the way for implementing dense switch tables as a jump table by
handling switches directly in the instruction selector.
This functionality is currently only enabled on x86, but should be safe for
every target. In anticipation of making it the default, the cfg is now
properly updated in the x86, ppc, and sparc select lowering code.
llvm-svn: 27156
2006-03-27 01:32:24 +00:00
6961fc76bb
Codegen vector predicate compares.
...
llvm-svn: 27151
2006-03-26 10:06:40 +00:00
b1ddc988af
Remove PPC:isZeroVector, use ISD::isBuildVectorAllZeros instead
...
llvm-svn: 27149
2006-03-26 09:52:32 +00:00
Chris Lattner