1 //===---------------------------------------------------------------------===//
2 // Random ideas for the X86 backend: SSE-specific stuff.
3 //===---------------------------------------------------------------------===//
5 - Consider eliminating the unaligned SSE load intrinsics, replacing them with
6 unaligned LLVM load instructions.
8 //===---------------------------------------------------------------------===//
10 Expand libm rounding functions inline: Significant speedups possible.
11 http://gcc.gnu.org/ml/gcc-patches/2006-10/msg00909.html
13 //===---------------------------------------------------------------------===//
15 When compiled with unsafemath enabled, "main" should enable SSE DAZ mode and
18 //===---------------------------------------------------------------------===//
20 Think about doing i64 math in SSE regs.
22 //===---------------------------------------------------------------------===//
24 This testcase should have no SSE instructions in it, and only one load from
27 double %test3(bool %B) {
28 %C = select bool %B, double 123.412, double 523.01123123
32 Currently, the select is being lowered, which prevents the dag combiner from
33 turning 'select (load CPI1), (load CPI2)' -> 'load (select CPI1, CPI2)'
35 The pattern isel got this one right.
37 //===---------------------------------------------------------------------===//
39 SSE doesn't have [mem] op= reg instructions. If we have an SSE instruction
44 and the register allocator decides to spill X, it is cheaper to emit this as:
55 ..and this uses one fewer register (so this should be done at load folding
56 time, not at spiller time). *Note* however that this can only be done
57 if Y is dead. Here's a testcase:
59 @.str_3 = external global [15 x i8] ; <[15 x i8]*> [#uses=0]
60 declare void @printf(i32, ...)
65 no_exit.i7: ; preds = %no_exit.i7, %build_tree.exit
66 %tmp.0.1.0.i9 = phi double [ 0.000000e+00, %build_tree.exit ], [ %tmp.34.i18, %no_exit.i7 ] ; <double> [#uses=1]
67 %tmp.0.0.0.i10 = phi double [ 0.000000e+00, %build_tree.exit ], [ %tmp.28.i16, %no_exit.i7 ] ; <double> [#uses=1]
68 %tmp.28.i16 = add double %tmp.0.0.0.i10, 0.000000e+00 ; <double> [#uses=1]
69 %tmp.34.i18 = add double %tmp.0.1.0.i9, 0.000000e+00 ; <double> [#uses=2]
70 br i1 false, label %Compute_Tree.exit23, label %no_exit.i7
72 Compute_Tree.exit23: ; preds = %no_exit.i7
73 tail call void (i32, ...)* @printf( i32 0 )
74 store double %tmp.34.i18, double* null
83 *** movsd %XMM2, QWORD PTR [%ESP + 8]
84 *** addsd %XMM2, %XMM1
85 *** movsd QWORD PTR [%ESP + 8], %XMM2
86 jmp .BBmain_1 # no_exit.i7
88 This is a bugpoint reduced testcase, which is why the testcase doesn't make
89 much sense (e.g. its an infinite loop). :)
91 //===---------------------------------------------------------------------===//
93 SSE should implement 'select_cc' using 'emulated conditional moves' that use
94 pcmp/pand/pandn/por to do a selection instead of a conditional branch:
96 double %X(double %Y, double %Z, double %A, double %B) {
97 %C = setlt double %A, %B
98 %z = add double %Z, 0.0 ;; select operand is not a load
99 %D = select bool %C, double %Y, double %z
108 addsd 24(%esp), %xmm0
109 movsd 32(%esp), %xmm1
110 movsd 16(%esp), %xmm2
111 ucomisd 40(%esp), %xmm1
121 //===---------------------------------------------------------------------===//
123 It's not clear whether we should use pxor or xorps / xorpd to clear XMM
124 registers. The choice may depend on subtarget information. We should do some
125 more experiments on different x86 machines.
127 //===---------------------------------------------------------------------===//
129 Lower memcpy / memset to a series of SSE 128 bit move instructions when it's
132 //===---------------------------------------------------------------------===//
135 if (copysign(1.0, x) == copysign(1.0, y))
140 //===---------------------------------------------------------------------===//
142 Use movhps to update upper 64-bits of a v4sf value. Also movlps on lower half
145 //===---------------------------------------------------------------------===//
147 Better codegen for vector_shuffles like this { x, 0, 0, 0 } or { x, 0, x, 0}.
148 Perhaps use pxor / xorp* to clear a XMM register first?
150 //===---------------------------------------------------------------------===//
152 How to decide when to use the "floating point version" of logical ops? Here are
155 movaps LCPI5_5, %xmm2
158 mulps 8656(%ecx), %xmm3
159 addps 8672(%ecx), %xmm3
165 movaps LCPI5_5, %xmm1
168 mulps 8656(%ecx), %xmm3
169 addps 8672(%ecx), %xmm3
173 movaps %xmm3, 112(%esp)
176 Due to some minor source change, the later case ended up using orps and movaps
177 instead of por and movdqa. Does it matter?
179 //===---------------------------------------------------------------------===//
181 X86RegisterInfo::copyRegToReg() returns X86::MOVAPSrr for VR128. Is it possible
182 to choose between movaps, movapd, and movdqa based on types of source and
185 How about andps, andpd, and pand? Do we really care about the type of the packed
186 elements? If not, why not always use the "ps" variants which are likely to be
189 //===---------------------------------------------------------------------===//
191 External test Nurbs exposed some problems. Look for
192 __ZN15Nurbs_SSE_Cubic17TessellateSurfaceE, bb cond_next140. This is what icc
195 movaps (%edx), %xmm2 #59.21
196 movaps (%edx), %xmm5 #60.21
197 movaps (%edx), %xmm4 #61.21
198 movaps (%edx), %xmm3 #62.21
199 movl 40(%ecx), %ebp #69.49
200 shufps $0, %xmm2, %xmm5 #60.21
201 movl 100(%esp), %ebx #69.20
202 movl (%ebx), %edi #69.20
203 imull %ebp, %edi #69.49
204 addl (%eax), %edi #70.33
205 shufps $85, %xmm2, %xmm4 #61.21
206 shufps $170, %xmm2, %xmm3 #62.21
207 shufps $255, %xmm2, %xmm2 #63.21
208 lea (%ebp,%ebp,2), %ebx #69.49
210 lea -3(%edi,%ebx), %ebx #70.33
212 addl 32(%ecx), %ebx #68.37
213 testb $15, %bl #91.13
214 jne L_B1.24 # Prob 5% #91.13
216 This is the llvm code after instruction scheduling:
218 cond_next140 (0xa910740, LLVM BB @0xa90beb0):
219 %reg1078 = MOV32ri -3
220 %reg1079 = ADD32rm %reg1078, %reg1068, 1, %NOREG, 0
221 %reg1037 = MOV32rm %reg1024, 1, %NOREG, 40
222 %reg1080 = IMUL32rr %reg1079, %reg1037
223 %reg1081 = MOV32rm %reg1058, 1, %NOREG, 0
224 %reg1038 = LEA32r %reg1081, 1, %reg1080, -3
225 %reg1036 = MOV32rm %reg1024, 1, %NOREG, 32
226 %reg1082 = SHL32ri %reg1038, 4
227 %reg1039 = ADD32rr %reg1036, %reg1082
228 %reg1083 = MOVAPSrm %reg1059, 1, %NOREG, 0
229 %reg1034 = SHUFPSrr %reg1083, %reg1083, 170
230 %reg1032 = SHUFPSrr %reg1083, %reg1083, 0
231 %reg1035 = SHUFPSrr %reg1083, %reg1083, 255
232 %reg1033 = SHUFPSrr %reg1083, %reg1083, 85
233 %reg1040 = MOV32rr %reg1039
234 %reg1084 = AND32ri8 %reg1039, 15
236 JE mbb<cond_next204,0xa914d30>
238 Still ok. After register allocation:
240 cond_next140 (0xa910740, LLVM BB @0xa90beb0):
242 %EDX = MOV32rm <fi#3>, 1, %NOREG, 0
243 ADD32rm %EAX<def&use>, %EDX, 1, %NOREG, 0
244 %EDX = MOV32rm <fi#7>, 1, %NOREG, 0
245 %EDX = MOV32rm %EDX, 1, %NOREG, 40
246 IMUL32rr %EAX<def&use>, %EDX
247 %ESI = MOV32rm <fi#5>, 1, %NOREG, 0
248 %ESI = MOV32rm %ESI, 1, %NOREG, 0
249 MOV32mr <fi#4>, 1, %NOREG, 0, %ESI
250 %EAX = LEA32r %ESI, 1, %EAX, -3
251 %ESI = MOV32rm <fi#7>, 1, %NOREG, 0
252 %ESI = MOV32rm %ESI, 1, %NOREG, 32
254 SHL32ri %EDI<def&use>, 4
255 ADD32rr %EDI<def&use>, %ESI
256 %XMM0 = MOVAPSrm %ECX, 1, %NOREG, 0
257 %XMM1 = MOVAPSrr %XMM0
258 SHUFPSrr %XMM1<def&use>, %XMM1, 170
259 %XMM2 = MOVAPSrr %XMM0
260 SHUFPSrr %XMM2<def&use>, %XMM2, 0
261 %XMM3 = MOVAPSrr %XMM0
262 SHUFPSrr %XMM3<def&use>, %XMM3, 255
263 SHUFPSrr %XMM0<def&use>, %XMM0, 85
265 AND32ri8 %EBX<def&use>, 15
267 JE mbb<cond_next204,0xa914d30>
269 This looks really bad. The problem is shufps is a destructive opcode. Since it
270 appears as operand two in more than one shufps ops. It resulted in a number of
271 copies. Note icc also suffers from the same problem. Either the instruction
272 selector should select pshufd or The register allocator can made the two-address
273 to three-address transformation.
275 It also exposes some other problems. See MOV32ri -3 and the spills.
277 //===---------------------------------------------------------------------===//
279 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25500
281 LLVM is producing bad code.
283 LBB_main_4: # cond_true44
294 jne LBB_main_4 # cond_true44
296 There are two problems. 1) No need to two loop induction variables. We can
297 compare against 262144 * 16. 2) Known register coalescer issue. We should
298 be able eliminate one of the movaps:
300 addps %xmm2, %xmm1 <=== Commute!
303 movaps %xmm1, %xmm1 <=== Eliminate!
310 jne LBB_main_4 # cond_true44
312 //===---------------------------------------------------------------------===//
316 __m128 test(float a) {
317 return _mm_set_ps(0.0, 0.0, 0.0, a*a);
328 Because mulss doesn't modify the top 3 elements, the top elements of
329 xmm1 are already zero'd. We could compile this to:
335 //===---------------------------------------------------------------------===//
337 Here's a sick and twisted idea. Consider code like this:
339 __m128 test(__m128 a) {
340 float b = *(float*)&A;
342 return _mm_set_ps(0.0, 0.0, 0.0, b);
345 This might compile to this code:
347 movaps c(%esp), %xmm1
352 Now consider if the ... code caused xmm1 to get spilled. This might produce
355 movaps c(%esp), %xmm1
356 movaps %xmm1, c2(%esp)
360 movaps c2(%esp), %xmm1
364 However, since the reload is only used by these instructions, we could
365 "fold" it into the uses, producing something like this:
367 movaps c(%esp), %xmm1
368 movaps %xmm1, c2(%esp)
371 movss c2(%esp), %xmm0
374 ... saving two instructions.
376 The basic idea is that a reload from a spill slot, can, if only one 4-byte
377 chunk is used, bring in 3 zeros the the one element instead of 4 elements.
378 This can be used to simplify a variety of shuffle operations, where the
379 elements are fixed zeros.
381 //===---------------------------------------------------------------------===//
385 #include <emmintrin.h>
386 void test(__m128d *r, __m128d *A, double B) {
387 *r = _mm_loadl_pd(*A, &B);
393 movsd 24(%esp), %xmm0
405 movl 4(%esp), %edx #3.6
406 movl 8(%esp), %eax #3.6
407 movapd (%eax), %xmm0 #4.22
408 movlpd 12(%esp), %xmm0 #4.8
409 movapd %xmm0, (%edx) #4.3
412 So icc is smart enough to know that B is in memory so it doesn't load it and
413 store it back to stack.
415 This should be fixed by eliminating the llvm.x86.sse2.loadl.pd intrinsic,
416 lowering it to a load+insertelement instead. Already match the load+shuffle
417 as movlpd, so this should be easy. We already get optimal code for:
419 define void @test2(<2 x double>* %r, <2 x double>* %A, double %B) {
421 %tmp2 = load <2 x double>* %A, align 16
422 %tmp8 = insertelement <2 x double> %tmp2, double %B, i32 0
423 store <2 x double> %tmp8, <2 x double>* %r, align 16
427 //===---------------------------------------------------------------------===//
429 __m128d test1( __m128d A, __m128d B) {
430 return _mm_shuffle_pd(A, B, 0x3);
435 shufpd $3, %xmm1, %xmm0
437 Perhaps it's better to use unpckhpd instead?
439 unpckhpd %xmm1, %xmm0
441 Don't know if unpckhpd is faster. But it is shorter.
443 //===---------------------------------------------------------------------===//
445 This code generates ugly code, probably due to costs being off or something:
447 define void @test(float* %P, <4 x float>* %P2 ) {
448 %xFloat0.688 = load float* %P
449 %tmp = load <4 x float>* %P2
450 %inFloat3.713 = insertelement <4 x float> %tmp, float 0.0, i32 3
451 store <4 x float> %inFloat3.713, <4 x float>* %P2
462 shufps $50, %xmm1, %xmm2
463 shufps $132, %xmm2, %xmm0
467 Would it be better to generate:
473 pinsrw $6, %eax, %xmm0
474 pinsrw $7, %eax, %xmm0
480 //===---------------------------------------------------------------------===//
482 Some useful information in the Apple Altivec / SSE Migration Guide:
484 http://developer.apple.com/documentation/Performance/Conceptual/
485 Accelerate_sse_migration/index.html
487 e.g. SSE select using and, andnot, or. Various SSE compare translations.
489 //===---------------------------------------------------------------------===//
491 Add hooks to commute some CMPP operations.
493 //===---------------------------------------------------------------------===//
495 Apply the same transformation that merged four float into a single 128-bit load
496 to loads from constant pool.
498 //===---------------------------------------------------------------------===//
500 Floating point max / min are commutable when -enable-unsafe-fp-path is
501 specified. We should turn int_x86_sse_max_ss and X86ISD::FMIN etc. into other
502 nodes which are selected to max / min instructions that are marked commutable.
504 //===---------------------------------------------------------------------===//
506 We should compile this:
507 #include <xmmintrin.h>
513 void swizzle (const void *a, vector4_t * b, vector4_t * c) {
514 b->v = _mm_loadl_pi (b->v, (__m64 *) a);
515 c->v = _mm_loadl_pi (c->v, ((__m64 *) a) + 1);
526 movlps 8(%eax), %xmm0
540 movlps 8(%ecx), %xmm0
544 //===---------------------------------------------------------------------===//
546 These functions should produce the same code:
548 #include <emmintrin.h>
550 typedef long long __m128i __attribute__ ((__vector_size__ (16)));
552 int foo(__m128i* val) {
553 return __builtin_ia32_vec_ext_v4si(*val, 1);
555 int bar(__m128i* val) {
563 We currently produce (with -m64):
566 pshufd $1, (%rdi), %xmm0
573 //===---------------------------------------------------------------------===//
575 We should materialize vector constants like "all ones" and "signbit" with
578 cmpeqps xmm1, xmm1 ; xmm1 = all-ones
581 cmpeqps xmm1, xmm1 ; xmm1 = all-ones
582 psrlq xmm1, 31 ; xmm1 = all 100000000000...
584 instead of using a load from the constant pool. The later is important for
585 ABS/NEG/copysign etc.
587 //===---------------------------------------------------------------------===//
591 #include <xmmintrin.h>
593 void x(unsigned short n) {
594 a = _mm_slli_epi32 (a, n);
597 a = _mm_slli_epi32 (a, n);
600 compile to ( -O3 -static -fomit-frame-pointer):
615 "y" looks good, but "x" does silly movzwl stuff around into a GPR. It seems
616 like movd would be sufficient in both cases as the value is already zero
617 extended in the 32-bit stack slot IIRC. For signed short, it should also be
618 save, as a really-signed value would be undefined for pslld.
621 //===---------------------------------------------------------------------===//
624 int t1(double d) { return signbit(d); }
626 This currently compiles to:
628 movsd 16(%esp), %xmm0
635 We should use movmskp{s|d} instead.
637 //===---------------------------------------------------------------------===//
639 CodeGen/X86/vec_align.ll tests whether we can turn 4 scalar loads into a single
640 (aligned) vector load. This functionality has a couple of problems.
642 1. The code to infer alignment from loads of globals is in the X86 backend,
643 not the dag combiner. This is because dagcombine2 needs to be able to see
644 through the X86ISD::Wrapper node, which DAGCombine can't really do.
645 2. The code for turning 4 x load into a single vector load is target
646 independent and should be moved to the dag combiner.
647 3. The code for turning 4 x load into a vector load can only handle a direct
648 load from a global or a direct load from the stack. It should be generalized
649 to handle any load from P, P+4, P+8, P+12, where P can be anything.
650 4. The alignment inference code cannot handle loads from globals in non-static
651 mode because it doesn't look through the extra dyld stub load. If you try
652 vec_align.ll without -relocation-model=static, you'll see what I mean.
654 //===---------------------------------------------------------------------===//
656 We should lower store(fneg(load p), q) into an integer load+xor+store, which
657 eliminates a constant pool load. For example, consider:
659 define i64 @ccosf(float %z.0, float %z.1) nounwind readonly {
661 %tmp6 = sub float -0.000000e+00, %z.1 ; <float> [#uses=1]
662 %tmp20 = tail call i64 @ccoshf( float %tmp6, float %z.0 ) nounwind readonly ; <i64> [#uses=1]
666 This currently compiles to:
668 LCPI1_0: # <4 x float>
669 .long 2147483648 # float -0
670 .long 2147483648 # float -0
671 .long 2147483648 # float -0
672 .long 2147483648 # float -0
675 movss 16(%esp), %xmm0
677 movss 20(%esp), %xmm0
684 Note the load into xmm0, then xor (to negate), then store. In PIC mode,
685 this code computes the pic base and does two loads to do the constant pool
686 load, so the improvement is much bigger.
688 The tricky part about this xform is that the argument load/store isn't exposed
689 until post-legalize, and at that point, the fneg has been custom expanded into
690 an X86 fxor. This means that we need to handle this case in the x86 backend
691 instead of in target independent code.
693 //===---------------------------------------------------------------------===//
695 Non-SSE4 insert into 16 x i8 is atrociously bad.
697 //===---------------------------------------------------------------------===//
699 <2 x i64> extract is substantially worse than <2 x f64>, even if the destination
702 //===---------------------------------------------------------------------===//
704 SSE4 extract-to-mem ops aren't being pattern matched because of the AssertZext
705 sitting between the truncate and the extract.
707 //===---------------------------------------------------------------------===//
709 INSERTPS can match any insert (extract, imm1), imm2 for 4 x float, and insert
710 any number of 0.0 simultaneously. Currently we only use it for simple
713 See comments in LowerINSERT_VECTOR_ELT_SSE4.
715 //===---------------------------------------------------------------------===//
717 On a random note, SSE2 should declare insert/extract of 2 x f64 as legal, not
718 Custom. All combinations of insert/extract reg-reg, reg-mem, and mem-reg are
719 legal, it'll just take a few extra patterns written in the .td file.
721 Note: this is not a code quality issue; the custom lowered code happens to be
722 right, but we shouldn't have to custom lower anything. This is probably related
723 to <2 x i64> ops being so bad.
725 //===---------------------------------------------------------------------===//
727 'select' on vectors and scalars could be a whole lot better. We currently
728 lower them to conditional branches. On x86-64 for example, we compile this:
730 double test(double a, double b, double c, double d) { return a<b ? c : d; }
752 For unpredictable branches, the later is much more efficient. This should
753 just be a matter of having scalar sse map to SELECT_CC and custom expanding