From 9aea89ba73280bcaeee67f6bd01f725a9698b344 Mon Sep 17 00:00:00 2001 From: Kartik Agaram Date: Mon, 28 Sep 2020 22:19:43 -0700 Subject: 6896 Readme-driven development for Mu's floating-point operations. --- mu.md | 95 ++++++++++++++++++++++++++++++++++++++++++++++++++++++--- mu_instructions | 77 ++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 167 insertions(+), 5 deletions(-) diff --git a/mu.md b/mu.md index aadb78e8..9c26dd67 100644 --- a/mu.md +++ b/mu.md @@ -127,16 +127,24 @@ var name/reg: type <- ... Variables on the stack are never initialized. (They're always implicitly zeroed them out.) Variables in registers are always initialized. -Register variables can go in 6 registers: `eax`, `ebx`, `ecx`, `edx`, `esi` -and `edi`. Defining a variable in a register either clobbers the previous -variable (if it was defined in the same block) or shadows it temporarily (if -it was defined in an outer block). +Register variables can go in 6 integer registers: `eax`, `ebx`, `ecx`, `edx`, +`esi` and `edi`. Floating-point values can also go in 8 other registers: +`xmm0`, `xmm1`, `xmm2`, `xmm3`, `xmm4`, `xmm5`, `xmm6` and `xmm7`. + +Defining a variable in a register either clobbers the previous variable (if it +was defined in the same block) or shadows it temporarily (if it was defined in +an outer block). Variables exist from their definition until the end of their containing block. Register variables may also die earlier if their register is clobbered by a new variable. -## Arithmetic primitives +Variables on the stack can be of many types (but not `byte`). Variables in +integer registers can only contain 32-bit values: `int`, `boolean`, `(addr +...)`. Variables in floating-point registers can only contain values of type +`float`. + +## Integer primitives Here is the list of arithmetic primitive operations supported by Mu. The name `n` indicates a literal integer rather than a variable, and `var/reg` indicates @@ -208,6 +216,83 @@ Excluding dereferences, the above statements must operate on non-address primitive types: `int` or `boolean`. (Booleans are really just `int`s, and Mu assumes any value but `0` is true.) +## Floating-point primitives + +These instructions may use the floating-point registers `xmm0` ... `xmm7` +(denoted by `/xreg2` or `/xrm32`). They also use integer values on occasion +(`/rm32` and `/r32`). They can't take literal floating-point values. + +``` +var/xreg <- add var2/xreg2 +var/xreg <- add var2 +var/xreg <- add *var2/reg2 + +var/xreg <- subtract var2/xreg2 +var/xreg <- subtract var2 +var/xreg <- subtract *var2/reg2 + +var/xreg <- multiply var2/xreg2 +var/xreg <- multiply var2 +var/xreg <- multiply *var2/reg2 + +var/xreg <- divide var2/xreg2 +var/xreg <- divide var2 +var/xreg <- divide *var2/reg2 + +var/xreg <- reciprocal var2/xreg2 +var/xreg <- reciprocal var2 +var/xreg <- reciprocal *var2/reg2 + +var/xreg <- square-root var2/xreg2 +var/xreg <- square-root var2 +var/xreg <- square-root *var2/reg2 + +var/xreg <- inverse-square-root var2/xreg2 +var/xreg <- inverse-square-root var2 +var/xreg <- inverse-square-root *var2/reg2 + +var/xreg <- min var2/xreg2 +var/xreg <- min var2 +var/xreg <- min *var2/reg2 + +var/xreg <- max var2/xreg2 +var/xreg <- max var2 +var/xreg <- max *var2/reg2 + +Remember, when these instructions use indirect mode, they still use an integer +register. Floating-point registers can't hold addresses. + +Most instructions operate exclusively on integer or floating-point operands. +The only exceptions are the instructions for converting between integers and +floating-point numbers. + +var/xreg <- convert var2/reg2 +var/xreg <- convert var2 +var/xreg <- convert *var2/reg2 + +var/reg <- convert var2/xreg2 +var/reg <- convert var2 +var/reg <- convert *var2/reg2 + +There are no instructions accepting floating-point literals. To obtain integer +literals in floating-point registers, copy them to general-purpose registers +and then convert them to floating-point. + +One pattern you may have noticed above is that the floating-point instructions +above always write to registers. The only exceptions are `copy` instructions, +which can write to memory locations. + +var/xreg <- copy var2/xreg2 +copy-to var1, var2/xreg +var/xreg <- copy var2 +var/xreg <- copy *var2/reg2 + +Floating-point comparisons always put a register on the left-hand side: + +compare var1/xreg1, var2/xreg2 +compare var1/xreg1, var2 +``` + ## Operating on individual bytes A special-case is variables of type 'byte'. Mu is a 32-bit platform so for the diff --git a/mu_instructions b/mu_instructions index 4566669c..fe11aaba 100644 --- a/mu_instructions +++ b/mu_instructions @@ -262,4 +262,81 @@ read-from-stream s: (addr stream T), out: (addr T) write-to-stream s: (addr stream T), in: (addr T) => "(write-to-stream " s " " in " " size-of(T) ")" +# Floating-point operations + +All the instructions so far use Intel's general-purpose integer registers. +However, some of them translate to different SubX if their arguments are in +floating-point registers. + +var/xreg <- add var2/xreg2 => "f3 0f 58/add 3/mod " xreg2 "/xm32 " xreg1 "/x32" +var/xreg <- add var2 => "f3 0f 58/add *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- add *var2/reg2 => "f3 0f 58/add *" reg2 " " xreg "/x32" + +var/xreg <- subtract var2/xreg2 => "f3 0f 5c/subtract 3/mod " xreg2 "/xm32 " xreg1 "/x32" +var/xreg <- subtract var2 => "f3 0f 5c/subtract *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- subtract *var2/reg2 => "f3 0f 5c/subtract *" reg2 " " xreg "/x32" + +var/xreg <- multiply var2/xreg2 => "f3 0f 59/multiply 3/mod " xreg2 "/xm32 " xreg1 "/x32" +var/xreg <- multiply var2 => "f3 0f 59/multiply *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- multiply *var2/reg2 => "f3 0f 59/multiply *" reg2 " " xreg "/x32" + +var/xreg <- divide var2/xreg2 => "f3 0f 5e/divide 3/mod " xreg2 "/xm32 " xreg1 "/x32" +var/xreg <- divide var2 => "f3 0f 5e/divide *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- divide *var2/reg2 => "f3 0f 5e/divide *" reg2 " " xreg "/x32" + +There are also some exclusively floating-point instructions: + +var/xreg <- reciprocal var2/xreg2 => "f3 0f 53/reciprocal 3/mod " xreg2 "/xm32 " xreg1 "/x32" +var/xreg <- reciprocal var2 => "f3 0f 53/reciprocal *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- reciprocal *var2/reg2 => "f3 0f 53/reciprocal *" reg2 " " xreg "/x32" + +var/xreg <- square-root var2/xreg2 => "f3 0f 51/square-root 3/mod " xreg2 "/xm32 " xreg1 "/x32" +var/xreg <- square-root var2 => "f3 0f 51/square-root *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- square-root *var2/reg2 => "f3 0f 51/square-root *" reg2 " " xreg "/x32" + +var/xreg <- inverse-square-root var2/xreg2 => "f3 0f 52/inverse-square-root 3/mod " xreg2 "/xm32 " xreg1 "/x32" +var/xreg <- inverse-square-root var2 => "f3 0f 52/inverse-square-root *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- inverse-square-root *var2/reg2 => "f3 0f 52/inverse-square-root *" reg2 " " xreg "/x32" + +var/xreg <- min var2/xreg2 => "f3 0f 5d/min 3/mod " xreg2 "/xm32 " xreg1 "/x32" +var/xreg <- min var2 => "f3 0f 5d/min *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- min *var2/reg2 => "f3 0f 5d/min *" reg2 " " xreg "/x32" + +var/xreg <- max var2/xreg2 => "f3 0f 5f/max 3/mod " xreg2 "/xm32 " xreg1 "/x32" +var/xreg <- max var2 => "f3 0f 5f/max *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- max *var2/reg2 => "f3 0f 5f/max *" reg2 " " xreg "/x32" + +Remember, when these instructions use indirect mode, they still use an integer +register. Floating-point registers can't hold addresses. + +Most instructions operate exclusively on integer or floating-point operands. +The only exceptions are the instructions for converting between integers and +floating-point numbers. + +var/xreg <- convert var2/reg2 => "f3 0f 2a/convert-to-float %" reg2 " " xreg "/x32" +var/xreg <- convert var2 => "f3 0f 2a/convert-to-float *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- convert *var2/reg2 => "f3 0f 2a/convert-to-float *" reg2 " " xreg "/x32" + +var/reg <- convert var2/xreg2 => "f3 0f 2d/convert-to-int 3/mod " xreg2 "/xm32 " reg "/r32" +var/reg <- convert var2 => "f3 0f 2d/convert-to-int *(ebp+" var2.stack-offset ") " reg "/r32" +var/reg <- convert *var2/reg2 => "f3 0f 2d/convert-to-int *" reg2 " " reg "/r32" + +There are no instructions accepting floating-point literals. To obtain integer +literals in floating-point registers, copy them to general-purpose registers +and then convert them to floating-point. + +One pattern you may have noticed above is that the floating-point instructions +above always write to registers. The only exceptions are `copy` instructions, +which can write to memory locations. + +var/xreg <- copy var2/xreg2 => "f3 0f 11/<- 3/mod " xreg "/xm32 " xreg2 "/x32" +copy-to var1, var2/xreg => "f3 0f 11/<- *(ebp+" var1.stack-offset ") " xreg "/x32" +var/xreg <- copy var2 => "f3 0f 10/-> *(ebp+" var2.stack-offset ") " xreg "/x32" +var/xreg <- copy *var2/reg2 => "f3 0f 10/-> *" reg2 " " xreg "/x32" + +Comparisons must always start with a register: + +compare var1/xreg1, var2/xreg2 => "0f 2f 3/mod " xreg2 "/xm32 " xreg1 "/x32" +compare var1/xreg1, var2 => "0f 2f 2/mod *(ebp+" var2.stack-offset ") " xreg1 "/x32" + vim:ft=mu:nowrap:textwidth=0 -- cgit 1.4.1-2-gfad0