1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-30 07:22:55 +01:00
Commit Graph

9897 Commits

Author SHA1 Message Date
Nate Begeman
ce63e383b8 Add a couple missing transforms in getSetCC that were triggering assertions
in the PPC Pattern ISel

llvm-svn: 21297
2005-04-14 08:56:52 +00:00
Duraid Madina
eb4c15b42c we have zextloads, not sextloads!
llvm-svn: 21296
2005-04-14 08:37:32 +00:00
Nate Begeman
b707ec16b4 Add the necessary support to codegen condition register logical ops with
register allocated condition registers.  Make sure that the printed
  output is gas compatible.

llvm-svn: 21295
2005-04-14 03:20:38 +00:00
Nate Begeman
99a9840b56 Start allocating condition registers. Almost all explicit uses of CR0 are
now gone.  Next step is to get rid of the remaining ones and then start
allocating bools to CRs where appropriate.

llvm-svn: 21294
2005-04-13 23:15:44 +00:00
Nate Begeman
ae49d52006 Implement the fold shift X, zext(Y) -> shift X, Y at the target level,
where it is safe to do so.

llvm-svn: 21293
2005-04-13 22:14:14 +00:00
Nate Begeman
20b3399465 Disbale the broken fold of shift + sz[ext] for now
Move the transform for select (a < 0) ? b : 0 into the dag from ppc isel
Enable the dag to fold and (setcc, 1) -> setcc for targets where setcc
  always produces zero or one.

llvm-svn: 21291
2005-04-13 21:23:31 +00:00
Chris Lattner
89f7e115a4 fix an infinite loop
llvm-svn: 21289
2005-04-13 20:06:29 +00:00
Chris Lattner
475fe85ddf fix some serious miscompiles on ia64, alpha, and ppc
llvm-svn: 21288
2005-04-13 19:53:40 +00:00
Chris Lattner
03d675414e avoid work when possible, perhaps fix the problem nate and andrew are seeing
with != 0 comparisons vanishing.

llvm-svn: 21287
2005-04-13 19:41:05 +00:00
Andrew Lenharth
cbf7a52768 WOW, function calls still seem to work after this.
llvm-svn: 21286
2005-04-13 17:17:28 +00:00
Andrew Lenharth
d1fee6d24a prepare for func call optimization
llvm-svn: 21285
2005-04-13 16:19:50 +00:00
Duraid Madina
b9d2d9ac63 * add the shladd instruction
* fold left shifts of 1, 2, 3 or 4 bits into adds

  This doesn't save much now, but should get a serious workout once
  multiplies by constants get converted to shift/add/sub sequences.
  Hold on! :)

llvm-svn: 21282
2005-04-13 06:12:04 +00:00
Andrew Lenharth
ec33ab6a2f add matches for SxADDL and company, as well as simplify the SxADDQ code
llvm-svn: 21281
2005-04-13 05:19:55 +00:00
Chris Lattner
9540cf8c7e Implement expansion of unsigned i64 -> FP.
Note that this probably only works for little endian targets, but is enough
to get siod working :)

llvm-svn: 21280
2005-04-13 05:09:42 +00:00
Duraid Madina
67e553e521 * if ANDing with a constant of the form:
0x00000..00FFF..FF
      ^      ^
      ^      ^
    any number of
    0's followed by
    some number of
    1's

    then we use dep.z to just paste zeros over the input. For the special
    cases where this is zxt1/zxt2/zxt4, we use those instructions instead,
    because we're all about readability!!!
    that's what it's about!! readability!

  *twitch* ;D

llvm-svn: 21279
2005-04-13 04:50:54 +00:00
Andrew Lenharth
510db15268 added all flavors of zap for anding
llvm-svn: 21276
2005-04-13 03:47:03 +00:00
Chris Lattner
1a6247ff51 Make expansion of uint->fp cast assert out instead of infinitely recurse.
llvm-svn: 21275
2005-04-13 03:42:14 +00:00
Chris Lattner
85c6a7bed0 Fix some mysteriously missing {}'s which cause the miscompilation of
Olden/mst, Ptrdist/bc, Obsequi, etc.

llvm-svn: 21274
2005-04-13 03:29:53 +00:00
Chris Lattner
63450e87d9 add back the optimization that Nate added for shl X, (zext_inreg y)
llvm-svn: 21273
2005-04-13 02:58:13 +00:00
Chris Lattner
759afe07d7 Oops, remove these too.
llvm-svn: 21272
2005-04-13 02:47:57 +00:00
Chris Lattner
8489ac991d remove one more occurance of this that snuck in
llvm-svn: 21271
2005-04-13 02:46:17 +00:00
Chris Lattner
5fdb103328 Remove support for ZERO_EXTEND_INREG. This pessimizes code, genering stuff
like this:

        ldah $1,1($31)
        lda $1,-1($1)
        and $0,$1,$24

instead of this:

        zap $0,252,$24

To get this back, the selector should recognize the ISD::AND case where this
happens and emit the appropriate ZAP instruction.

llvm-svn: 21270
2005-04-13 02:43:40 +00:00
Chris Lattner
a2e92e69da Remove special handling of ZERO_EXTEND_INREG. This pessimizes code, causing
things like this:

       mov r9 = 65535;;
       and r8 = r8, r9;;

To be emitted instead of:

        zxt2 r8 = r8;;

To get this back, the selector for ISD::AND should recognize this case.

llvm-svn: 21269
2005-04-13 02:41:52 +00:00
Chris Lattner
26c7c9150a Elimate handling of ZERO_EXTEND_INREG. This causes the PPC backend to emit
andi instructions instead of rlwinm instructions for zero extend, but they
seem like they would take the same time.

llvm-svn: 21268
2005-04-13 02:40:26 +00:00
Chris Lattner
f25fefd9cf Z_E_I is gone
llvm-svn: 21267
2005-04-13 02:39:05 +00:00
Chris Lattner
4f188f949c Instead of making ZERO_EXTEND_INREG nodes, use the helper method in
SelectionDAG to do the job with AND.  Don't legalize Z_E_I anymore as
it is gone

llvm-svn: 21266
2005-04-13 02:38:47 +00:00
Chris Lattner
bce0030a88 Remove all foldings of ZERO_EXTEND_INREG, moving them to work for AND nodes
instead.  OVerall, this increases the amount of folding we can do.

llvm-svn: 21265
2005-04-13 02:38:18 +00:00
Nate Begeman
38d8248a9e Fold shift x, [sz]ext(y) -> shift x, y
llvm-svn: 21262
2005-04-12 23:32:28 +00:00
Nate Begeman
a56527ea5f Fold shift by size larger than type size to undef
Make llvm undef values generate ISD::UNDEF nodes

llvm-svn: 21261
2005-04-12 23:12:17 +00:00
Nate Begeman
79c8b8fd1c Implement setcc op, -1 sequences
Remove dead setcc op, 0 sequences
Coming later: generalization of op, imm

llvm-svn: 21260
2005-04-12 21:22:28 +00:00
Chris Lattner
58f72ab722 promote extload i1 -> extload i8
llvm-svn: 21258
2005-04-12 20:30:10 +00:00
Chris Lattner
7c88662870 add an argument to allow avoiding deleting phi nodes.
llvm-svn: 21255
2005-04-12 18:52:14 +00:00
Chris Lattner
ee06161a63 Get rid of this for_each loop
llvm-svn: 21253
2005-04-12 18:51:33 +00:00
Duraid Madina
39fcec1541 * OK, after changing to use liveIn/liveOut instead of IDEFs,
to avoid redundant mov out3=r44 type instructions, we need to
tell the register allocator the truth about out? registers.

FIXME: unfortunately, since the list of allocatable registers is immutable,
we can't simply 'delete r127' from the allocation order, say, if 'out0' is
used. The only correct thing we can do is have a linear order of regs:

out7, out6 ... out2, out1, out0, r32, r33, r34 ... r126, r127

and slide a 'window' of 96 registers along this line, depending on how many
of the out? regs a function actually uses. The only downside of this is
that the out? registers will be allocated _first_, which makes the
resulting assembly ugly. :( Note this in the README. Hope this gets fixed
soon. :) (note the 3rd person speech there)

llvm-svn: 21252
2005-04-12 18:42:59 +00:00
Andrew Lenharth
174d44f223 Get rid of idefs for arguments (oops)
llvm-svn: 21251
2005-04-12 17:47:57 +00:00
Andrew Lenharth
1b8a8331c9 Get rid of idefs for arguments
llvm-svn: 21250
2005-04-12 17:35:16 +00:00
Chris Lattner
be2dceff48 Put out* into the allocation order, allowing the register allocator to
coallesce moves into outgoing args.

llvm-svn: 21249
2005-04-12 15:12:51 +00:00
Chris Lattner
37712352c0 Make sure to realize that calls use their argument regs
llvm-svn: 21248
2005-04-12 15:12:19 +00:00
Duraid Madina
2821f99f19 stop emitting IDEFs for args - change to using liveIn/liveOut
llvm-svn: 21247
2005-04-12 14:54:44 +00:00
Nate Begeman
f96b42f1b6 Initial support for allocation condition registers
llvm-svn: 21246
2005-04-12 07:04:16 +00:00
Chris Lattner
f8d9224d8c Fix a crash analyzing MultiSource/Benchmarks/MallocBench/gs
llvm-svn: 21245
2005-04-12 03:59:27 +00:00
Chris Lattner
cfc7093ca6 Remove some redundant checks, add a couple of new ones. This allows us to
compile this:

int foo (unsigned long a, unsigned long long g) {
  return a >= g;
}

To:

foo:
        movl 8(%esp), %eax
        cmpl %eax, 4(%esp)
        setae %al
        cmpl $0, 12(%esp)
        sete %cl
        andb %al, %cl
        movzbl %cl, %eax
        ret

instead of:

foo:
        movl 8(%esp), %eax
        cmpl %eax, 4(%esp)
        setae %al
        movzbw %al, %cx
        movl 12(%esp), %edx
        cmpl $0, %edx
        sete %al
        movzbw %al, %ax
        cmpl $0, %edx
        cmove %cx, %ax
        movzbl %al, %eax
        ret

llvm-svn: 21244
2005-04-12 02:54:39 +00:00
Chris Lattner
61f353dbdc Emit comparisons against the sign bit better. Codegen this:
bool %test1(long %X) {
        %A = setlt long %X, 0
        ret bool %A
}

like this:

test1:
        cmpl $0, 8(%esp)
        setl %al
        movzbl %al, %eax
        ret

instead of:

test1:
        movl 8(%esp), %ecx
        cmpl $0, %ecx
        setl %al
        movzbw %al, %ax
        cmpl $0, 4(%esp)
        setb %dl
        movzbw %dl, %dx
        cmpl $0, %ecx
        cmove %dx, %ax
        movzbl %al, %eax
        ret

llvm-svn: 21243
2005-04-12 02:19:10 +00:00
Chris Lattner
6cbbb55967 Emit long comparison against -1 better. Instead of this (x86):
test2:
        movl 8(%esp), %eax
        notl %eax
        movl 4(%esp), %ecx
        notl %ecx
        orl %eax, %ecx
        cmpl $0, %ecx
        sete %al
        movzbl %al, %eax
        ret

or this (PPC):

_test2:
        nor r2, r4, r4
        nor r3, r3, r3
        or r2, r2, r3
        cntlzw r2, r2
        srwi r3, r2, 5
        blr

Emit this:

test2:
        movl 8(%esp), %eax
        andl 4(%esp), %eax
        cmpl $-1, %eax
        sete %al
        movzbl %al, %eax
        ret

or this:

_test2:
.LBB_test2_0:   ;
        and r2, r4, r3
        cmpwi cr0, r2, -1
        li r3, 1
        li r2, 0
        beq .LBB_test2_2        ;
.LBB_test2_1:   ;
        or r3, r2, r2
.LBB_test2_2:   ;
        blr

it seems like the PPC isel could do better for R32 == -1 case.

llvm-svn: 21242
2005-04-12 01:46:05 +00:00
Chris Lattner
37534d43d0 canonicalize x <u 1 -> x == 0. On this testcase:
unsigned long long g;
unsigned long foo (unsigned long a) {
  return (a >= g) ? 1 : 0;
}

It changes the ppc code from:

_foo:
.LBB_foo_0:     ; entry
        mflr r11
        stw r11, 8(r1)
        bl "L00000$pb"
"L00000$pb":
        mflr r2
        addis r2, r2, ha16(L_g$non_lazy_ptr-"L00000$pb")
        lwz r2, lo16(L_g$non_lazy_ptr-"L00000$pb")(r2)
        lwz r4, 0(r2)
        lwz r2, 4(r2)
        cmplw cr0, r3, r2
        li r2, 1
        li r3, 0
        bge .LBB_foo_2  ; entry
.LBB_foo_1:     ; entry
        or r2, r3, r3
.LBB_foo_2:     ; entry
        cmplwi cr0, r4, 1
        li r3, 1
        li r5, 0
        blt .LBB_foo_4  ; entry
.LBB_foo_3:     ; entry
        or r3, r5, r5
.LBB_foo_4:     ; entry
        cmpwi cr0, r4, 0
        beq .LBB_foo_6  ; entry
.LBB_foo_5:     ; entry
        or r2, r3, r3
.LBB_foo_6:     ; entry
        rlwinm r3, r2, 0, 31, 31
        lwz r11, 8(r1)
        mtlr r11
        blr


to:

_foo:
.LBB_foo_0:     ; entry
        mflr r11
        stw r11, 8(r1)
        bl "L00000$pb"
"L00000$pb":
        mflr r2
        addis r2, r2, ha16(L_g$non_lazy_ptr-"L00000$pb")
        lwz r2, lo16(L_g$non_lazy_ptr-"L00000$pb")(r2)
        lwz r4, 0(r2)
        lwz r2, 4(r2)
        cmplw cr0, r3, r2
        li r2, 1
        li r3, 0
        bge .LBB_foo_2  ; entry
.LBB_foo_1:     ; entry
        or r2, r3, r3
.LBB_foo_2:     ; entry
        cntlzw r3, r4
        srwi r3, r3, 5
        cmpwi cr0, r4, 0
        beq .LBB_foo_4  ; entry
.LBB_foo_3:     ; entry
        or r2, r3, r3
.LBB_foo_4:     ; entry
        rlwinm r3, r2, 0, 31, 31
        lwz r11, 8(r1)
        mtlr r11
        blr

llvm-svn: 21241
2005-04-12 00:28:49 +00:00
Nate Begeman
a154deaaff Implement bitfield clears
Implement divide by negative power of two

llvm-svn: 21240
2005-04-12 00:10:02 +00:00
Nate Begeman
f31b58f145 Update PPC readme. Remove things that are done or aren't ppc specific
llvm-svn: 21232
2005-04-11 20:48:57 +00:00
Chris Lattner
7f0f0854fa Teach the dag mechanism that this:
long long test2(unsigned A, unsigned B) {
        return ((unsigned long long)A << 32) + B;
}

is equivalent to this:

long long test1(unsigned A, unsigned B) {
        return ((unsigned long long)A << 32) | B;
}

Now they are both codegen'd to this on ppc:

_test2:
        blr

or this on x86:

test2:
        movl 4(%esp), %edx
        movl 8(%esp), %eax
        ret

llvm-svn: 21231
2005-04-11 20:29:59 +00:00
Chris Lattner
71f3d4ce57 Fix expansion of shifts by exactly NVT bits on arch's (like X86) that have
masking shifts.

This fixes the miscompilation of this:

long long test1(unsigned A, unsigned B) {
        return ((unsigned long long)A << 32) | B;
}

into this:

test1:
        movl 4(%esp), %edx
        movl %edx, %eax
        orl 8(%esp), %eax
        ret

allowing us to generate this instead:

test1:
        movl 4(%esp), %edx
        movl 8(%esp), %eax
        ret

llvm-svn: 21230
2005-04-11 20:08:52 +00:00
Chris Lattner
55e620f08d IA64 supports this operation.
llvm-svn: 21228
2005-04-11 18:55:36 +00:00
Chris Lattner
ee715b2abc ORo sets CR0
llvm-svn: 21227
2005-04-11 15:03:48 +00:00
Chris Lattner
7d11f40ee2 Revert the previous patch, which I didn't mean to check in.
llvm-svn: 21226
2005-04-11 15:03:41 +00:00
Chris Lattner
d925c74452 Fix a minor bug (ORo didn't mark that it set CR0).
Refactor how . instructions are handled.  In particular, instead of passing
the RC flag all the way up the inheritance hierarchy, just make a new tblgen
class 'DOT' which can be added to an instruction definition.

For example, instead of this:

-def AND  : XForm_6<31,  28, 0, 0, 0, (ops GPRC:$rA, GPRC:$rS, GPRC:$rB),
-let Defs = [CR0] in
-def ANDo : XForm_6<31,  28, 1, 0, 0, (ops GPRC:$rA, GPRC:$rS, GPRC:$rB),
-                   "and. $rA, $rS, $rB">;

We now have this:

+def AND  : XForm_6<31,  28, 0, 0, (ops GPRC:$rA, GPRC:$rS, GPRC:$rB),
                    "and $rA, $rS, $rB">;

llvm-svn: 21225
2005-04-11 15:01:39 +00:00
Duraid Madina
01aaf77792 hmm, should probably change addImm() to take 64-bit arguments one day anyway.
llvm-svn: 21224
2005-04-11 07:16:39 +00:00
Nate Begeman
783fe2108e Add recording variants of ISD::AND and ISD::OR. This kills almost 1000
(1.5%) instructions in 186.crafty

llvm-svn: 21222
2005-04-11 06:34:10 +00:00
Duraid Madina
d2ae9221c7 assorted fixes:
* clean up immediates (we use 14, 22 and 64 bit immediates now. sane.)
  * fold r0/f0/f1 registers into comparisons against 0/0.0/1.0
  * fix nasty thinko - didn't use two-address form of conditional add
    for extending bools to integers, so occasionally there would be
    garbage in the result. it's amazing how often zeros are just
    sitting around in registers ;) - this should fix a bunch of tests.

llvm-svn: 21221
2005-04-11 05:55:56 +00:00
Jeff Cohen
1b89da675f Eliminate tabs
llvm-svn: 21216
2005-04-11 03:44:22 +00:00
Nate Begeman
32163963cb Fix libcall code to not pass a NULL Chain to LowerCallTo
Fix libcall code to not crash or assert looking for an ADJCALLSTACKUP node
  when it is known that there is no ADJCALLSTACKDOWN to match.
Expand i64 multiply when ISD::MULHU is legal for the target.

llvm-svn: 21214
2005-04-11 03:01:51 +00:00
Chris Lattner
4f26677dc9 Don't bother sign/zext_inreg'ing the result of an and operation if we know
the result does change as a result of the extend.

This improves codegen for Alpha on this testcase:

int %a(ushort* %i) {
        %tmp.1 = load ushort* %i
        %tmp.2 = cast ushort %tmp.1 to int
        %tmp.4 = and int %tmp.2, 1
        ret int %tmp.4
}

Generating:

a:
        ldgp $29, 0($27)
        ldwu $0,0($16)
        and $0,1,$0
        ret $31,($26),1

instead of:

a:
        ldgp $29, 0($27)
        ldwu $0,0($16)
        and $0,1,$0
        addl $0,0,$0
        ret $31,($26),1

btw, alpha really should switch to livein/outs for args :)

llvm-svn: 21213
2005-04-10 23:37:16 +00:00
Chris Lattner
c730ea00e2 Teach legalize to deal with targets that don't support some SEXTLOAD/ZEXTLOADs
llvm-svn: 21212
2005-04-10 22:54:25 +00:00
Chris Lattner
1b9e1e26cb don't zextload fp values!
llvm-svn: 21209
2005-04-10 17:40:35 +00:00
Nate Begeman
34aa7ec9cb Fix another fixme: factor out the constant fp generation code.
llvm-svn: 21207
2005-04-10 06:06:10 +00:00
Nate Begeman
b6c9b326e3 Fix 64 bit argument loading that straddles the args in regs / args on stack
boundary.

llvm-svn: 21206
2005-04-10 05:53:14 +00:00
Chris Lattner
0c089eae41 Until we have a dag combiner, promote using zextload's instead of extloads.
This gives the optimizer a bit of information about the top-part of the
value.

llvm-svn: 21205
2005-04-10 04:33:47 +00:00
Chris Lattner
9d13d0b958 Fold zext_inreg(zextload), likewise for sext's
llvm-svn: 21204
2005-04-10 04:33:08 +00:00
Chris Lattner
9c8fe594e5 add a simple xform
llvm-svn: 21203
2005-04-10 04:04:49 +00:00
Nate Begeman
0a43e95718 Remove unnecessary Implicit Defs. Since r0 is not in allocation, we do not
have to inform the register allocator it might be stepped on.

llvm-svn: 21202
2005-04-10 03:59:42 +00:00
Nate Begeman
f5cedbc812 Make sure that BRCOND branches can be converted into long branches too.
llvm-svn: 21198
2005-04-10 01:48:29 +00:00
Nate Begeman
ab8e705a52 Don't hand ISD::CALL nodes off to SelectExprFP. This fixes siod.
llvm-svn: 21197
2005-04-10 01:14:13 +00:00
Chris Lattner
b3518a838c Fix a thinko. If the operand is promoted, pass the promoted value into
the new zero extend, not the original operand.  This fixes cast bool -> long
on ppc.

Add an unrelated fixme

llvm-svn: 21196
2005-04-10 01:13:15 +00:00
Chris Lattner
c1bacbff9d rename getPPCOpcodeForSetCCNumber -> getPPCOpcodeForSetCCOpode to be more
correct.  Remove the EmitComparison retvalue, as it is always the first arg.

Fix a place where we incorrectly passed in the setcc opcode instead of the
setcc number, causing us to miscompile crafty.  Crafty now works!

llvm-svn: 21195
2005-04-10 01:03:31 +00:00
Nate Begeman
a2374d39df fix ISD::BRCONDTWOWAY codegen to not deference the end() iterator
llvm-svn: 21193
2005-04-09 23:35:05 +00:00
Chris Lattner
17c60891c1 Fix CodeGen/Generic/2005-05-09-GlobalInPHI.ll, which was reduced from 254.gap.
This caused the "use before a def" assertion on some programs.

With this patch, 254.gap now passes with the PPC backend.

llvm-svn: 21191
2005-04-09 22:05:17 +00:00
Chris Lattner
034716de24 add a little peephole optimization. This allows us to codegen:
int a(short i) {
        return i & 1;
}

as

_a:
        andi. r3, r3, 1
        blr

instead of:

_a:
        rlwinm r2, r3, 0, 16, 31
        andi. r3, r2, 1
        blr

on ppc.  It should also help the other risc targets.

llvm-svn: 21189
2005-04-09 21:43:54 +00:00
Chris Lattner
b630949c2e do not set the root to null if an argument is dead
llvm-svn: 21188
2005-04-09 21:23:24 +00:00
Nate Begeman
dda6155d19 Add rlwnm instruction for variable rotate
Generate rotate left/right immediate
Generate code for brcondtwoway
Use new livein/liveout functionality

llvm-svn: 21187
2005-04-09 20:09:12 +00:00
Chris Lattner
72b1964108 Fix a crash on 173.applu by asking for a constant bigger than 32-bits.
llvm-svn: 21185
2005-04-09 19:47:21 +00:00
Chris Lattner
c97f9f403f Switch this instruction selector over to using liveins and liveouts, eliminating
implicit defs on entry to the function.  yaay :)

llvm-svn: 21184
2005-04-09 16:32:30 +00:00
Chris Lattner
77ab286605 there is no need to remove this instruction, linscan does it already as it
removes noop moves.

llvm-svn: 21183
2005-04-09 16:24:20 +00:00
Chris Lattner
f408e9a07b Adjust live intervals to support a livein set
llvm-svn: 21182
2005-04-09 16:17:50 +00:00
Chris Lattner
fd4d6e0847 Use live out sets for return values instead of imp_defs, which is cleaner and faster.
llvm-svn: 21181
2005-04-09 15:23:56 +00:00
Chris Lattner
1a9c8fc64a Consider the livein/out set for a function, allowing targets to not have to
use ugly imp_def/imp_uses for arguments and return values.

llvm-svn: 21180
2005-04-09 15:23:25 +00:00
Duraid Madina
2d3e52576d ok, the "ia64 has a boatload of registers" joke stopped being funny today ;)
* fix overallocation of integer (stacked) registers: we can't allocate
  registers for local use if they are required as output registers

this fixes 'toast' in the test suite, and all sorts of larger programs
like bzip2 etc.

llvm-svn: 21178
2005-04-09 11:53:00 +00:00
Nate Begeman
98bcb13bfa Optimize FSEL a bit for fneg arguments. This fixes the recently added test
case so that we emit

_test_fneg_sel:
.LBB_test_fneg_sel_0:   ;
        fsel f1, f1, f3, f2
        blr

instead of:

_test_fneg_sel:
.LBB_test_fneg_sel_0:   ;
        fneg f0, f1
        fneg f0, f0
        fsel f1, f0, f3, f2
        blr

llvm-svn: 21177
2005-04-09 09:33:07 +00:00
Chris Lattner
3fde68ea7a Fix CodeGen/SparcV9/2005-05-09-GEP-Crash.ll a crash on some specfp program
lets hope this doesn't break other programs with induced entropy

llvm-svn: 21174
2005-04-09 06:27:14 +00:00
Chris Lattner
afa0001d54 recognize some patterns as fabs operations, so that fabs at the source level
is deconstructed then reconstructed here.  This catches 19 fabs's in 177.mesa
9 in 168.wupwise, 5 in 171.swim, 3 in 172.mgrid, and 14 in 173.applu out of
specfp2000.

This allows the X86 code generator to make MUCH better code than before for
each of these and saves one instr on ppc.

This depends on the previous CFE patch to expose these correctly.

llvm-svn: 21171
2005-04-09 05:15:53 +00:00
Chris Lattner
8e6eafa8e1 Emit BRCONDTWOWAY when possible.
llvm-svn: 21167
2005-04-09 03:30:29 +00:00
Chris Lattner
55b73bda6c Legalize BRCONDTWOWAY into a BRCOND/BR pair if a target doesn't support it.
llvm-svn: 21166
2005-04-09 03:30:19 +00:00
Chris Lattner
da902bdf1b print and fold BRCONDTWOWAY correctly
llvm-svn: 21165
2005-04-09 03:27:28 +00:00
Chris Lattner
39f963f968 This target does not support/want ISD::BRCONDTWOWAY
llvm-svn: 21164
2005-04-09 03:22:37 +00:00
Chris Lattner
c80baf5567 This target does not yet support ISD::BRCONDTWOWAY
llvm-svn: 21163
2005-04-09 03:22:30 +00:00
Nate Begeman
99fb6814bd 64b: Expand S/UREM
32b: No longer pattern match fneg(fsub(fmul)) as fnmsub
     Pattern match fsub a, mul(b, c) as fnmsub
     Pattern match fadd a, mul(b, c) as fmadd
Those changes speed up hydro2d by 2.5%, distray by 6%, and scimark by 8%

llvm-svn: 21161
2005-04-09 03:05:51 +00:00
Chris Lattner
31170cd2ec canonicalize a bunch of operations involving fneg
llvm-svn: 21160
2005-04-09 03:02:46 +00:00
Nate Begeman
95e1b860a1 Fix 64b shifts
llvm-svn: 21159
2005-04-08 23:45:01 +00:00
Nate Begeman
3fca499b8d Match Mac OS X 64 bit calling conventions
llvm-svn: 21157
2005-04-08 21:26:05 +00:00
Andrew Lenharth
6eff2083b5 collect a few statistics, factor constants (constant loading and mult), fix logic operation pattern matchs, supress FP div when int dividing by a constant
llvm-svn: 21156
2005-04-08 17:28:49 +00:00
Duraid Madina
e7412561bf fix bogus division-by-power-of-2 (was wrong for negative input, adds extr insn)
fix hack in division (clean up frcpa instruction)

llvm-svn: 21153
2005-04-08 10:01:48 +00:00
Chris Lattner
f82999fabe Fix bug: InstCombine/2005-05-07-UDivSelectCrash.ll
llvm-svn: 21152
2005-04-08 04:03:26 +00:00
Nate Begeman
6875356db1 Optimized code sequences for setcc reg, 0
Optimized code sequence for (a < 0) ? b : 0

llvm-svn: 21150
2005-04-07 20:30:01 +00:00
Andrew Lenharth
b1b7ef0979 Alpha zero extends setcc results
llvm-svn: 21149
2005-04-07 20:11:32 +00:00
Chris Lattner
9a56ef5693 If a target zero or sign extends the result of its setcc, allow folding of
this into sign/zero extension instructions later.

On PPC, for example, this testcase:

%G = external global sbyte
implementation
void %test(int %X, int %Y) {
  %C = setlt int %X, %Y
  %D = cast bool %C to sbyte
  store sbyte %D, sbyte* %G
  ret void
}

Now codegens to:

        cmpw cr0, r3, r4
        li r3, 1
        li r4, 0
        blt .LBB_test_2 ;
.LBB_test_1:    ;
        or r3, r4, r4
.LBB_test_2:    ;
        addis r2, r2, ha16(L_G$non_lazy_ptr-"L00000$pb")
        lwz r2, lo16(L_G$non_lazy_ptr-"L00000$pb")(r2)
        stb r3, 0(r2)

instead of:

        cmpw cr0, r3, r4
        li r3, 1
        li r4, 0
        blt .LBB_test_2 ;
.LBB_test_1:    ;
        or r3, r4, r4
.LBB_test_2:    ;
***     rlwinm r3, r3, 0, 31, 31
        addis r2, r2, ha16(L_G$non_lazy_ptr-"L00000$pb")
        lwz r2, lo16(L_G$non_lazy_ptr-"L00000$pb")(r2)
        stb r3, 0(r2)

llvm-svn: 21148
2005-04-07 19:43:53 +00:00
Chris Lattner
352dd3e579 PowerPC zero extends setcc results
llvm-svn: 21147
2005-04-07 19:41:49 +00:00
Chris Lattner
583dcc66f2 X86 zero extends setcc results
llvm-svn: 21146
2005-04-07 19:41:46 +00:00
Chris Lattner
bbe0e9e9db Remove somethign I had for testing
llvm-svn: 21144
2005-04-07 18:58:54 +00:00
Andrew Lenharth
c3e5c42b86 fix a small optimization opertunity and make gcc happy
llvm-svn: 21143
2005-04-07 18:15:28 +00:00
Chris Lattner
ee836c7b32 This patch does two things. First, it canonicalizes 'X >= C' -> 'X > C-1'
(likewise for <= >=u >=u).

Second, it implements a special case hack to turn 'X gtu SINTMAX' -> 'X lt 0'

On powerpc, for example, this changes this:

        lis r2, 32767
        ori r2, r2, 65535
        cmplw cr0, r3, r2
        bgt .LBB_test_2

into:

        cmpwi cr0, r3, 0
        blt .LBB_test_2

llvm-svn: 21142
2005-04-07 18:14:58 +00:00
Andrew Lenharth
469244a2b4 fixup magic constant making code. tested by thousands of random divisions.... by 10000. ok, so random divisors would be good too, but this at least fixes some things
llvm-svn: 21140
2005-04-07 17:19:16 +00:00
Andrew Lenharth
61215a78fb lowercase instructions, makes diff happier
llvm-svn: 21139
2005-04-07 17:17:48 +00:00
Chris Lattner
05d79f36b1 Implement the following xforms:
(X-Y)-X --> -Y
A + (B - A) --> B
(B - A) + A --> B

llvm-svn: 21138
2005-04-07 17:14:51 +00:00
Chris Lattner
5b31ada26d Implement InstCombine/add.ll:test28, transforming C1-(X+C2) --> (C1-C2)-X.
This occurs several dozen times in specint2k, particularly in crafty and gcc
apparently.

llvm-svn: 21136
2005-04-07 16:28:01 +00:00
Chris Lattner
ea1752ddf5 Transform X-(X+Y) == -Y and X-(Y+X) == -Y
llvm-svn: 21134
2005-04-07 16:15:25 +00:00
Andrew Lenharth
3d1228500d It wasn't happy about this either
llvm-svn: 21133
2005-04-07 14:18:13 +00:00
Andrew Lenharth
ce0dcea720 Yea, it wasn't happy
llvm-svn: 21132
2005-04-07 13:55:53 +00:00
Duraid Madina
ab6e2cfaeb teach asmprinter to print s8/s14 operands
llvm-svn: 21131
2005-04-07 12:34:36 +00:00
Duraid Madina
ed8eb83203 codegen immediate forms of add/sub/shift
llvm-svn: 21130
2005-04-07 12:33:38 +00:00
Duraid Madina
5c4a7b68b3 add immediate forms of add, sub, shift
llvm-svn: 21129
2005-04-07 12:32:24 +00:00
Chris Lattner
22bbc2351e Fix a really scary bug that Nate found where we weren't deleting the right
elements auto of the autoCSE maps.

llvm-svn: 21128
2005-04-07 00:30:13 +00:00
Nate Begeman
b890f32ac9 Pattern match bitfield insert, which helps shift long by immediate, among
other things.

llvm-svn: 21127
2005-04-06 23:51:40 +00:00
Nate Begeman
6c5e4c3bb1 Fix some shift bugs
llvm-svn: 21126
2005-04-06 22:42:08 +00:00
Alkis Evlogimenos
242fe17bc4 Make these 64 bit constants so that this compiles on x86-32 as well.
llvm-svn: 21125
2005-04-06 22:09:40 +00:00
Andrew Lenharth
abf46a4f5e added sdiv by 2^k and works for neg divisors also
llvm-svn: 21124
2005-04-06 22:03:13 +00:00
Chris Lattner
49d166c6b6 Don't make this require loopsimplify. It works BETTER with loop simplify
but should not require it.

llvm-svn: 21123
2005-04-06 21:45:00 +00:00
Nate Begeman
7898fc8cc8 Teach ExpandShift how to handle shifts by a constant. This allows targets
like PowerPC to codegen long shifts in many fewer instructions.

llvm-svn: 21122
2005-04-06 21:13:14 +00:00
Andrew Lenharth
5c71402202 fix copy/paste errors, and add imm support to SxADDQ and SxSUBQ
llvm-svn: 21121
2005-04-06 20:59:59 +00:00
Chris Lattner
40ee8a062b Fix SingleSource/Regression/C/2005-05-06-LongLongSignedShift.c, we were not
properly sign extending the top of the result of a 64-bit shift right by
a constant > 32.

llvm-svn: 21120
2005-04-06 20:59:35 +00:00
Andrew Lenharth
9f65102f9b Added Nate's div by constant stuff, also scaled operations!
llvm-svn: 21116
2005-04-06 20:25:34 +00:00
Chris Lattner
7c0786dda8 Fix a namespace issue, reported by Vladimir Merzliakov!
llvm-svn: 21115
2005-04-06 19:45:39 +00:00
Duraid Madina
5fa854a7fe steal sampo's div-by-constant-power-of-2 stuff
thanks sampo!!

llvm-svn: 21113
2005-04-06 09:55:17 +00:00
Duraid Madina
1d156293cb add fms instruction
llvm-svn: 21112
2005-04-06 09:54:09 +00:00
Nate Begeman
98251d6a1c Fixed version of optimized integer divide is now fixed. Calculate the
quotient, not the remainder.  Also, make sure to remove the old div operand
from the ExprMap and let SelectExpr insert the new one.

llvm-svn: 21111
2005-04-06 06:44:57 +00:00
Duraid Madina
2db29b8016 lie a bit and say that r1/r12 (GP/SP) _aren't_ callee-save, as we take
care of this ourselves

llvm-svn: 21110
2005-04-06 06:18:36 +00:00
Duraid Madina
62c4de8d62 make sure 'special' registers don't get allocated
llvm-svn: 21109
2005-04-06 06:17:54 +00:00
Chris Lattner
c87ff88efb Add (untested) support for MULHS and MULHU.
llvm-svn: 21107
2005-04-06 04:21:07 +00:00
Chris Lattner
ba7cdbebb1 add signed versions of the extra precision multiplies
llvm-svn: 21106
2005-04-06 04:19:22 +00:00
Nate Begeman
aee0f81849 Turn off the div -> mul optimization until it works correctly 100% of the
time.

llvm-svn: 21105
2005-04-06 03:36:33 +00:00
Nate Begeman
b44597771c Add support for MULHS and MULHU nodes
Have LegalizeDAG handle SREM and UREM for us
Codegen SDIV and UDIV by constant as a multiply by magic constant instead
of integer divide, which is very slow.

llvm-svn: 21104
2005-04-06 00:25:27 +00:00
Nate Begeman
4457b4994c Expand SREM and UREM for targets that claim not to have them, like PowerPC
llvm-svn: 21103
2005-04-06 00:23:54 +00:00
Nate Begeman
12af81407b Add MULHU and MULHS nodes for the high part of an (un)signed 32x32=64b
multiply.

llvm-svn: 21102
2005-04-05 22:36:56 +00:00
Andrew Lenharth
613a940af8 added lowerargs support for varargs
llvm-svn: 21101
2005-04-05 20:51:46 +00:00
Nate Begeman
82ff41c342 Behold, rlwinm with certain immediate arguments is printed as the much more
readable slwi or srwi (shift left/right word immediate).

llvm-svn: 21099
2005-04-05 18:19:50 +00:00
Nate Begeman
581553fd21 Fix cut & paste errors (32->64), and codegen float->int more optimally.
llvm-svn: 21098
2005-04-05 17:32:30 +00:00
Tanya Lattner
4d8f553a82 Updated to use dep analyzer.
llvm-svn: 21097
2005-04-05 16:36:44 +00:00
Nate Begeman
152dbbe856 Remove 64 bit simple ISel, it never worked correctly
Add initial (buggy) implementation of 64 bit pattern ISel

llvm-svn: 21096
2005-04-05 08:51:15 +00:00
Nate Begeman
a18a26f47c Back out the previous change to SelectBranchCC, since there are cases it
could miscompile.  A correct solution will be found in the near future.

llvm-svn: 21095
2005-04-05 04:32:16 +00:00
Nate Begeman
358dee806e Rename canUseAsImmediateForOpcode to getImmediateForOpcode to better
indicate that it is not a boolean function.
Properly emit the pseudo instruction for conditional branch, so that we
  can fix up conditional branches whose displacements are too large.
Reserve the right amount of opcode space for said pseudo instructions.

llvm-svn: 21094
2005-04-05 04:22:58 +00:00
Chris Lattner
c5b8fbe7f9 do not crash when using -debug
llvm-svn: 21092
2005-04-05 01:12:03 +00:00
Nate Begeman
ede4abc899 Implement SDIV by power of 2 as srawi/addze rather than load imm, divw
llvm-svn: 21091
2005-04-05 00:15:08 +00:00
Nate Begeman
00002553ba Pattern match fp mul-add, mul-sub, neg-mul-add, and neg-mul-sub
llvm-svn: 21090
2005-04-04 23:40:36 +00:00
Nate Begeman
682fd51f9c Add support for multiply-add, multiply-sub, and their negated versions
llvm-svn: 21089
2005-04-04 23:01:51 +00:00
Chris Lattner
57ea4daa2e do not dereference an extra layer of pointers to determine if an external
call can modify a memory location.  This fixes
test/Regression/Analysis/Andersens/modreftest.ll

llvm-svn: 21088
2005-04-04 22:23:21 +00:00