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 For functions that use altivec AND have calls, we are VRSAVE'ing all call
45 //===----------------------------------------------------------------------===//
47 Implement passing vectors by value into calls and receiving them as arguments.
49 //===----------------------------------------------------------------------===//
51 GCC apparently tries to codegen { C1, C2, Variable, C3 } as a constant pool load
52 of C1/C2/C3, then a load and vperm of Variable.
54 //===----------------------------------------------------------------------===//
56 We need a way to teach tblgen that some operands of an intrinsic are required to
57 be constants. The verifier should enforce this constraint.
59 //===----------------------------------------------------------------------===//
61 We currently codegen SCALAR_TO_VECTOR as a store of the scalar to a 16-byte
62 aligned stack slot, followed by a load/vperm. We should probably just store it
63 to a scalar stack slot, then use lvsl/vperm to load it. If the value is already
64 in memory this is a big win.
66 //===----------------------------------------------------------------------===//
68 extract_vector_elt of an arbitrary constant vector can be done with the
69 following instructions:
71 vTemp = vec_splat(v0,2); // 2 is the element the src is in.
72 vec_ste(&destloc,0,vTemp);
74 We can do an arbitrary non-constant value by using lvsr/perm/ste.
76 //===----------------------------------------------------------------------===//
78 If we want to tie instruction selection into the scheduler, we can do some
79 constant formation with different instructions. For example, we can generate
80 "vsplti -1" with "vcmpequw R,R" and 1,1,1,1 with "vsubcuw R,R", and 0,0,0,0 with
81 "vsplti 0" or "vxor", each of which use different execution units, thus could
84 This is probably only reasonable for a post-pass scheduler.
86 //===----------------------------------------------------------------------===//
90 void test(vector float *A, vector float *B) {
91 vector float C = (vector float)vec_cmpeq(*A, *B);
92 if (!vec_any_eq(*A, *B))
93 *B = (vector float){0,0,0,0};
97 we get the following basic block:
104 bne cr6, LBB1_2 ; cond_next
106 The vcmpeqfp/vcmpeqfp. instructions currently cannot be merged when the
107 vcmpeqfp. result is used by a branch. This can be improved.
109 //===----------------------------------------------------------------------===//
111 The code generated for this is truly aweful:
113 vector float test(float a, float b) {
114 return (vector float){ 0.0, a, 0.0, 0.0};
126 lis r3, ha16(LCPI1_0)
130 lfs f0, lo16(LCPI1_0)(r3)
140 //===----------------------------------------------------------------------===//
142 int foo(vector float *x, vector float *y) {
143 if (vec_all_eq(*x,*y)) return 3245;
147 A predicate compare being used in a select_cc should have the same peephole
148 applied to it as a predicate compare used by a br_cc. There should be no
161 rlwinm r3, r3, 25, 31, 31
163 bne cr0, LBB1_2 ; entry
171 //===----------------------------------------------------------------------===//
173 CodeGen/PowerPC/vec_constants.ll has an and operation that should be
174 codegen'd to andc. The issue is that the 'all ones' build vector is
175 SelectNodeTo'd a VSPLTISB instruction node before the and/xor is selected
176 which prevents the vnot pattern from matching.
179 //===----------------------------------------------------------------------===//
181 An alternative to the store/store/load approach for illegal insert element
184 1. store element to any ol' slot
186 3. lvsl 0; splat index; vcmpeq to generate a select mask
187 4. lvsl slot + x; vperm to rotate result into correct slot
188 5. vsel result together.
190 //===----------------------------------------------------------------------===//
192 Should codegen branches on vec_any/vec_all to avoid mfcr. Two examples:
195 int f(vector float a, vector float b)
198 if (vec_all_ge(a, b))
205 vector float f(vector float a, vector float b) {
206 if (vec_any_eq(a, b))
212 //===----------------------------------------------------------------------===//
214 We should do a little better with eliminating dead stores.
215 The stores to the stack are dead since %a and %b are not needed
217 ; Function Attrs: nounwind
218 define <16 x i8> @test_vpmsumb() #0 {
220 %a = alloca <16 x i8>, align 16
221 %b = alloca <16 x i8>, align 16
222 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
223 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
224 %0 = load <16 x i8>* %a, align 16
225 %1 = load <16 x i8>* %b, align 16
226 %2 = call <16 x i8> @llvm.ppc.altivec.crypto.vpmsumb(<16 x i8> %0, <16 x i8> %1)
231 ; Function Attrs: nounwind readnone
232 declare <16 x i8> @llvm.ppc.altivec.crypto.vpmsumb(<16 x i8>, <16 x i8>) #1
235 Produces the following code with -mtriple=powerpc64-unknown-linux-gnu:
237 addis 3, 2, .LCPI0_0@toc@ha
238 addis 4, 2, .LCPI0_1@toc@ha
239 addi 3, 3, .LCPI0_0@toc@l
240 addi 4, 4, .LCPI0_1@toc@l
254 The two stxvw4x instructions are not needed.
255 With -mtriple=powerpc64le-unknown-linux-gnu, the associated permutes
258 //===----------------------------------------------------------------------===//
260 The following example is found in test/CodeGen/PowerPC/vec_add_sub_doubleword.ll:
262 define <2 x i64> @increment_by_val(<2 x i64> %x, i64 %val) nounwind {
263 %tmpvec = insertelement <2 x i64> <i64 0, i64 0>, i64 %val, i32 0
264 %tmpvec2 = insertelement <2 x i64> %tmpvec, i64 %val, i32 1
265 %result = add <2 x i64> %x, %tmpvec2
266 ret <2 x i64> %result
268 This will generate the following instruction sequence:
277 This will almost certainly cause a load-hit-store hazard.
278 Since val is a value parameter, it should not need to be saved onto
279 the stack, unless it's being done set up the vector register. Instead,
280 it would be better to splat the value into a vector register, and then
281 remove the (dead) stores to the stack.
283 //===----------------------------------------------------------------------===//
285 At the moment we always generate a lxsdx in preference to lfd, or stxsdx in
286 preference to stfd. When we have a reg-immediate addressing mode, this is a
287 poor choice, since we have to load the address into an index register. This
288 should be fixed for P7/P8.
290 //===----------------------------------------------------------------------===//
292 Right now, ShuffleKind 0 is supported only on BE, and ShuffleKind 2 only on LE.
293 However, we could actually support both kinds on either endianness, if we check
294 for the appropriate shufflevector pattern for each case ... this would cause
295 some additional shufflevectors to be recognized and implemented via the
298 //===----------------------------------------------------------------------===//
300 There is a utility program called PerfectShuffle that generates a table of the
301 shortest instruction sequence for implementing a shufflevector operation on
302 PowerPC. However, this was designed for big-endian code generation. We could
303 modify this program to create a little endian version of the table. The table
304 is used in PPCISelLowering.cpp, PPCTargetLowering::LOWERVECTOR_SHUFFLE().
306 //===----------------------------------------------------------------------===//
308 Opportunies to use instructions from PPCInstrVSX.td during code gen
309 - Conversion instructions (Sections 7.6.1.5 and 7.6.1.6 of ISA 2.07)
310 - Scalar comparisons (xscmpodp and xscmpudp)
311 - Min and max (xsmaxdp, xsmindp, xvmaxdp, xvmindp, xvmaxsp, xvminsp)
313 Related to this: we currently do not generate the lxvw4x instruction for either
314 v4f32 or v4i32, probably because adding a dag pattern to the recognizer requires
315 a single target type. This should probably be addressed in the PPCISelDAGToDAG logic.
317 //===----------------------------------------------------------------------===//
319 Currently EXTRACT_VECTOR_ELT and INSERT_VECTOR_ELT are type-legal only
320 for v2f64 with VSX available. We should create custom lowering
321 support for the other vector types. Without this support, we generate
322 sequences with load-hit-store hazards.
324 v4f32 can be supported with VSX by shifting the correct element into
325 big-endian lane 0, using xscvspdpn to produce a double-precision
326 representation of the single-precision value in big-endian
327 double-precision lane 0, and reinterpreting lane 0 as an FPR or
328 vector-scalar register.
330 v2i64 can be supported with VSX and P8Vector in the same manner as
331 v2f64, followed by a direct move to a GPR.
333 v4i32 can be supported with VSX and P8Vector by shifting the correct
334 element into big-endian lane 1, using a direct move to a GPR, and
335 sign-extending the 32-bit result to 64 bits.
337 v8i16 can be supported with VSX and P8Vector by shifting the correct
338 element into big-endian lane 3, using a direct move to a GPR, and
339 sign-extending the 16-bit result to 64 bits.
341 v16i8 can be supported with VSX and P8Vector by shifting the correct
342 element into big-endian lane 7, using a direct move to a GPR, and
343 sign-extending the 8-bit result to 64 bits.