| Issue |
123727
|
| Summary |
[X86][AVX512] Generate `valignq` instead of `vpermi2pd`
|
| Labels |
question,
backend:X86
|
| Assignees |
abhishek-kaushik22
|
| Reporter |
abhishek-kaushik22
|
Consider this ir:
```llvm
define <8 x double> @bar(i64 %x, ptr %arr, <8 x double> %val1, <8 x double> %val2) {
entry:
%gep1 = getelementptr double, ptr %arr, i64 %x
%gepload1 = load <8 x double>, ptr %gep1, align 16
%z = add i64 %x, 8
%gep3 = getelementptr double, ptr %arr, i64 %z
%gepload3 = load <8 x double>, ptr %gep3, align 16
%shuffle1 = shufflevector <8 x double> %gepload1, <8 x double> %gepload3, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>
store <8 x double> %shuffle1, ptr %gep1, align 16
ret <8 x double> %gepload3
}
```
Currently, we generate
```asm
.LCPI0_1:
.byte 1 # 0x1
.byte 2 # 0x2
.byte 3 # 0x3
.byte 4 # 0x4
.byte 5 # 0x5
.byte 6 # 0x6
.byte 7 # 0x7
.byte 8 # 0x8
bar: # @bar
vmovupd zmm1, zmmword ptr [rsi + 8*rdi]
vmovupd zmm0, zmmword ptr [rsi + 8*rdi + 64]
vpmovsxbq zmm2, qword ptr [rip + .LCPI0_1] # zmm2 = [1,2,3,4,5,6,7,8]
vpermi2pd zmm2, zmm1, zmm0
vmovupd zmmword ptr [rsi + 8*rdi], zmm2
ret
```
but can't we generate a `valignq` instead of the full permute?
```asm
bar: # @bar
vmovupd zmm1, zmmword ptr [rsi + 8*rdi]
vmovupd zmm0, zmmword ptr [rsi + 8*rdi + 64]
valignq zmm2, zmm1, zmm0, 1
vmovupd zmmword ptr [rsi + 8*rdi], zmm2
ret
```
This saves us from saving the pattern `.LCPI0_1` in the binary.
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
