+//===- README.txt - Notes for improving PowerPC-specific code gen ---------===//
+
TODO:
* gpr0 allocation
* implement do-loop -> bdnz transform
* implement powerpc-64 for darwin
-* use stfiwx in float->int
-
-* Fold add and sub with constant into non-extern, non-weak addresses so this:
- lis r2, ha16(l2__ZTV4Cell)
- la r2, lo16(l2__ZTV4Cell)(r2)
- addi r2, r2, 8
-becomes:
- lis r2, ha16(l2__ZTV4Cell+8)
- la r2, lo16(l2__ZTV4Cell+8)(r2)
+===-------------------------------------------------------------------------===
-* Teach LLVM how to codegen this:
-unsigned short foo(float a) { return a; }
-as:
-_foo:
- fctiwz f0,f1
- stfd f0,-8(r1)
- lhz r3,-2(r1)
- blr
-not:
-_foo:
- fctiwz f0, f1
- stfd f0, -8(r1)
- lwz r2, -4(r1)
- rlwinm r3, r2, 0, 16, 31
- blr
+Support 'update' load/store instructions. These are cracked on the G5, but are
+still a codesize win.
-* Support 'update' load/store instructions. These are cracked on the G5, but
- are still a codesize win.
+===-------------------------------------------------------------------------===
-* should hint to the branch select pass that it doesn't need to print the
- second unconditional branch, so we don't end up with things like:
- b .LBBl42__2E_expand_function_8_674 ; loopentry.24
- b .LBBl42__2E_expand_function_8_42 ; NewDefault
- b .LBBl42__2E_expand_function_8_42 ; NewDefault
+Teach the .td file to pattern match PPC::BR_COND to appropriate bc variant, so
+we don't have to always run the branch selector for small functions.
===-------------------------------------------------------------------------===
as:
xoris r0,r3,0x1234
- cmpwi cr0,r0,0x5678
+ cmplwi cr0,r0,0x5678
beq cr0,L6
not:
off of the register to avoid the lis's. This is even more important in PIC
mode.
+Note that this (and the static variable version) is discussed here for GCC:
+http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
+
===-------------------------------------------------------------------------===
-Implement Newton-Rhapson method for improving estimate instructions to the
-correct accuracy, and implementing divide as multiply by reciprocal when it has
-more than one use. Itanium will want this too.
+PIC Code Gen IPO optimization:
-===-------------------------------------------------------------------------===
+Squish small scalar globals together into a single global struct, allowing the
+address of the struct to be CSE'd, avoiding PIC accesses (also reduces the size
+of the GOT on targets with one).
-int foo(int a, int b) { return a == b ? 16 : 0; }
-_foo:
- cmpw cr7, r3, r4
- mfcr r2
- rlwinm r2, r2, 31, 31, 31
- slwi r3, r2, 4
- blr
+Note that this is discussed here for GCC:
+http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
-If we exposed the srl & mask ops after the MFCR that we are doing to select
-the correct CR bit, then we could fold the slwi into the rlwinm before it.
+===-------------------------------------------------------------------------===
+
+Implement Newton-Rhapson method for improving estimate instructions to the
+correct accuracy, and implementing divide as multiply by reciprocal when it has
+more than one use. Itanium will want this too.
===-------------------------------------------------------------------------===
void AdjustBitfields(program* prog, unsigned int fmt1)
{
- unsigned int shift = 0;
- unsigned int texCount = 0;
- unsigned int i;
-
- for (i = 0; i < 8; i++)
- {
- prog->array[i].bitfields.field0 = texCount;
- prog->array[i].bitfields.field1 = texCount + 1;
- prog->array[i].bitfields.field2 = texCount + 2;
- prog->array[i].bitfields.field3 = texCount + 3;
-
- texCount += (fmt1 >> shift) & 0x7;
- shift += 3;
- }
+ prog->array[0].bitfields.field0 = fmt1;
+ prog->array[0].bitfields.field1 = fmt1 + 1;
}
-In the loop above, the bitfield adds get generated as
-(add (shl bitfield, C1), (shl C2, C1)) where C2 is 1, 2 or 3.
+We currently generate:
+
+_AdjustBitfields:
+ lwz r2, 0(r3)
+ addi r5, r4, 1
+ rlwinm r2, r2, 0, 0, 19
+ rlwinm r5, r5, 6, 20, 25
+ rlwimi r2, r4, 0, 26, 31
+ or r2, r2, r5
+ stw r2, 0(r3)
+ blr
-Since the input to the (or and, and) is an (add) rather than a (shl), the shift
-doesn't get folded into the rlwimi instruction. We should ideally see through
-things like this, rather than forcing llvm to generate the equivalent
+We should teach someone that or (rlwimi, rlwinm) with disjoint masks can be
+turned into rlwimi (rlwimi)
-(shl (add bitfield, C2), C1) with some kind of mask.
+The better codegen would be:
+
+_AdjustBitfields:
+ lwz r0,0(r3)
+ rlwinm r4,r4,0,0xff
+ rlwimi r0,r4,0,26,31
+ addi r4,r4,1
+ rlwimi r0,r4,6,20,25
+ stw r0,0(r3)
+ blr
===-------------------------------------------------------------------------===
===-------------------------------------------------------------------------===
-176.gcc contains a bunch of code like this (this occurs dozens of times):
-
-int %test(uint %mode.0.i.0) {
- %tmp.79 = cast uint %mode.0.i.0 to sbyte ; <sbyte> [#uses=1]
- %tmp.80 = cast sbyte %tmp.79 to int ; <int> [#uses=1]
- %tmp.81 = shl int %tmp.80, ubyte 16 ; <int> [#uses=1]
- %tmp.82 = and int %tmp.81, 16711680
- ret int %tmp.82
-}
-
-which we compile to:
-
-_test:
- extsb r2, r3
- rlwinm r3, r2, 16, 8, 15
- blr
-
-The extsb is obviously dead. This can be handled by a future thing like
-MaskedValueIsZero that checks to see if bits are ever demanded (in this case,
-the sign bits are never used, so we can fold the sext_inreg to nothing).
-
-I'm seeing code like this:
-
- srwi r3, r3, 16
- extsb r3, r3
- rlwimi r4, r3, 16, 8, 15
-
-in which the extsb is preventing the srwi from being nuked.
-
-===-------------------------------------------------------------------------===
-
-Another example that occurs is:
-
-uint %test(int %specbits.6.1) {
- %tmp.2540 = shr int %specbits.6.1, ubyte 11 ; <int> [#uses=1]
- %tmp.2541 = cast int %tmp.2540 to uint ; <uint> [#uses=1]
- %tmp.2542 = shl uint %tmp.2541, ubyte 13 ; <uint> [#uses=1]
- %tmp.2543 = and uint %tmp.2542, 8192 ; <uint> [#uses=1]
- ret uint %tmp.2543
-}
-
-which we codegen as:
-
-l1_test:
- srawi r2, r3, 11
- rlwinm r3, r2, 13, 18, 18
- blr
-
-the srawi can be nuked by turning the SAR into a logical SHR (the sext bits are
-dead), which I think can then be folded into the rlwinm.
-
-===-------------------------------------------------------------------------===
-
Compile offsets from allocas:
int *%test() {
===-------------------------------------------------------------------------===
-For this:
-
-int h(int i, int j, int k) {
- return (i==0||j==0||k == 0);
-}
-
-We currently emit this:
-
-_h:
- cntlzw r2, r3
- cntlzw r3, r4
- cntlzw r4, r5
- srwi r2, r2, 5
- srwi r3, r3, 5
- srwi r4, r4, 5
- or r2, r3, r2
- or r3, r2, r4
- blr
-
-The ctlz/shift instructions are created by the isel, so the dag combiner doesn't
-have a chance to pull the shifts through the or's (eliminating two
-instructions). SETCC nodes should be custom lowered in this case, not expanded
-by the isel.
-
-===-------------------------------------------------------------------------===
-
Darwin Stub LICM optimization:
Loops like this:
cmpwi cr0, r7, 0
bne cr0, LBB_compare_4 ; loopexit
+FreeBench/mason has a basic block that looks like this:
+
+ %tmp.130 = seteq int %p.0__, 5 ; <bool> [#uses=1]
+ %tmp.134 = seteq int %p.1__, 6 ; <bool> [#uses=1]
+ %tmp.139 = seteq int %p.2__, 12 ; <bool> [#uses=1]
+ %tmp.144 = seteq int %p.3__, 13 ; <bool> [#uses=1]
+ %tmp.149 = seteq int %p.4__, 14 ; <bool> [#uses=1]
+ %tmp.154 = seteq int %p.5__, 15 ; <bool> [#uses=1]
+ %bothcond = and bool %tmp.134, %tmp.130 ; <bool> [#uses=1]
+ %bothcond123 = and bool %bothcond, %tmp.139 ; <bool>
+ %bothcond124 = and bool %bothcond123, %tmp.144 ; <bool>
+ %bothcond125 = and bool %bothcond124, %tmp.149 ; <bool>
+ %bothcond126 = and bool %bothcond125, %tmp.154 ; <bool>
+ br bool %bothcond126, label %shortcirc_next.5, label %else.0
+
+This is a particularly important case where handling CRs better will help.
+
===-------------------------------------------------------------------------===
Simple IPO for argument passing, change:
===-------------------------------------------------------------------------===
-Code Gen IPO optimization:
+Generate lwbrx and other byteswapping load/store instructions when reasonable.
-Squish small scalar globals together into a single global struct, allowing the
-address of the struct to be CSE'd, avoiding PIC accesses (also reduces the size
-of the GOT on targets with one).
+===-------------------------------------------------------------------------===
+
+Compile this:
+
+int foo(int a) {
+ int b = (a < 8);
+ if (b) {
+ return b * 3; // ignore the fact that this is always 3.
+ } else {
+ return 2;
+ }
+}
+
+into something not this:
+
+_foo:
+1) cmpwi cr7, r3, 8
+ mfcr r2, 1
+ rlwinm r2, r2, 29, 31, 31
+1) cmpwi cr0, r3, 7
+ bgt cr0, LBB1_2 ; UnifiedReturnBlock
+LBB1_1: ; then
+ rlwinm r2, r2, 0, 31, 31
+ mulli r3, r2, 3
+ blr
+LBB1_2: ; UnifiedReturnBlock
+ li r3, 2
+ blr
+
+In particular, the two compares (marked 1) could be shared by reversing one.
+This could be done in the dag combiner, by swapping a BR_CC when a SETCC of the
+same operands (but backwards) exists. In this case, this wouldn't save us
+anything though, because the compares still wouldn't be shared.
===-------------------------------------------------------------------------===
-Generate lwbrx and other byteswapping load/store instructions when reasonable.
+The legalizer should lower this:
+
+bool %test(ulong %x) {
+ %tmp = setlt ulong %x, 4294967296
+ ret bool %tmp
+}
+
+into "if x.high == 0", not:
+
+_test:
+ addi r2, r3, -1
+ cntlzw r2, r2
+ cntlzw r3, r3
+ srwi r2, r2, 5
+ srwi r4, r3, 5
+ li r3, 0
+ cmpwi cr0, r2, 0
+ bne cr0, LBB1_2 ;
+LBB1_1:
+ or r3, r4, r4
+LBB1_2:
+ blr
+
+noticed in 2005-05-11-Popcount-ffs-fls.c.
+
+
+===-------------------------------------------------------------------------===
+
+We should custom expand setcc instead of pretending that we have it. That
+would allow us to expose the access of the crbit after the mfcr, allowing
+that access to be trivially folded into other ops. A simple example:
+
+int foo(int a, int b) { return (a < b) << 4; }
+
+compiles into:
+
+_foo:
+ cmpw cr7, r3, r4
+ mfcr r2, 1
+ rlwinm r2, r2, 29, 31, 31
+ slwi r3, r2, 4
+ blr
+
+===-------------------------------------------------------------------------===
+
+Fold add and sub with constant into non-extern, non-weak addresses so this:
+
+static int a;
+void bar(int b) { a = b; }
+void foo(unsigned char *c) {
+ *c = a;
+}
+
+So that
+
+_foo:
+ lis r2, ha16(_a)
+ la r2, lo16(_a)(r2)
+ lbz r2, 3(r2)
+ stb r2, 0(r3)
+ blr
+
+Becomes
+
+_foo:
+ lis r2, ha16(_a+3)
+ lbz r2, lo16(_a+3)(r2)
+ stb r2, 0(r3)
+ blr
+
+===-------------------------------------------------------------------------===
+
+We generate really bad code for this:
+
+int f(signed char *a, _Bool b, _Bool c) {
+ signed char t = 0;
+ if (b) t = *a;
+ if (c) *a = t;
+}
+
+===-------------------------------------------------------------------------===
+
+This:
+int test(unsigned *P) { return *P >> 24; }
+
+Should compile to:
+
+_test:
+ lbz r3,0(r3)
+ blr
+
+not:
+
+_test:
+ lwz r2, 0(r3)
+ srwi r3, r2, 24
+ blr
+
+===-------------------------------------------------------------------------===
+
+On the G5, logical CR operations are more expensive in their three
+address form: ops that read/write the same register are half as expensive as
+those that read from two registers that are different from their destination.
+
+We should model this with two separate instructions. The isel should generate
+the "two address" form of the instructions. When the register allocator
+detects that it needs to insert a copy due to the two-addresness of the CR
+logical op, it will invoke PPCInstrInfo::convertToThreeAddress. At this point
+we can convert to the "three address" instruction, to save code space.
+
+This only matters when we start generating cr logical ops.
+
+===-------------------------------------------------------------------------===
+
+We should compile these two functions to the same thing:
+
+#include <stdlib.h>
+void f(int a, int b, int *P) {
+ *P = (a-b)>=0?(a-b):(b-a);
+}
+void g(int a, int b, int *P) {
+ *P = abs(a-b);
+}
+
+Further, they should compile to something better than:
+
+_g:
+ subf r2, r4, r3
+ subfic r3, r2, 0
+ cmpwi cr0, r2, -1
+ bgt cr0, LBB2_2 ; entry
+LBB2_1: ; entry
+ mr r2, r3
+LBB2_2: ; entry
+ stw r2, 0(r5)
+ blr
+
+GCC produces:
+
+_g:
+ subf r4,r4,r3
+ srawi r2,r4,31
+ xor r0,r2,r4
+ subf r0,r2,r0
+ stw r0,0(r5)
+ blr
+
+... which is much nicer.
+
+This theoretically may help improve twolf slightly (used in dimbox.c:142?).
+
+===-------------------------------------------------------------------------===
+
+int foo(int N, int ***W, int **TK, int X) {
+ int t, i;
+
+ for (t = 0; t < N; ++t)
+ for (i = 0; i < 4; ++i)
+ W[t / X][i][t % X] = TK[i][t];
+
+ return 5;
+}
+
+We generate relatively atrocious code for this loop compared to gcc.
+
+We could also strength reduce the rem and the div:
+http://www.lcs.mit.edu/pubs/pdf/MIT-LCS-TM-600.pdf
+
+===-------------------------------------------------------------------------===
+
+float foo(float X) { return (int)(X); }
+
+Currently produces:
+
+_foo:
+ fctiwz f0, f1
+ stfd f0, -8(r1)
+ lwz r2, -4(r1)
+ extsw r2, r2
+ std r2, -16(r1)
+ lfd f0, -16(r1)
+ fcfid f0, f0
+ frsp f1, f0
+ blr
+
+We could use a target dag combine to turn the lwz/extsw into an lwa when the
+lwz has a single use. Since LWA is cracked anyway, this would be a codesize
+win only.
+
+===-------------------------------------------------------------------------===
+
+We generate ugly code for this:
+
+void func(unsigned int *ret, float dx, float dy, float dz, float dw) {
+ unsigned code = 0;
+ if(dx < -dw) code |= 1;
+ if(dx > dw) code |= 2;
+ if(dy < -dw) code |= 4;
+ if(dy > dw) code |= 8;
+ if(dz < -dw) code |= 16;
+ if(dz > dw) code |= 32;
+ *ret = code;
+}
+
+===-------------------------------------------------------------------------===