1 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s
3 ; The fundamental problem: an add separated from other arithmetic by a sext can't
4 ; be combined with the later instructions. However, if the first add is 'nsw',
5 ; then we can promote the sext ahead of that add to allow optimizations.
7 define i64 @add_nsw_consts(i32 %i) {
8 ; CHECK-LABEL: add_nsw_consts:
10 ; CHECK-NEXT: movslq %edi, %rax
11 ; CHECK-NEXT: addq $12, %rax
14 %add = add nsw i32 %i, 5
15 %ext = sext i32 %add to i64
16 %idx = add i64 %ext, 7
20 ; An x86 bonus: If we promote the sext ahead of the 'add nsw',
21 ; we allow LEA formation and eliminate an add instruction.
23 define i64 @add_nsw_sext_add(i32 %i, i64 %x) {
24 ; CHECK-LABEL: add_nsw_sext_add:
26 ; CHECK-NEXT: movslq %edi, %rax
27 ; CHECK-NEXT: leaq 5(%rax,%rsi), %rax
30 %add = add nsw i32 %i, 5
31 %ext = sext i32 %add to i64
32 %idx = add i64 %x, %ext
36 ; Throw in a scale (left shift) because an LEA can do that too.
37 ; Use a negative constant (LEA displacement) to verify that's handled correctly.
39 define i64 @add_nsw_sext_lsh_add(i32 %i, i64 %x) {
40 ; CHECK-LABEL: add_nsw_sext_lsh_add:
42 ; CHECK-NEXT: movslq %edi, %rax
43 ; CHECK-NEXT: leaq -40(%rsi,%rax,8), %rax
46 %add = add nsw i32 %i, -5
47 %ext = sext i32 %add to i64
48 %shl = shl i64 %ext, 3
49 %idx = add i64 %x, %shl
53 ; Don't promote the sext if it has no users. The wider add instruction needs an
54 ; extra byte to encode.
56 define i64 @add_nsw_sext(i32 %i, i64 %x) {
57 ; CHECK-LABEL: add_nsw_sext:
59 ; CHECK-NEXT: addl $5, %edi
60 ; CHECK-NEXT: movslq %edi, %rax
63 %add = add nsw i32 %i, 5
64 %ext = sext i32 %add to i64
68 ; The typical use case: a 64-bit system where an 'int' is used as an index into an array.
70 define i8* @gep8(i32 %i, i8* %x) {
73 ; CHECK-NEXT: movslq %edi, %rax
74 ; CHECK-NEXT: leaq 5(%rax,%rsi), %rax
77 %add = add nsw i32 %i, 5
78 %ext = sext i32 %add to i64
79 %idx = getelementptr i8, i8* %x, i64 %ext
83 define i16* @gep16(i32 %i, i16* %x) {
86 ; CHECK-NEXT: movslq %edi, %rax
87 ; CHECK-NEXT: leaq -10(%rsi,%rax,2), %rax
90 %add = add nsw i32 %i, -5
91 %ext = sext i32 %add to i64
92 %idx = getelementptr i16, i16* %x, i64 %ext
96 define i32* @gep32(i32 %i, i32* %x) {
99 ; CHECK-NEXT: movslq %edi, %rax
100 ; CHECK-NEXT: leaq 20(%rsi,%rax,4), %rax
103 %add = add nsw i32 %i, 5
104 %ext = sext i32 %add to i64
105 %idx = getelementptr i32, i32* %x, i64 %ext
109 define i64* @gep64(i32 %i, i64* %x) {
110 ; CHECK-LABEL: gep64:
112 ; CHECK-NEXT: movslq %edi, %rax
113 ; CHECK-NEXT: leaq -40(%rsi,%rax,8), %rax
116 %add = add nsw i32 %i, -5
117 %ext = sext i32 %add to i64
118 %idx = getelementptr i64, i64* %x, i64 %ext
122 ; LEA can't scale by 16, but the adds can still be combined into an LEA.
124 define i128* @gep128(i32 %i, i128* %x) {
125 ; CHECK-LABEL: gep128:
127 ; CHECK-NEXT: movslq %edi, %rax
128 ; CHECK-NEXT: shlq $4, %rax
129 ; CHECK-NEXT: leaq 80(%rax,%rsi), %rax
132 %add = add nsw i32 %i, 5
133 %ext = sext i32 %add to i64
134 %idx = getelementptr i128, i128* %x, i64 %ext
138 ; A bigger win can be achieved when there is more than one use of the
139 ; sign extended value. In this case, we can eliminate sign extension
140 ; instructions plus use more efficient addressing modes for memory ops.
142 define void @PR20134(i32* %a, i32 %i) {
143 ; CHECK-LABEL: PR20134:
145 ; CHECK-NEXT: movslq %esi, %rax
146 ; CHECK-NEXT: movl 4(%rdi,%rax,4), %ecx
147 ; CHECK-NEXT: addl 8(%rdi,%rax,4), %ecx
148 ; CHECK-NEXT: movl %ecx, (%rdi,%rax,4)
151 %add1 = add nsw i32 %i, 1
152 %idx1 = sext i32 %add1 to i64
153 %gep1 = getelementptr i32, i32* %a, i64 %idx1
154 %load1 = load i32, i32* %gep1, align 4
156 %add2 = add nsw i32 %i, 2
157 %idx2 = sext i32 %add2 to i64
158 %gep2 = getelementptr i32, i32* %a, i64 %idx2
159 %load2 = load i32, i32* %gep2, align 4
161 %add3 = add i32 %load1, %load2
162 %idx3 = sext i32 %i to i64
163 %gep3 = getelementptr i32, i32* %a, i64 %idx3
164 store i32 %add3, i32* %gep3, align 4