1 //===- README_X86_64.txt - Notes for X86-64 code gen ----------------------===//
3 AMD64 Optimization Manual 8.2 has some nice information about optimizing integer
4 multiplication by a constant. How much of it applies to Intel's X86-64
5 implementation? There are definite trade-offs to consider: latency vs. register
6 pressure vs. code size.
8 //===---------------------------------------------------------------------===//
10 Are we better off using branches instead of cmove to implement FP to
14 ucomiss LC0(%rip), %xmm0
15 cvttss2siq %xmm0, %rdx
17 subss LC0(%rip), %xmm0
18 movabsq $-9223372036854775808, %rax
19 cvttss2siq %xmm0, %rdx
28 movss LCPI1_0(%rip), %xmm1
29 cvttss2siq %xmm0, %rcx
32 cvttss2siq %xmm2, %rax
33 movabsq $-9223372036854775808, %rdx
39 Seems like the jb branch has high likelyhood of being taken. It would have
40 saved a few instructions.
42 //===---------------------------------------------------------------------===//
49 memset(X, b, 2*sizeof(X[0]));
53 movq _b@GOTPCREL(%rip), %rax
63 movq _X@GOTPCREL(%rip), %rdx
69 movq _b@GOTPCREL(%rip), %rax
70 movabsq $72340172838076673, %rdx
73 movq _X@GOTPCREL(%rip), %rdx
77 And the codegen is even worse for the following
78 (from http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33103):
79 void fill1(char *s, int a)
81 __builtin_memset(s, a, 15);
84 For this version, we duplicate the computation of the constant to store.
86 //===---------------------------------------------------------------------===//
88 It's not possible to reference AH, BH, CH, and DH registers in an instruction
89 requiring REX prefix. However, divb and mulb both produce results in AH. If isel
90 emits a CopyFromReg which gets turned into a movb and that can be allocated a
93 To get around this, isel emits a CopyFromReg from AX and then right shift it
94 down by 8 and truncate it. It's not pretty but it works. We need some register
95 allocation magic to make the hack go away (e.g. putting additional constraints
96 on the result of the movb).
98 //===---------------------------------------------------------------------===//
100 The x86-64 ABI for hidden-argument struct returns requires that the
101 incoming value of %rdi be copied into %rax by the callee upon return.
103 The idea is that it saves callers from having to remember this value,
104 which would often require a callee-saved register. Callees usually
105 need to keep this value live for most of their body anyway, so it
106 doesn't add a significant burden on them.
108 We currently implement this in codegen, however this is suboptimal
109 because it means that it would be quite awkward to implement the
110 optimization for callers.
112 A better implementation would be to relax the LLVM IR rules for sret
113 arguments to allow a function with an sret argument to have a non-void
114 return type, and to have the front-end to set up the sret argument value
115 as the return value of the function. The front-end could more easily
116 emit uses of the returned struct value to be in terms of the function's
117 lowered return value, and it would free non-C frontends from a
118 complication only required by a C-based ABI.
120 //===---------------------------------------------------------------------===//
122 We get a redundant zero extension for code like this:
125 int foo(unsigned x) {
138 imull $78, %edi, %eax
140 movl %eax, %eax <----
141 movq _mask@GOTPCREL(%rip), %rcx
142 movl (%rcx,%rax,4), %eax
145 imull $45, %edi, %eax
148 Before regalloc, we have:
150 %reg1025<def> = IMUL32rri8 %reg1024, 45, %EFLAGS<imp-def>
151 JMP mbb<bb2,0x203afb0>
152 Successors according to CFG: 0x203afb0 (#3)
154 bb1: 0x203af60, LLVM BB @0x1e02310, ID#2:
155 Predecessors according to CFG: 0x203aec0 (#0)
156 %reg1026<def> = IMUL32rri8 %reg1024, 78, %EFLAGS<imp-def>
157 Successors according to CFG: 0x203afb0 (#3)
159 bb2: 0x203afb0, LLVM BB @0x1e02340, ID#3:
160 Predecessors according to CFG: 0x203af10 (#1) 0x203af60 (#2)
161 %reg1027<def> = PHI %reg1025, mbb<bb,0x203af10>,
162 %reg1026, mbb<bb1,0x203af60>
163 %reg1029<def> = MOVZX64rr32 %reg1027
165 so we'd have to know that IMUL32rri8 leaves the high word zero extended and to
166 be able to recognize the zero extend. This could also presumably be implemented
167 if we have whole-function selectiondags.
169 //===---------------------------------------------------------------------===//
171 Take the following C code
172 (from http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43640):
180 float foo(struct u1 u)
185 Optimizes to the following IR:
186 define float @foo(double %u.0) nounwind readnone {
188 %tmp8 = bitcast double %u.0 to i64 ; <i64> [#uses=2]
189 %tmp6 = trunc i64 %tmp8 to i32 ; <i32> [#uses=1]
190 %tmp7 = bitcast i32 %tmp6 to float ; <float> [#uses=1]
191 %tmp2 = lshr i64 %tmp8, 32 ; <i64> [#uses=1]
192 %tmp3 = trunc i64 %tmp2 to i32 ; <i32> [#uses=1]
193 %tmp4 = bitcast i32 %tmp3 to float ; <float> [#uses=1]
194 %0 = fadd float %tmp7, %tmp4 ; <float> [#uses=1]
198 And current llvm-gcc/clang output:
206 We really shouldn't move the floats to RAX, only to immediately move them
207 straight back to the XMM registers.
209 There really isn't any good way to handle this purely in IR optimizers; it
210 could possibly be handled by changing the output of the fronted, though. It
211 would also be feasible to add a x86-specific DAGCombine to optimize the
212 bitcast+trunc+(lshr+)bitcast combination.
214 //===---------------------------------------------------------------------===//
216 Take the following code
217 (from http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34653):
218 extern unsigned long table[];
219 unsigned long foo(unsigned char *p) {
220 unsigned long tag = *p;
221 return table[tag >> 4] + table[tag & 0xf];
224 Current code generated:
230 movq table(,%rax,8), %rax
231 addq table(%rcx), %rax
235 1. First movq should be movl; saves a byte.
236 2. Both andq's should be andl; saves another two bytes. I think this was
237 implemented at one point, but subsequently regressed.
238 3. shrq should be shrl; saves another byte.
239 4. The first andq can be completely eliminated by using a slightly more
240 expensive addressing mode.
242 //===---------------------------------------------------------------------===//
244 Consider the following (contrived testcase, but contains common factors):
247 int test(int x, ...) {
251 for (i = 0; i < x; i++)
252 sum += va_arg(l, int);
257 Testcase given in C because fixing it will likely involve changing the IR
258 generated for it. The primary issue with the result is that it doesn't do any
259 of the optimizations which are possible if we know the address of a va_list
260 in the current function is never taken:
261 1. We shouldn't spill the XMM registers because we only call va_arg with "int".
262 2. It would be nice if we could scalarrepl the va_list.
263 3. Probably overkill, but it'd be cool if we could peel off the first five
264 iterations of the loop.
266 Other optimizations involving functions which use va_arg on floats which don't
267 have the address of a va_list taken:
268 1. Conversely to the above, we shouldn't spill general registers if we only
269 call va_arg on "double".
270 2. If we know nothing more than 64 bits wide is read from the XMM registers,
271 we can change the spilling code to reduce the amount of stack used by half.
273 //===---------------------------------------------------------------------===//