; trunc + sext + icmp ne <- not canonical
; shl + ashr + icmp ne
; add + icmp ult
-; add + icmp uge
+; add + icmp uge/ugt
; However only the simplest form (with two shifts) gets lowered best.
; ---------------------------------------------------------------------------- ;
ret i1 %tmp1
}
+; Slightly more canonical variant
+define i1 @add_ugtcmp_i16_i8(i16 %x) nounwind {
+; CHECK-LABEL: add_ugtcmp_i16_i8:
+; CHECK: // %bb.0:
+; CHECK-NEXT: add w8, w0, #128 // =128
+; CHECK-NEXT: and w8, w8, #0xffff
+; CHECK-NEXT: cmp w8, #255 // =255
+; CHECK-NEXT: cset w0, hi
+; CHECK-NEXT: ret
+ %tmp0 = add i16 %x, 128 ; 1U << (8-1)
+ %tmp1 = icmp ugt i16 %tmp0, 255 ; (1U << 8) - 1
+ ret i1 %tmp1
+}
+
; Negative tests
; ---------------------------------------------------------------------------- ;
%tmp1 = icmp uge i24 %tmp0, 256 ; 1U << 8
ret i1 %tmp1
}
+
+; Slightly more canonical variant
+define i1 @add_ugtcmp_bad_i16_i8(i16 %x) nounwind {
+; CHECK-LABEL: add_ugtcmp_bad_i16_i8:
+; CHECK: // %bb.0:
+; CHECK-NEXT: mov w0, wzr
+; CHECK-NEXT: ret
+ %tmp0 = add i16 %x, 128 ; 1U << (8-1)
+ %tmp1 = icmp ugt i16 %tmp0, -1 ; when we +1 it, it will wrap to 0
+ ret i1 %tmp1
+}
; trunc + sext + icmp eq <- not canonical
; shl + ashr + icmp eq
; add + icmp uge
-; add + icmp ult
+; add + icmp ult/ule
; However only the simplest form (with two shifts) gets lowered best.
; ---------------------------------------------------------------------------- ;
ret i1 %tmp1
}
+; Slightly more canonical variant
+define i1 @add_ulecmp_i16_i8(i16 %x) nounwind {
+; CHECK-LABEL: add_ulecmp_i16_i8:
+; CHECK: // %bb.0:
+; CHECK-NEXT: sxtb w8, w0
+; CHECK-NEXT: and w8, w8, #0xffff
+; CHECK-NEXT: cmp w8, w0, uxth
+; CHECK-NEXT: cset w0, eq
+; CHECK-NEXT: ret
+ %tmp0 = add i16 %x, 128 ; 1U << (8-1)
+ %tmp1 = icmp ule i16 %tmp0, 255 ; (1U << 8) - 1
+ ret i1 %tmp1
+}
+
; Negative tests
; ---------------------------------------------------------------------------- ;
%tmp1 = icmp ult i24 %tmp0, 256 ; 1U << 8
ret i1 %tmp1
}
+
+define i1 @add_ulecmp_bad_i16_i8(i16 %x) nounwind {
+; CHECK-LABEL: add_ulecmp_bad_i16_i8:
+; CHECK: // %bb.0:
+; CHECK-NEXT: orr w0, wzr, #0x1
+; CHECK-NEXT: ret
+ %tmp0 = add i16 %x, 128 ; 1U << (8-1)
+ %tmp1 = icmp ule i16 %tmp0, -1 ; when we +1 it, it will wrap to 0
+ ret i1 %tmp1
+}
; trunc + sext + icmp ne <- not canonical
; shl + ashr + icmp ne
; add + icmp ult
-; add + icmp uge
+; add + icmp uge/ugt
; However only the simplest form (with two shifts) gets lowered best.
; ---------------------------------------------------------------------------- ;
ret i1 %tmp1
}
+; Slightly more canonical variant
+define i1 @add_ugtcmp_i16_i8(i16 %x) nounwind {
+; X86-LABEL: add_ugtcmp_i16_i8:
+; X86: # %bb.0:
+; X86-NEXT: movl $128, %eax
+; X86-NEXT: addl {{[0-9]+}}(%esp), %eax
+; X86-NEXT: movzwl %ax, %eax
+; X86-NEXT: cmpl $255, %eax
+; X86-NEXT: seta %al
+; X86-NEXT: retl
+;
+; X64-LABEL: add_ugtcmp_i16_i8:
+; X64: # %bb.0:
+; X64-NEXT: subl $-128, %edi
+; X64-NEXT: movzwl %di, %eax
+; X64-NEXT: cmpl $255, %eax
+; X64-NEXT: seta %al
+; X64-NEXT: retq
+ %tmp0 = add i16 %x, 128 ; 1U << (8-1)
+ %tmp1 = icmp ugt i16 %tmp0, 255 ; (1U << 8) - 1
+ ret i1 %tmp1
+}
+
; Negative tests
; ---------------------------------------------------------------------------- ;
%tmp1 = icmp uge i24 %tmp0, 256 ; 1U << 8
ret i1 %tmp1
}
+
+; Slightly more canonical variant
+define i1 @add_ugtcmp_bad_i16_i8(i16 %x) nounwind {
+; CHECK-LABEL: add_ugtcmp_bad_i16_i8:
+; CHECK: # %bb.0:
+; CHECK-NEXT: xorl %eax, %eax
+; CHECK-NEXT: ret{{[l|q]}}
+ %tmp0 = add i16 %x, 128 ; 1U << (8-1)
+ %tmp1 = icmp ugt i16 %tmp0, -1 ; when we +1 it, it will wrap to 0
+ ret i1 %tmp1
+}
; trunc + sext + icmp eq <- not canonical
; shl + ashr + icmp eq
; add + icmp uge
-; add + icmp ult
+; add + icmp ult/ule
; However only the simplest form (with two shifts) gets lowered best.
; ---------------------------------------------------------------------------- ;
ret i1 %tmp1
}
+; Slightly more canonical variant
+define i1 @add_ulecmp_i16_i8(i16 %x) nounwind {
+; X86-LABEL: add_ulecmp_i16_i8:
+; X86: # %bb.0:
+; X86-NEXT: movzwl {{[0-9]+}}(%esp), %eax
+; X86-NEXT: movsbl %al, %ecx
+; X86-NEXT: cmpw %ax, %cx
+; X86-NEXT: sete %al
+; X86-NEXT: retl
+;
+; X64-LABEL: add_ulecmp_i16_i8:
+; X64: # %bb.0:
+; X64-NEXT: movsbl %dil, %eax
+; X64-NEXT: cmpw %di, %ax
+; X64-NEXT: sete %al
+; X64-NEXT: retq
+ %tmp0 = add i16 %x, 128 ; 1U << (8-1)
+ %tmp1 = icmp ule i16 %tmp0, 255 ; (1U << 8) - 1
+ ret i1 %tmp1
+}
+
; Negative tests
; ---------------------------------------------------------------------------- ;
%tmp1 = icmp ult i24 %tmp0, 256 ; 1U << 8
ret i1 %tmp1
}
+
+define i1 @add_ulecmp_bad_i16_i8(i16 %x) nounwind {
+; CHECK-LABEL: add_ulecmp_bad_i16_i8:
+; CHECK: # %bb.0:
+; CHECK-NEXT: movb $1, %al
+; CHECK-NEXT: ret{{[l|q]}}
+ %tmp0 = add i16 %x, 128 ; 1U << (8-1)
+ %tmp1 = icmp ule i16 %tmp0, -1 ; when we +1 it, it will wrap to 0
+ ret i1 %tmp1
+}