Add support for prefetch with locality levels.

This patch adds support for several new primitive operations which
support using processor-specific instructions to help guide data and
cache locality decisions. We have levels ranging from [0..3]

For LLVM, we generate llvm.prefetch intrinsics at the proper locality
level (similar to GCC.)

For x86 we generate prefetch{NTA, t2, t1, t0} instructions. On SPARC and
PowerPC, the locality levels are ignored.

This closes #8256.

Authored-by: Carter Tazio Schonwald <carter.schonwald@gmail.com>
Signed-off-by: Austin Seipp <austin@well-typed.com>

3 files changed, 24 insertions, 8 deletions

diff --git a/compiler/cmm/CmmMachOp.hs b/compiler/cmm/CmmMachOp.hs
index c009d15e25..684a4b9729 100644
--- a/compiler/cmm/CmmMachOp.hs
+++ b/compiler/cmm/CmmMachOp.hs
@@ -107,10 +107,10 @@ data MachOp
   -- Vector element insertion and extraction operations
   | MO_V_Insert  Length Width   -- Insert scalar into vector
   | MO_V_Extract Length Width   -- Extract scalar from vector
-  
+
   -- Integer vector operations
-  | MO_V_Add Length Width  
-  | MO_V_Sub Length Width  
+  | MO_V_Add Length Width
+  | MO_V_Sub Length Width
   | MO_V_Mul Length Width
 
   -- Signed vector multiply/divide
@@ -127,8 +127,8 @@ data MachOp
   | MO_VF_Extract Length Width   -- Extract scalar from vector
 
   -- Floating point vector operations
-  | MO_VF_Add  Length Width  
-  | MO_VF_Sub  Length Width  
+  | MO_VF_Add  Length Width
+  | MO_VF_Sub  Length Width
   | MO_VF_Neg  Length Width             -- unary -
   | MO_VF_Mul  Length Width
   | MO_VF_Quot Length Width
@@ -528,8 +528,14 @@ data CallishMachOp
   | MO_Touch         -- Keep variables live (when using interior pointers)
 
   -- Prefetch
-  | MO_Prefetch_Data -- Prefetch hint. May change program performance but not
+  | MO_Prefetch_Data Int -- Prefetch hint. May change program performance but not
                      -- program behavior.
+                     -- the Int can be 0-3. Needs to be known at compile time
+                     -- to interact with code generation correctly.
+                     --  TODO: add support for prefetch WRITES,
+                     --  currently only exposes prefetch reads, which
+                     -- would the majority of use cases in ghc anyways
+
 
   -- Note that these three MachOps all take 1 extra parameter than the
   -- standard C lib versions. The extra (last) parameter contains
diff --git a/compiler/cmm/CmmParse.y b/compiler/cmm/CmmParse.y
index ebd9278e15..a0c9bc4eb5 100644
--- a/compiler/cmm/CmmParse.y
+++ b/compiler/cmm/CmmParse.y
@@ -952,8 +952,16 @@ callishMachOps = listToUFM $
         ( "write_barrier", MO_WriteBarrier ),
         ( "memcpy", MO_Memcpy ),
         ( "memset", MO_Memset ),
-        ( "memmove", MO_Memmove )
+        ( "memmove", MO_Memmove ),
+
+        ("prefetch0",MO_Prefetch_Data 0),
+        ("prefetch1",MO_Prefetch_Data 1),
+        ("prefetch2",MO_Prefetch_Data 2),
+        ("prefetch3",MO_Prefetch_Data 3)
+
         -- ToDo: the rest, maybe
+        -- edit: which rest?
+        -- also: how do we tell CMM Lint how to type check callish macops?
     ]
 
 parseSafety :: String -> P Safety
diff --git a/compiler/cmm/PprC.hs b/compiler/cmm/PprC.hs
index c468161c73..32fd8b4feb 100644
--- a/compiler/cmm/PprC.hs
+++ b/compiler/cmm/PprC.hs
@@ -759,7 +759,9 @@ pprCallishMachOp_for_C mop
         MO_Add2       {} -> unsupported
         MO_U_Mul2     {} -> unsupported
         MO_Touch         -> unsupported
-        MO_Prefetch_Data -> unsupported
+        (MO_Prefetch_Data _ ) -> unsupported
+        --- we could support prefetch via "__builtin_prefetch"
+        --- Not adding it for now
     where unsupported = panic ("pprCallishMachOp_for_C: " ++ show mop
                             ++ " not supported!")