Sanjay Patel | bab5d6c | 2015-09-20 15:58:00 +0000 | [diff] [blame] | 1 | ; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=sse2 < %s | FileCheck %s |
| 2 | |
| 3 | ; PR22428: https://llvm.org/bugs/show_bug.cgi?id=22428 |
| 4 | ; f1, f2, f3, and f4 should use an integer logic instruction. |
| 5 | ; f9 and f10 should use an FP (SSE) logic instruction. |
| 6 | ; |
| 7 | ; f5, f6, f7, and f8 are less clear. |
| 8 | ; |
| 9 | ; For f5 and f6, we can save a register move by using an FP logic instruction, |
| 10 | ; but we may need to calculate the relative costs of an SSE op vs. int op vs. |
| 11 | ; scalar <-> SSE register moves. |
| 12 | ; |
| 13 | ; For f7 and f8, the SSE instructions don't take immediate operands, so if we |
| 14 | ; use one of those, we either have to load a constant from memory or move the |
| 15 | ; scalar immediate value from an integer register over to an SSE register. |
| 16 | ; Optimizing for size may affect that decision. Also, note that there are no |
| 17 | ; scalar versions of the FP logic ops, so if we want to fold a load into a |
| 18 | ; logic op, we have to load or splat a 16-byte vector constant. |
| 19 | |
| 20 | ; 1 FP operand, 1 int operand, int result |
| 21 | |
| 22 | define i32 @f1(float %x, i32 %y) { |
| 23 | ; CHECK-LABEL: f1: |
| 24 | ; CHECK: # BB#0: |
| 25 | ; CHECK-NEXT: movd %xmm0, %eax |
| 26 | ; CHECK-NEXT: andl %edi, %eax |
| 27 | ; CHECK-NEXT: retq |
| 28 | |
| 29 | %bc1 = bitcast float %x to i32 |
| 30 | %and = and i32 %bc1, %y |
| 31 | ret i32 %and |
| 32 | } |
| 33 | |
| 34 | ; Swap operands of the logic op. |
| 35 | |
| 36 | define i32 @f2(float %x, i32 %y) { |
| 37 | ; CHECK-LABEL: f2: |
| 38 | ; CHECK: # BB#0: |
| 39 | ; CHECK-NEXT: movd %xmm0, %eax |
| 40 | ; CHECK-NEXT: andl %edi, %eax |
| 41 | ; CHECK-NEXT: retq |
| 42 | |
| 43 | %bc1 = bitcast float %x to i32 |
| 44 | %and = and i32 %y, %bc1 |
| 45 | ret i32 %and |
| 46 | } |
| 47 | |
| 48 | ; 1 FP operand, 1 constant operand, int result |
| 49 | |
| 50 | define i32 @f3(float %x) { |
| 51 | ; CHECK-LABEL: f3: |
| 52 | ; CHECK: # BB#0: |
| 53 | ; CHECK-NEXT: movd %xmm0, %eax |
| 54 | ; CHECK-NEXT: andl $1, %eax |
| 55 | ; CHECK-NEXT: retq |
| 56 | |
| 57 | %bc1 = bitcast float %x to i32 |
| 58 | %and = and i32 %bc1, 1 |
| 59 | ret i32 %and |
| 60 | } |
| 61 | |
| 62 | ; Swap operands of the logic op. |
| 63 | |
| 64 | define i32 @f4(float %x) { |
| 65 | ; CHECK-LABEL: f4: |
| 66 | ; CHECK: # BB#0: |
| 67 | ; CHECK-NEXT: movd %xmm0, %eax |
| 68 | ; CHECK-NEXT: andl $2, %eax |
| 69 | ; CHECK-NEXT: retq |
| 70 | |
| 71 | %bc1 = bitcast float %x to i32 |
| 72 | %and = and i32 2, %bc1 |
| 73 | ret i32 %and |
| 74 | } |
| 75 | |
| 76 | ; 1 FP operand, 1 integer operand, FP result |
| 77 | |
| 78 | define float @f5(float %x, i32 %y) { |
| 79 | ; CHECK-LABEL: f5: |
| 80 | ; CHECK: # BB#0: |
| 81 | ; CHECK-NEXT: movd %xmm0, %eax |
| 82 | ; CHECK-NEXT: andl %edi, %eax |
| 83 | ; CHECK-NEXT: movd %eax, %xmm0 |
| 84 | ; CHECK-NEXT: retq |
| 85 | |
| 86 | %bc1 = bitcast float %x to i32 |
| 87 | %and = and i32 %bc1, %y |
| 88 | %bc2 = bitcast i32 %and to float |
| 89 | ret float %bc2 |
| 90 | } |
| 91 | |
| 92 | ; Swap operands of the logic op. |
| 93 | |
| 94 | define float @f6(float %x, i32 %y) { |
| 95 | ; CHECK-LABEL: f6: |
| 96 | ; CHECK: # BB#0: |
| 97 | ; CHECK-NEXT: movd %xmm0, %eax |
| 98 | ; CHECK-NEXT: andl %edi, %eax |
| 99 | ; CHECK-NEXT: movd %eax, %xmm0 |
| 100 | ; CHECK-NEXT: retq |
| 101 | |
| 102 | %bc1 = bitcast float %x to i32 |
| 103 | %and = and i32 %y, %bc1 |
| 104 | %bc2 = bitcast i32 %and to float |
| 105 | ret float %bc2 |
| 106 | } |
| 107 | |
| 108 | ; 1 FP operand, 1 constant operand, FP result |
| 109 | |
| 110 | define float @f7(float %x) { |
| 111 | ; CHECK-LABEL: f7: |
| 112 | ; CHECK: # BB#0: |
Sanjay Patel | 309c4f9 | 2015-10-27 01:28:07 +0000 | [diff] [blame] | 113 | ; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero |
| 114 | ; CHECK-NEXT: andps %xmm1, %xmm0 |
Sanjay Patel | bab5d6c | 2015-09-20 15:58:00 +0000 | [diff] [blame] | 115 | ; CHECK-NEXT: retq |
| 116 | |
| 117 | %bc1 = bitcast float %x to i32 |
| 118 | %and = and i32 %bc1, 3 |
| 119 | %bc2 = bitcast i32 %and to float |
| 120 | ret float %bc2 |
| 121 | } |
| 122 | |
| 123 | ; Swap operands of the logic op. |
| 124 | |
| 125 | define float @f8(float %x) { |
| 126 | ; CHECK-LABEL: f8: |
| 127 | ; CHECK: # BB#0: |
Sanjay Patel | 309c4f9 | 2015-10-27 01:28:07 +0000 | [diff] [blame] | 128 | ; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero |
| 129 | ; CHECK-NEXT: andps %xmm1, %xmm0 |
Sanjay Patel | bab5d6c | 2015-09-20 15:58:00 +0000 | [diff] [blame] | 130 | ; CHECK-NEXT: retq |
| 131 | |
| 132 | %bc1 = bitcast float %x to i32 |
| 133 | %and = and i32 4, %bc1 |
| 134 | %bc2 = bitcast i32 %and to float |
| 135 | ret float %bc2 |
| 136 | } |
| 137 | |
| 138 | ; 2 FP operands, int result |
| 139 | |
| 140 | define i32 @f9(float %x, float %y) { |
| 141 | ; CHECK-LABEL: f9: |
| 142 | ; CHECK: # BB#0: |
Sanjay Patel | df2495f | 2015-09-23 17:00:06 +0000 | [diff] [blame] | 143 | ; CHECK-NEXT: andps %xmm1, %xmm0 |
| 144 | ; CHECK-NEXT: movd %xmm0, %eax |
Sanjay Patel | bab5d6c | 2015-09-20 15:58:00 +0000 | [diff] [blame] | 145 | ; CHECK-NEXT: retq |
| 146 | |
| 147 | %bc1 = bitcast float %x to i32 |
| 148 | %bc2 = bitcast float %y to i32 |
| 149 | %and = and i32 %bc1, %bc2 |
| 150 | ret i32 %and |
| 151 | } |
| 152 | |
| 153 | ; 2 FP operands, FP result |
| 154 | |
| 155 | define float @f10(float %x, float %y) { |
| 156 | ; CHECK-LABEL: f10: |
| 157 | ; CHECK: # BB#0: |
Sanjay Patel | df2495f | 2015-09-23 17:00:06 +0000 | [diff] [blame] | 158 | ; CHECK-NEXT: andps %xmm1, %xmm0 |
Sanjay Patel | bab5d6c | 2015-09-20 15:58:00 +0000 | [diff] [blame] | 159 | ; CHECK-NEXT: retq |
| 160 | |
| 161 | %bc1 = bitcast float %x to i32 |
| 162 | %bc2 = bitcast float %y to i32 |
| 163 | %and = and i32 %bc1, %bc2 |
| 164 | %bc3 = bitcast i32 %and to float |
| 165 | ret float %bc3 |
| 166 | } |
| 167 | |
| 168 | define float @or(float %x, float %y) { |
| 169 | ; CHECK-LABEL: or: |
| 170 | ; CHECK: # BB#0: |
Sanjay Patel | aba3755 | 2015-09-23 18:19:07 +0000 | [diff] [blame] | 171 | ; CHECK-NEXT: orps %xmm1, %xmm0 |
Sanjay Patel | bab5d6c | 2015-09-20 15:58:00 +0000 | [diff] [blame] | 172 | ; CHECK-NEXT: retq |
| 173 | |
| 174 | %bc1 = bitcast float %x to i32 |
| 175 | %bc2 = bitcast float %y to i32 |
| 176 | %and = or i32 %bc1, %bc2 |
| 177 | %bc3 = bitcast i32 %and to float |
| 178 | ret float %bc3 |
| 179 | } |
| 180 | |
| 181 | define float @xor(float %x, float %y) { |
| 182 | ; CHECK-LABEL: xor: |
| 183 | ; CHECK: # BB#0: |
Sanjay Patel | 1a65346 | 2015-09-23 18:33:42 +0000 | [diff] [blame] | 184 | ; CHECK-NEXT: xorps %xmm1, %xmm0 |
Sanjay Patel | bab5d6c | 2015-09-20 15:58:00 +0000 | [diff] [blame] | 185 | ; CHECK-NEXT: retq |
| 186 | |
| 187 | %bc1 = bitcast float %x to i32 |
| 188 | %bc2 = bitcast float %y to i32 |
| 189 | %and = xor i32 %bc1, %bc2 |
| 190 | %bc3 = bitcast i32 %and to float |
| 191 | ret float %bc3 |
| 192 | } |
| 193 | |
Sanjay Patel | 28d1598 | 2015-10-26 23:52:42 +0000 | [diff] [blame] | 194 | define float @f7_or(float %x) { |
| 195 | ; CHECK-LABEL: f7_or: |
| 196 | ; CHECK: # BB#0: |
Sanjay Patel | 309c4f9 | 2015-10-27 01:28:07 +0000 | [diff] [blame] | 197 | ; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero |
| 198 | ; CHECK-NEXT: orps %xmm1, %xmm0 |
Sanjay Patel | 28d1598 | 2015-10-26 23:52:42 +0000 | [diff] [blame] | 199 | ; CHECK-NEXT: retq |
| 200 | |
| 201 | %bc1 = bitcast float %x to i32 |
| 202 | %and = or i32 %bc1, 3 |
| 203 | %bc2 = bitcast i32 %and to float |
| 204 | ret float %bc2 |
| 205 | } |
| 206 | |
| 207 | define float @f7_xor(float %x) { |
| 208 | ; CHECK-LABEL: f7_xor: |
| 209 | ; CHECK: # BB#0: |
Sanjay Patel | 309c4f9 | 2015-10-27 01:28:07 +0000 | [diff] [blame] | 210 | ; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero |
| 211 | ; CHECK-NEXT: xorps %xmm1, %xmm0 |
Sanjay Patel | 28d1598 | 2015-10-26 23:52:42 +0000 | [diff] [blame] | 212 | ; CHECK-NEXT: retq |
| 213 | |
| 214 | %bc1 = bitcast float %x to i32 |
| 215 | %and = xor i32 %bc1, 3 |
| 216 | %bc2 = bitcast i32 %and to float |
| 217 | ret float %bc2 |
| 218 | } |
| 219 | |
Sanjay Patel | bab5d6c | 2015-09-20 15:58:00 +0000 | [diff] [blame] | 220 | ; Make sure that doubles work too. |
| 221 | |
| 222 | define double @doubles(double %x, double %y) { |
| 223 | ; CHECK-LABEL: doubles: |
| 224 | ; CHECK: # BB#0: |
Sanjay Patel | df2495f | 2015-09-23 17:00:06 +0000 | [diff] [blame] | 225 | ; CHECK-NEXT: andpd %xmm1, %xmm0 |
Sanjay Patel | bab5d6c | 2015-09-20 15:58:00 +0000 | [diff] [blame] | 226 | ; CHECK-NEXT: retq |
| 227 | |
| 228 | %bc1 = bitcast double %x to i64 |
| 229 | %bc2 = bitcast double %y to i64 |
| 230 | %and = and i64 %bc1, %bc2 |
| 231 | %bc3 = bitcast i64 %and to double |
| 232 | ret double %bc3 |
| 233 | } |
| 234 | |
Sanjay Patel | 28d1598 | 2015-10-26 23:52:42 +0000 | [diff] [blame] | 235 | define double @f7_double(double %x) { |
| 236 | ; CHECK-LABEL: f7_double: |
| 237 | ; CHECK: # BB#0: |
Sanjay Patel | 309c4f9 | 2015-10-27 01:28:07 +0000 | [diff] [blame] | 238 | ; CHECK-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero |
| 239 | ; CHECK-NEXT: andpd %xmm1, %xmm0 |
Sanjay Patel | 28d1598 | 2015-10-26 23:52:42 +0000 | [diff] [blame] | 240 | ; CHECK-NEXT: retq |
| 241 | |
| 242 | %bc1 = bitcast double %x to i64 |
| 243 | %and = and i64 %bc1, 3 |
| 244 | %bc2 = bitcast i64 %and to double |
| 245 | ret double %bc2 |
| 246 | } |
| 247 | |
| 248 | ; Grabbing the sign bit is a special case that could be handled |
| 249 | ; by movmskps/movmskpd, but if we're not shifting it over, then |
| 250 | ; a simple FP logic op is cheaper. |
| 251 | |
| 252 | define float @movmsk(float %x) { |
| 253 | ; CHECK-LABEL: movmsk: |
| 254 | ; CHECK: # BB#0: |
Sanjay Patel | 309c4f9 | 2015-10-27 01:28:07 +0000 | [diff] [blame] | 255 | ; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero |
| 256 | ; CHECK-NEXT: andps %xmm1, %xmm0 |
Sanjay Patel | 28d1598 | 2015-10-26 23:52:42 +0000 | [diff] [blame] | 257 | ; CHECK-NEXT: retq |
| 258 | |
| 259 | %bc1 = bitcast float %x to i32 |
| 260 | %and = and i32 %bc1, 2147483648 |
| 261 | %bc2 = bitcast i32 %and to float |
| 262 | ret float %bc2 |
| 263 | } |
| 264 | |