The AArch64 provides a handful of branchless conditional instructions.<\/p>\n
First up are the conditional assignments.<\/p>\n
; condition select\r\n ; Rd = cond ? Rn : Rm\r\n csel Rd\/zr, Rn\/zr, Rm\/zr, cond\r\n\r\n ; conditional select invert\r\n ; Rd = cond ? Rn : ~Rm\r\n csinv Rd\/zr, Rn\/zr, Rm\/zr, cond\r\n\r\n ; conditional select negate\r\n ; Rd = cond ? Rn : -Rm\r\n csneg Rd\/zr, Rn\/zr, Rm\/zr, cond\r\n\r\n ; conditional select increment\r\n ; Rd = cond ? Rn : (Rm + 1)\r\n csinc Rd\/zr, Rn\/zr, Rm\/zr, cond\r\n<\/pre>\nThese operations assign a value based on a condition. If the condition is met, then the first input operand is assigned to the destination. Otherwise, some function of the second input operand is assigned.<\/p>\n
The condition is any of the same condition codes used by the conditional branch instruction.<\/p>\n
By passing the same register as both input operands, you get some interesting pseudo-instructions:<\/p>\n
; conditional invert\r\n ; Rd = cond ? Rn : ~Rn \r\n cinv Rd\/zr, Rn\/zr, cond ; csinv Rd, Rn, Rn, cond\r\n\r\n ; conditional increment\r\n ; Rd = cond ? (Rn + 1) : Rn\r\n cinc Rd\/zr, Rn\/zr, cond ; csinc Rd, Rn, Rn, !cond\r\n\r\n ; conditional negate\r\n ; Rd = cond ? Rn : -Rn\r\n cneg Rd\/zr, Rn\/zr, cond ; csneg Rd, Rn, Rn, !cond\r\n<\/pre>\nSince the interesting operation occurs to the second input operand, we have to reverse the sense of the condition. (The assembler doesn’t accept
!<\/code> to negate the condition. You’ll have to write it out by hand.)<\/p>\nFinally, we get some interesting pseudo-instructions if we hard-code both input registers to zero.<\/p>\n
; conditional set\r\n ; Rd = cond ? 1 : 0\r\n cset Rd\/zr, cond ; csinc Rd, zr, zr, !cond\r\n\r\n ; conditional set mask\r\n ; Rd = cond ? -1 : 0 ; -1 is all bits set\r\n csetm Rd\/zr, cond ; csinv Rd, zr, zr, !cond\r\n<\/pre>\nThe next set of conditional operations is the conditional comparisons, which let you combine the results of multiple comparisons so you can perform a single test at the end.<\/p>\n
Recall that Itanium accomplished this by predicating a comparison instruction<\/a>, which had the effect of accumulating (either by AND or OR) multiple predicates into a single predicate register. And PowerPC did this by having eight sets of flags on which you can perform boolean operations<\/a>, so that you can combine the flags in the way you like to produce a single result bit at the end.<\/p>\n
AArch64 does it by letting you make a comparison instruction conditional and also specify the artificial result if the condition is not met.<\/p>\n
; conditional compare\r\n ; if (cond) then set flags as if \"cmp a, b\"\r\n ; else set flags to #nzcv\r\n ccmp Rd\/zr, #imm5, #nzcv, cond\r\n ccmp Rd\/zr, Rn\/zr, #nzcv, cond\r\n\r\n ; conditional compare negative\r\n ; if (cond) then set flags as if \"cmn a, b\"\r\n ; else set flags to #nzcv\r\n ccmn Rd\/zr, #imm5, #nzcv, cond\r\n ccmn Rd\/zr, Rn\/zr, #nzcv, cond\r\n<\/pre>\nThe immediate is an unsigned 5-bit value, so it can cover the range 0 \u2026 31.<\/p>\n
If the condition is met, then the flags are set according to the underlying comparison instruction. And if the condition is not met, then the flags are set to the bits you specify. The flags are expressed as a 4-bit value, corresponding to this arrangement of the flag bits:<\/p>\n
\n\n
\n N<\/td>\n Z<\/td>\n C<\/td>\n V<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n The pattern for combining two results via AND is<\/p>\n
; branch if a1 op1 b1 && a2 op2 b2\r\n\r\n cmp a1, b1\r\n ccmp a2, b2, #op2-fail, op1\r\n bop2 both_true\r\n<\/pre>\nYou start with the first comparison. Then you follow up with a
CCMP<\/code> where the condition is the thing you want the first comparison to be. The register operands are the arguments to the second comparison. And thenzvc<\/code> value is chosen so that it fails theBop2<\/code>.<\/p>\nFor example,<\/p>\n
; branch if r0 ge 0 and r1 lt 5\r\n cmp r0, #0\r\n ccmp r1, #5, #0, ge\r\n blt both_true\r\n<\/pre>\nLet’s walk through this code. The important aspect of the magic value
#0<\/code> is that it corresponds to N<\/var> = 0 and V<\/var> = 0, which is the flags result of a comparison that reports “greater than or equal to”. (You can consult the condition chart from last time to see what each condition tests.)<\/p>\n\n\n
\n Instruction<\/th>\n Flags<\/th>\n<\/tr>\n \n If r0 \u2265 0<\/var><\/th>\n If r0 < 0<\/var><\/th>\n<\/tr>\n \n cmp r0, #0<\/code><\/td>\nGE<\/td>\n LT<\/td>\n<\/tr>\n \n ccmp r1, #5, #0, ge<\/code><\/td>\ncmp r1, #5<\/code><\/td>\nGE<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n If the first comparison results in
GE<\/code>, then we perform the second comparison, and if it results inLT<\/code> then we branch, satisfied that both conditions were met.<\/p>\nIf the first comparison does not produce
GE<\/code>, then we force the nzcv<\/var> to zero, which acts likeGE<\/code>, and do not perform the second comparison. We just force it to fail. The branch fails, because we forced the flags to the opposite ofLT<\/code>.<\/p>\nSimilarly, the pattern for combining two comparisons via OR is<\/p>\n
; branch if a1 op1 b1 && a2 op2 b2\r\n\r\n cmp a1, b1\r\n ccmp a2, b2, #op2-succeed, !op1\r\n bop2 either_true\r\n<\/pre>\nIf the first comparison is not the desired
op1<\/code>, then we try again with the second comparison. But if the first comparison was what we wanted, then we force the flags to be something that causes the conditional branch to succeed.<\/p>\nThis strikes me as a clever solution for allowing multiple conditions to be combined and tested with a single conditional branch at the end, and therefore consume only a single branch prediction slot. It gives you the results in a single flags register, rather than having to create multiple flags registers or predicates and then invent instructions that combine them. It works only for straight-line expressions (not things like
(a && b) || (c && d)<\/code>), but that’s probably good enough.<\/p>\nBonus chatter<\/b>: The Windows debugger disassembles these instructions differently from how they are listed in the ARM reference manual. Instead of putting the condition at the end of the instruction, the condition is appended to the opcode.<\/p>\n
csel w0, w8, wzr, eq ; ARM reference manual\r\n cseleq w0, w8, wzr ; Windows debugger\r\n\r\n ccmp x0, #0x1c, #0, le ; ARM reference manual\r\n ccmple x0, #0x1c, #0 ; Windows debugger\r\n<\/pre>\n","protected":false},"excerpt":{"rendered":"Making decisions.<\/p>\n","protected":false},"author":1069,"featured_media":111744,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[1],"tags":[2],"class_list":["post-106998","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-oldnewthing","tag-history"],"acf":[],"blog_post_summary":"
Making decisions.<\/p>\n","_links":{"self":[{"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/posts\/106998","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/users\/1069"}],"replies":[{"embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/comments?post=106998"}],"version-history":[{"count":0,"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/posts\/106998\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/media\/111744"}],"wp:attachment":[{"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/media?parent=106998"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/categories?post=106998"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/oldnewthing\/wp-json\/wp\/v2\/tags?post=106998"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}