Raymond Toy pushed to branch master at cmucl / cmucl

Commits: d51eb4b8 by Raymond Toy at 2023-03-12T19:05:39+00:00 Fix #175: Simplify branching in x86 float compare vops

- - - - - a7237e1d by Raymond Toy at 2023-03-12T19:05:39+00:00 Merge branch 'issue-175-simplify-float-compare-vops' into 'master'

Fix #175: Simplify branching in x86 float compare vops

Closes #175

See merge request cmucl/cmucl!129 - - - - -

2 changed files:

- src/compiler/x86/float-sse2.lisp - + tests/nan.lisp

Changes:

===================================== src/compiler/x86/float-sse2.lisp ===================================== @@ -945,7 +945,7 @@ (frob double ucomisd))

(macrolet - ((frob (op size inst yep nope) + ((frob (op size inst) (let ((ea (ecase size (single 'ea-for-sf-desc) @@ -953,28 +953,52 @@ 'ea-for-df-desc))) (name (symbolicate op "/" size "-FLOAT")) (sc-type (symbolicate size "-REG")) - (inherit (symbolicate size "-FLOAT-COMPARE"))) + (inherit (symbolicate size "-FLOAT-COMPARE")) + (reverse-args-p (eq op '<))) `(define-vop (,name ,inherit) + ;; The compare instructions take a reg argument for the + ;; first arg and reg or mem argument for the second. When + ;; inverting the arguments we must also invert which of + ;; the argument can be a mem argument. + (:args (x :scs (,sc-type ,@(when reverse-args-p 'descriptor-reg))) + (y :scs (,sc-type ,@(unless reverse-args-p 'descriptor-reg)))) (:translate ,op) (:info target not-p) (:generator 3 - (sc-case y - (,sc-type - (inst ,inst x y)) - (descriptor-reg - (inst ,inst x (,ea y)))) - (cond (not-p - (inst jmp :p target) - (inst jmp ,nope target)) - (t - (let ((not-lab (gen-label))) - (inst jmp :p not-lab) - (inst jmp ,yep target) - (emit-label not-lab))))))))) - (frob < single comiss :b :nb) - (frob > single comiss :a :na) - (frob < double comisd :b :nb) - (frob > double comisd :a :na)) + ;; Note: x < y is the same as y > x. We reverse the + ;; args to reduce the number of jump instructions + ;; needed. + ,(if reverse-args-p + `(sc-case x + (,sc-type + (inst ,inst y x)) + (descriptor-reg + (inst ,inst y (,ea x)))) + `(sc-case y + (,sc-type + (inst ,inst x y)) + (descriptor-reg + (inst ,inst x (,ea y))))) + ;; Consider the case of x > y. + ;; + ;; When a NaN occurs, comis sets ZF, PF, and CF = 1. In + ;; the normal case (not-p false), we want to jump to the + ;; target when x > y. This happens when CF = 0. Hence, + ;; we won't jump to the target when there's a NaN, as + ;; desired. + ;; + ;; For the not-p case, we want to jump to target when x + ;; <= y. This means CF = 1 or ZF = 1. But NaN sets + ;; these bits too, so we jump to the target for NaN or x + ;; <= y, as desired. + ;; + ;; For the case of x < y, we can use the equivalent y > + ;; x. Thus if we swap the args, the same logic applies. + (inst jmp (if (not not-p) :a :be) target)))))) + (frob > single comiss) + (frob > double comisd) + (frob < single comiss) + (frob < double comisd))

;;;; Conversion:

===================================== tests/nan.lisp ===================================== @@ -0,0 +1,209 @@ +;;; Tests for NaN comparisons. +(defpackage :nan-tests + (:use :cl :lisp-unit)) + +(in-package :nan-tests) + +(defparameter *single-float-nan* + (ext:with-float-traps-masked (:invalid :divide-by-zero) + (/ 0d0 0d0))) + +(defparameter *double-float-nan* + (ext:with-float-traps-masked (:invalid :divide-by-zero) + (/ 0d0 0d0))) + + +(eval-when (:compile-toplevel :load-toplevel :execute) + (macrolet + ((frob (ntype op) + (let* ((name (ext:symbolicate (if (eq ntype 'single-float) + "S" + "D") + "TST-" op)) + (name3 (ext:symbolicate name "3"))) + + `(progn + (defun ,name (x y) + (declare (,ntype x y)) + (,op x y)) + (defun ,name3 (x y z) + (declare (,ntype x y z)) + (,op x y z)))))) + (frob single-float <) + (frob single-float >) + (frob double-float <) + (frob double-float >) + (frob single-float =) + (frob double-float =))) + +(define-test nan-single.< + (:tag :nan) + ;; First just make sure it works with regular single-floats + (assert-true (stst-< 1f0 2f0)) + (assert-false (stst-< 1f0 1f0)) + (assert-false (stst-< 1f0 0f0)) + ;; Now try NaN. All comparisons should be false. + (ext:with-float-traps-masked (:invalid) + (assert-false (stst-< *single-float-nan* 1f0)) + (assert-false (stst-< 1f0 *single-float-nan*)) + (assert-false (stst-< *single-float-nan* *single-float-nan*)))) + +(define-test nan-double.< + (:tag :nan) + ;; First just make sure it works with regular single-floats + (assert-true (dtst-< 1d0 2d0)) + (assert-false (dtst-< 1d0 1d0)) + (assert-false (dtst-< 1d0 0d0)) + ;; Now try NaN. All comparisons should be false. + (ext:with-float-traps-masked (:invalid) + (assert-false (dtst-< *double-float-nan* 1d0)) + (assert-false (dtst-< 1d0 *double-float-nan*)) + (assert-false (dtst-< *double-float-nan* *double-float-nan*)))) + +(define-test nan-single.> + (:tag :nan) + ;; First just make sure it works with regular single-floats + (assert-true (stst-> 2f0 1f0)) + (assert-false (stst-> 1f0 1f0)) + (assert-false (stst-> 0f0 1f0)) + ;; Now try NaN. All comparisons should be false. + (ext:with-float-traps-masked (:invalid) + (assert-false (stst-> *single-float-nan* 1f0)) + (assert-false (stst-> 1f0 *single-float-nan*)) + (assert-false (stst-> *single-float-nan* *single-float-nan*)))) + +(define-test nan-double.> + (:tag :nan) + ;; First just make sure it works with regular single-floats + (assert-true (dtst-> 2d0 1d0)) + (assert-false (dtst-> 1d0 1d0)) + (assert-false (dtst-> 0d0 1d0)) + ;; Now try NaN. All comparisons should be false. + (ext:with-float-traps-masked (:invalid) + (assert-false (dtst-> *double-float-nan* 1d0)) + (assert-false (dtst-> 1d0 *double-float-nan*)) + (assert-false (dtst-> *double-float-nan* *double-float-nan*)))) + +(define-test nan-single.<3 + (:tag :nan) + ;; First just make sure it works with regular single-floats + (assert-true (stst-<3 1f0 2f0 3f0)) + (assert-false (stst-<3 1f0 2f0 2f0)) + (assert-false (stst-<3 1f0 1f0 2f0)) + (assert-false (stst-<3 1f0 0f0 2f0)) + ;; Now try NaN. Currently we can only test if there's NaN in the + ;; first two args. When NaN is the last arg, we return the + ;; incorrect value because of how multi-compare converts multiple + ;; args into paris of comparisons. + ;; + ;; When that is fixed, we can add additional tests. Nevertheless, + ;; this is useful because it tests the not-p case of the vops. + (ext:with-float-traps-masked (:invalid) + (assert-false (stst-<3 *single-float-nan* 2f0 3f0)) + (assert-false (stst-<3 1f0 *single-float-nan* 3f0)) + (assert-false (stst-<3 *single-float-nan* *single-float-nan* 3f0)))) + +(define-test nan-double.<3 + (:tag :nan) + ;; First just make sure it works with regular double-floats + (assert-true (dtst-<3 1d0 2d0 3d0)) + (assert-false (dtst-<3 1d0 2d0 2d0)) + (assert-false (dtst-<3 1d0 1d0 2d0)) + (assert-false (dtst-<3 1d0 0d0 2d0)) + ;; Now try NaN. Currently we can only test if there's NaN in the + ;; first two args. When NaN is the last arg, we return the + ;; incorrect value because of how multi-compare converts multiple + ;; args into paris of comparisons. + ;; + ;; When that is fixed, we can add additional tests. Nevertheless, + ;; this is useful because it tests the not-p case of the vops. + (ext:with-float-traps-masked (:invalid) + (assert-false (dtst-<3 *double-float-nan* 2d0 3d0)) + (assert-false (dtst-<3 1d0 *double-float-nan* 3d0)) + (assert-false (dtst-<3 *double-float-nan* *double-float-nan* 3d0)))) + +(define-test nan-single.>3 + (:tag :nan) + ;; First just make sure it works with regular single-floats + (assert-true (stst->3 3f0 2f0 1f0)) + (assert-false (stst->3 3f0 1f0 1f0)) + (assert-false (stst->3 2f0 2f0 1f0)) + (assert-false (stst->3 0f0 2f0 1f0)) + ;; Now try NaN. Currently we can only test if there's NaN in the + ;; first two args. When NaN is the last arg, we return the + ;; incorrect value because of how multi-compare converts multiple + ;; args into paris of comparisons. + ;; + ;; When that is fixed, we can add additional tests. Nevertheless, + ;; this is useful because it tests the not-p case of the vops. + (ext:with-float-traps-masked (:invalid) + (assert-false (stst->3 *single-float-nan* 2f0 3f0)) + (assert-false (stst->3 1f0 *single-float-nan* 3f0)) + (assert-false (stst->3 *single-float-nan* *single-float-nan* 3f0)))) + +(define-test nan-double.>3 + (:tag :nan) + ;; First just make sure it works with regular double-floats + (assert-true (dtst->3 3d0 2d0 1d0)) + (assert-false (dtst->3 3d0 1d0 1d0)) + (assert-false (dtst->3 2d0 2d0 1d0)) + (assert-false (dtst->3 0d0 2d0 1d0)) + ;; Now try NaN. Currently we can only test if there's NaN in the + ;; first two args. When NaN is the last arg, we return the + ;; incorrect value because of how multi-compare converts multiple + ;; args into paris of comparisons. + ;; + ;; When that is fixed, we can add additional tests. Nevertheless, + ;; this is useful because it tests the not-p case of the vops. + (ext:with-float-traps-masked (:invalid) + (assert-false (dtst->3 *double-float-nan* 2d0 3d0)) + (assert-false (dtst->3 1d0 *double-float-nan* 3d0)) + (assert-false (dtst->3 *double-float-nan* *double-float-nan* 3d0)))) + +(define-test nan-single.= + (:tag :nan) + ;; Basic tests with regular numbers. + (assert-true (stst-= 1f0 1f0)) + (assert-false (stst-= 2f0 1f0)) + (assert-false (stst-= 0f0 1f0)) + ;; Tests with NaN, where = should fail. + (ext:with-float-traps-masked (:invalid) + (assert-false (stst-= *single-float-nan* 1f0)) + (assert-false (stst-= 1f0 *single-float-nan*)) + (assert-false (stst-= *single-float-nan* *single-float-nan*)))) + +(define-test nan-double.= + (:tag :nan) + ;; Basic tests with regular numbers. + (assert-true (stst-= 1d0 1d0)) + (assert-false (stst-= 2d0 1d0)) + (assert-false (stst-= 0d0 1d0)) + ;; Tests with NaN, where = should fail. + (ext:with-float-traps-masked (:invalid) + (assert-false (stst-= *double-float-nan* 1d0)) + (assert-false (stst-= 1d0 *double-float-nan*)) + (assert-false (stst-= *double-float-nan* *double-float-nan*)))) + +(define-test nan-single.=3 + (:tag :nan) + ;; Basic tests with regular numbers. + (assert-true (stst-=3 1f0 1f0 1f0)) + (assert-false (stst-=3 1f0 1f0 0f0)) + (assert-false (stst-=3 0f0 1f0 1f0)) + ;; Tests with NaN, where = should fail. + (ext:with-float-traps-masked (:invalid) + (assert-false (stst-=3 *single-float-nan* 1f0 1f0)) + (assert-false (stst-=3 1f0 *single-float-nan* 1f0)) + (assert-false (stst-=3 1f0 1f0 *single-float-nan*)))) + +(define-test nan-double.=3 + (:tag :nan) + ;; Basic tests with regular numbers. + (assert-true (dtst-=3 1d0 1d0 1d0)) + (assert-false (dtst-=3 1d0 1d0 0d0)) + (assert-false (dtst-=3 0d0 1d0 1d0)) + ;; Tests with NaN, where = should fail. + (ext:with-float-traps-masked (:invalid) + (assert-false (dtst-=3 *double-float-nan* 1d0 1d0)) + (assert-false (dtst-=3 1d0 *double-float-nan* 1d0)) + (assert-false (dtst-=3 1d0 1d0 *double-float-nan*))))

View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/compare/0038d3d9a489ffcf47bb155...