6896

Readme-driven development for Mu's floating-point operations.

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