1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s 3 4; The fundamental problem: an add separated from other arithmetic by a sign or 5; zero extension can't be combined with the later instructions. However, if the 6; first add is 'nsw' or 'nuw' respectively, then we can promote the extension 7; ahead of that add to allow optimizations. 8 9define i64 @add_nsw_consts(i32 %i) { 10; CHECK-LABEL: add_nsw_consts: 11; CHECK: # %bb.0: 12; CHECK-NEXT: movslq %edi, %rax 13; CHECK-NEXT: addq $12, %rax 14; CHECK-NEXT: retq 15 16 %add = add nsw i32 %i, 5 17 %ext = sext i32 %add to i64 18 %idx = add i64 %ext, 7 19 ret i64 %idx 20} 21 22; An x86 bonus: If we promote the sext ahead of the 'add nsw', 23; we allow LEA formation and eliminate an add instruction. 24 25define i64 @add_nsw_sext_add(i32 %i, i64 %x) { 26; CHECK-LABEL: add_nsw_sext_add: 27; CHECK: # %bb.0: 28; CHECK-NEXT: movslq %edi, %rax 29; CHECK-NEXT: leaq 5(%rax,%rsi), %rax 30; CHECK-NEXT: retq 31 32 %add = add nsw i32 %i, 5 33 %ext = sext i32 %add to i64 34 %idx = add i64 %x, %ext 35 ret i64 %idx 36} 37 38; Throw in a scale (left shift) because an LEA can do that too. 39; Use a negative constant (LEA displacement) to verify that's handled correctly. 40 41define i64 @add_nsw_sext_lsh_add(i32 %i, i64 %x) { 42; CHECK-LABEL: add_nsw_sext_lsh_add: 43; CHECK: # %bb.0: 44; CHECK-NEXT: movslq %edi, %rax 45; CHECK-NEXT: leaq -40(%rsi,%rax,8), %rax 46; CHECK-NEXT: retq 47 48 %add = add nsw i32 %i, -5 49 %ext = sext i32 %add to i64 50 %shl = shl i64 %ext, 3 51 %idx = add i64 %x, %shl 52 ret i64 %idx 53} 54 55; Don't promote the sext if it has no users. The wider add instruction needs an 56; extra byte to encode. 57 58define i64 @add_nsw_sext(i32 %i, i64 %x) { 59; CHECK-LABEL: add_nsw_sext: 60; CHECK: # %bb.0: 61; CHECK-NEXT: addl $5, %edi 62; CHECK-NEXT: movslq %edi, %rax 63; CHECK-NEXT: retq 64 65 %add = add nsw i32 %i, 5 66 %ext = sext i32 %add to i64 67 ret i64 %ext 68} 69 70; The typical use case: a 64-bit system where an 'int' is used as an index into an array. 71 72define i8* @gep8(i32 %i, i8* %x) { 73; CHECK-LABEL: gep8: 74; CHECK: # %bb.0: 75; CHECK-NEXT: movslq %edi, %rax 76; CHECK-NEXT: leaq 5(%rax,%rsi), %rax 77; CHECK-NEXT: retq 78 79 %add = add nsw i32 %i, 5 80 %ext = sext i32 %add to i64 81 %idx = getelementptr i8, i8* %x, i64 %ext 82 ret i8* %idx 83} 84 85define i16* @gep16(i32 %i, i16* %x) { 86; CHECK-LABEL: gep16: 87; CHECK: # %bb.0: 88; CHECK-NEXT: movslq %edi, %rax 89; CHECK-NEXT: leaq -10(%rsi,%rax,2), %rax 90; CHECK-NEXT: retq 91 92 %add = add nsw i32 %i, -5 93 %ext = sext i32 %add to i64 94 %idx = getelementptr i16, i16* %x, i64 %ext 95 ret i16* %idx 96} 97 98define i32* @gep32(i32 %i, i32* %x) { 99; CHECK-LABEL: gep32: 100; CHECK: # %bb.0: 101; CHECK-NEXT: movslq %edi, %rax 102; CHECK-NEXT: leaq 20(%rsi,%rax,4), %rax 103; CHECK-NEXT: retq 104 105 %add = add nsw i32 %i, 5 106 %ext = sext i32 %add to i64 107 %idx = getelementptr i32, i32* %x, i64 %ext 108 ret i32* %idx 109} 110 111define i64* @gep64(i32 %i, i64* %x) { 112; CHECK-LABEL: gep64: 113; CHECK: # %bb.0: 114; CHECK-NEXT: movslq %edi, %rax 115; CHECK-NEXT: leaq -40(%rsi,%rax,8), %rax 116; CHECK-NEXT: retq 117 118 %add = add nsw i32 %i, -5 119 %ext = sext i32 %add to i64 120 %idx = getelementptr i64, i64* %x, i64 %ext 121 ret i64* %idx 122} 123 124; LEA can't scale by 16, but the adds can still be combined into an LEA. 125 126define i128* @gep128(i32 %i, i128* %x) { 127; CHECK-LABEL: gep128: 128; CHECK: # %bb.0: 129; CHECK-NEXT: movslq %edi, %rax 130; CHECK-NEXT: shlq $4, %rax 131; CHECK-NEXT: leaq 80(%rax,%rsi), %rax 132; CHECK-NEXT: retq 133 134 %add = add nsw i32 %i, 5 135 %ext = sext i32 %add to i64 136 %idx = getelementptr i128, i128* %x, i64 %ext 137 ret i128* %idx 138} 139 140; A bigger win can be achieved when there is more than one use of the 141; sign extended value. In this case, we can eliminate sign extension 142; instructions plus use more efficient addressing modes for memory ops. 143 144define void @PR20134(i32* %a, i32 %i) { 145; CHECK-LABEL: PR20134: 146; CHECK: # %bb.0: 147; CHECK-NEXT: movslq %esi, %rax 148; CHECK-NEXT: movl 4(%rdi,%rax,4), %ecx 149; CHECK-NEXT: addl 8(%rdi,%rax,4), %ecx 150; CHECK-NEXT: movl %ecx, (%rdi,%rax,4) 151; CHECK-NEXT: retq 152 153 %add1 = add nsw i32 %i, 1 154 %idx1 = sext i32 %add1 to i64 155 %gep1 = getelementptr i32, i32* %a, i64 %idx1 156 %load1 = load i32, i32* %gep1, align 4 157 158 %add2 = add nsw i32 %i, 2 159 %idx2 = sext i32 %add2 to i64 160 %gep2 = getelementptr i32, i32* %a, i64 %idx2 161 %load2 = load i32, i32* %gep2, align 4 162 163 %add3 = add i32 %load1, %load2 164 %idx3 = sext i32 %i to i64 165 %gep3 = getelementptr i32, i32* %a, i64 %idx3 166 store i32 %add3, i32* %gep3, align 4 167 ret void 168} 169 170; The same as @PR20134 but sign extension is replaced with zero extension 171define void @PR20134_zext(i32* %a, i32 %i) { 172; CHECK-LABEL: PR20134_zext: 173; CHECK: # %bb.0: 174; CHECK-NEXT: movl %esi, %eax 175; CHECK-NEXT: movl 4(%rdi,%rax,4), %ecx 176; CHECK-NEXT: addl 8(%rdi,%rax,4), %ecx 177; CHECK-NEXT: movl %ecx, (%rdi,%rax,4) 178; CHECK-NEXT: retq 179 180 %add1 = add nuw i32 %i, 1 181 %idx1 = zext i32 %add1 to i64 182 %gep1 = getelementptr i32, i32* %a, i64 %idx1 183 %load1 = load i32, i32* %gep1, align 4 184 185 %add2 = add nuw i32 %i, 2 186 %idx2 = zext i32 %add2 to i64 187 %gep2 = getelementptr i32, i32* %a, i64 %idx2 188 %load2 = load i32, i32* %gep2, align 4 189 190 %add3 = add i32 %load1, %load2 191 %idx3 = zext i32 %i to i64 192 %gep3 = getelementptr i32, i32* %a, i64 %idx3 193 store i32 %add3, i32* %gep3, align 4 194 ret void 195} 196