optimize ctpop codegen by width and narrow ctpop intrinsic in IR #43033
Description
| Bugzilla Link | 43688 |
| Version | trunk |
| OS | All |
| CC | @davidbolvansky,@RKSimon |
Extended Description
Spinning this off from bug 43656:
define i32 @zpop(i8 %x) {
%z = zext i8 %x to i32
%pop = tail call i32 @llvm.ctpop.i32(i32 %z)
ret i32 %pop
}
define i32 @popz(i8 %x) {
%pop = tail call i8 @llvm.ctpop.i8(i8 %x)
%z = zext i8 %pop to i32
ret i32 %z
}
declare i8 @llvm.ctpop.i8(i8)
declare i32 @llvm.ctpop.i32(i32)These are equivalent, so we should try to canonicalize them in IR. The narrow call is likely better for vectorization and would line up with our transforms of most math/logic ops.
But we don't have DAGCombiner and/or legalization to ensure that the narrow call is optimized in codegen.
For example on base x86-64:
_zpop: ## @zpop
movzbl %dil, %eax
movl %eax, %ecx
shrl %ecx
andl $-43, %ecx
subl %ecx, %eax
movl %eax, %ecx
andl $858993459, %ecx ## imm = 0x33333333
shrl $2, %eax
andl $858993459, %eax ## imm = 0x33333333
addl %ecx, %eax
movl %eax, %ecx
shrl $4, %ecx
addl %eax, %ecx
andl $252645135, %ecx ## imm = 0xF0F0F0F
imull $16843009, %ecx, %eax ## imm = 0x1010101
shrl $24, %eax
retq
_popz: ## @popz
movl %edi, %eax
shrb %al
andb $85, %al
subb %al, %dil
movl %edi, %eax
andb $51, %al
shrb $2, %dil
andb $51, %dil
addb %al, %dil
movl %edi, %eax
shrb $4, %al
addb %dil, %al
andb $15, %al
movzbl %al, %eax
retq