1 //===- README_ALTIVEC.txt - Notes for improving Altivec code gen ----------===//
3 Implement PPCInstrInfo::isLoadFromStackSlot/isStoreToStackSlot for vector
4 registers, to generate better spill code.
6 //===----------------------------------------------------------------------===//
8 The first should be a single lvx from the constant pool, the second should be
12 int x[8] __attribute__((aligned(128))) = { 1, 1, 1, 17, 1, 1, 1, 1 };
18 int x[8] __attribute__((aligned(128)));
19 memset (x, 0, sizeof (x));
23 //===----------------------------------------------------------------------===//
25 Altivec: Codegen'ing MUL with vector FMADD should add -0.0, not 0.0:
26 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=8763
28 When -ffast-math is on, we can use 0.0.
30 //===----------------------------------------------------------------------===//
34 v4f32 Vector2 = { Vector.X, Vector.X, Vector.X, Vector.X };
36 Since we know that "Vector" is 16-byte aligned and we know the element offset
37 of ".X", we should change the load into a lve*x instruction, instead of doing
38 a load/store/lve*x sequence.
40 //===----------------------------------------------------------------------===//
42 Implement passing vectors by value into calls and receiving them as arguments.
44 //===----------------------------------------------------------------------===//
46 GCC apparently tries to codegen { C1, C2, Variable, C3 } as a constant pool load
47 of C1/C2/C3, then a load and vperm of Variable.
49 //===----------------------------------------------------------------------===//
51 We need a way to teach tblgen that some operands of an intrinsic are required to
52 be constants. The verifier should enforce this constraint.
54 //===----------------------------------------------------------------------===//
56 We currently codegen SCALAR_TO_VECTOR as a store of the scalar to a 16-byte
57 aligned stack slot, followed by a load/vperm. We should probably just store it
58 to a scalar stack slot, then use lvsl/vperm to load it. If the value is already
59 in memory this is a big win.
61 //===----------------------------------------------------------------------===//
63 extract_vector_elt of an arbitrary constant vector can be done with the
64 following instructions:
66 vTemp = vec_splat(v0,2); // 2 is the element the src is in.
67 vec_ste(&destloc,0,vTemp);
69 We can do an arbitrary non-constant value by using lvsr/perm/ste.
71 //===----------------------------------------------------------------------===//
73 If we want to tie instruction selection into the scheduler, we can do some
74 constant formation with different instructions. For example, we can generate
75 "vsplti -1" with "vcmpequw R,R" and 1,1,1,1 with "vsubcuw R,R", and 0,0,0,0 with
76 "vsplti 0" or "vxor", each of which use different execution units, thus could
79 This is probably only reasonable for a post-pass scheduler.
81 //===----------------------------------------------------------------------===//
85 void test(vector float *A, vector float *B) {
86 vector float C = (vector float)vec_cmpeq(*A, *B);
87 if (!vec_any_eq(*A, *B))
88 *B = (vector float){0,0,0,0};
92 we get the following basic block:
99 bne cr6, LBB1_2 ; cond_next
101 The vcmpeqfp/vcmpeqfp. instructions currently cannot be merged when the
102 vcmpeqfp. result is used by a branch. This can be improved.
104 //===----------------------------------------------------------------------===//
106 The code generated for this is truly aweful:
108 vector float test(float a, float b) {
109 return (vector float){ 0.0, a, 0.0, 0.0};
121 lis r3, ha16(LCPI1_0)
125 lfs f0, lo16(LCPI1_0)(r3)
135 //===----------------------------------------------------------------------===//
137 int foo(vector float *x, vector float *y) {
138 if (vec_all_eq(*x,*y)) return 3245;
142 A predicate compare being used in a select_cc should have the same peephole
143 applied to it as a predicate compare used by a br_cc. There should be no
156 rlwinm r3, r3, 25, 31, 31
158 bne cr0, LBB1_2 ; entry
166 //===----------------------------------------------------------------------===//
168 CodeGen/PowerPC/vec_constants.ll has an and operation that should be
169 codegen'd to andc. The issue is that the 'all ones' build vector is
170 SelectNodeTo'd a VSPLTISB instruction node before the and/xor is selected
171 which prevents the vnot pattern from matching.
174 //===----------------------------------------------------------------------===//
176 An alternative to the store/store/load approach for illegal insert element
179 1. store element to any ol' slot
181 3. lvsl 0; splat index; vcmpeq to generate a select mask
182 4. lvsl slot + x; vperm to rotate result into correct slot
183 5. vsel result together.
185 //===----------------------------------------------------------------------===//
187 Should codegen branches on vec_any/vec_all to avoid mfcr. Two examples:
190 int f(vector float a, vector float b)
193 if (vec_all_ge(a, b))
200 vector float f(vector float a, vector float b) {
201 if (vec_any_eq(a, b))
207 //===----------------------------------------------------------------------===//
209 We should do a little better with eliminating dead stores.
210 The stores to the stack are dead since %a and %b are not needed
212 ; Function Attrs: nounwind
213 define <16 x i8> @test_vpmsumb() #0 {
215 %a = alloca <16 x i8>, align 16
216 %b = alloca <16 x i8>, align 16
217 store <16 x i8> <i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 9, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15, i8 16>, <16 x i8>* %a, align 16
218 store <16 x i8> <i8 113, i8 114, i8 115, i8 116, i8 117, i8 118, i8 119, i8 120, i8 121, i8 122, i8 123, i8 124, i8 125, i8 126, i8 127, i8 112>, <16 x i8>* %b, align 16
219 %0 = load <16 x i8>* %a, align 16
220 %1 = load <16 x i8>* %b, align 16
221 %2 = call <16 x i8> @llvm.ppc.altivec.crypto.vpmsumb(<16 x i8> %0, <16 x i8> %1)
226 ; Function Attrs: nounwind readnone
227 declare <16 x i8> @llvm.ppc.altivec.crypto.vpmsumb(<16 x i8>, <16 x i8>) #1
230 Produces the following code with -mtriple=powerpc64-unknown-linux-gnu:
232 addis 3, 2, .LCPI0_0@toc@ha
233 addis 4, 2, .LCPI0_1@toc@ha
234 addi 3, 3, .LCPI0_0@toc@l
235 addi 4, 4, .LCPI0_1@toc@l
249 The two stxvw4x instructions are not needed.
250 With -mtriple=powerpc64le-unknown-linux-gnu, the associated permutes
253 //===----------------------------------------------------------------------===//
255 The following example is found in test/CodeGen/PowerPC/vec_add_sub_doubleword.ll:
257 define <2 x i64> @increment_by_val(<2 x i64> %x, i64 %val) nounwind {
258 %tmpvec = insertelement <2 x i64> <i64 0, i64 0>, i64 %val, i32 0
259 %tmpvec2 = insertelement <2 x i64> %tmpvec, i64 %val, i32 1
260 %result = add <2 x i64> %x, %tmpvec2
261 ret <2 x i64> %result
263 This will generate the following instruction sequence:
272 This will almost certainly cause a load-hit-store hazard.
273 Since val is a value parameter, it should not need to be saved onto
274 the stack, unless it's being done set up the vector register. Instead,
275 it would be better to splat the value into a vector register, and then
276 remove the (dead) stores to the stack.
278 //===----------------------------------------------------------------------===//
280 At the moment we always generate a lxsdx in preference to lfd, or stxsdx in
281 preference to stfd. When we have a reg-immediate addressing mode, this is a
282 poor choice, since we have to load the address into an index register. This
283 should be fixed for P7/P8.
285 //===----------------------------------------------------------------------===//
287 Right now, ShuffleKind 0 is supported only on BE, and ShuffleKind 2 only on LE.
288 However, we could actually support both kinds on either endianness, if we check
289 for the appropriate shufflevector pattern for each case ... this would cause
290 some additional shufflevectors to be recognized and implemented via the
293 //===----------------------------------------------------------------------===//
295 There is a utility program called PerfectShuffle that generates a table of the
296 shortest instruction sequence for implementing a shufflevector operation on
297 PowerPC. However, this was designed for big-endian code generation. We could
298 modify this program to create a little endian version of the table. The table
299 is used in PPCISelLowering.cpp, PPCTargetLowering::LOWERVECTOR_SHUFFLE().
301 //===----------------------------------------------------------------------===//
303 Opportunies to use instructions from PPCInstrVSX.td during code gen
304 - Conversion instructions (Sections 7.6.1.5 and 7.6.1.6 of ISA 2.07)
305 - Scalar comparisons (xscmpodp and xscmpudp)
306 - Min and max (xsmaxdp, xsmindp, xvmaxdp, xvmindp, xvmaxsp, xvminsp)
308 Related to this: we currently do not generate the lxvw4x instruction for either
309 v4f32 or v4i32, probably because adding a dag pattern to the recognizer requires
310 a single target type. This should probably be addressed in the PPCISelDAGToDAG logic.
312 //===----------------------------------------------------------------------===//
314 Currently EXTRACT_VECTOR_ELT and INSERT_VECTOR_ELT are type-legal only
315 for v2f64 with VSX available. We should create custom lowering
316 support for the other vector types. Without this support, we generate
317 sequences with load-hit-store hazards.
319 v4f32 can be supported with VSX by shifting the correct element into
320 big-endian lane 0, using xscvspdpn to produce a double-precision
321 representation of the single-precision value in big-endian
322 double-precision lane 0, and reinterpreting lane 0 as an FPR or
323 vector-scalar register.
325 v2i64 can be supported with VSX and P8Vector in the same manner as
326 v2f64, followed by a direct move to a GPR.
328 v4i32 can be supported with VSX and P8Vector by shifting the correct
329 element into big-endian lane 1, using a direct move to a GPR, and
330 sign-extending the 32-bit result to 64 bits.
332 v8i16 can be supported with VSX and P8Vector by shifting the correct
333 element into big-endian lane 3, using a direct move to a GPR, and
334 sign-extending the 16-bit result to 64 bits.
336 v16i8 can be supported with VSX and P8Vector by shifting the correct
337 element into big-endian lane 7, using a direct move to a GPR, and
338 sign-extending the 8-bit result to 64 bits.