[Git][ghc/ghc][wip/uf-conf-codegen] 2 commits: CmmToAsm/X86: Improve code generation for MO_UF_Conv of literal

[Ben Gamari]

Ben Gamari pushed to branch wip/uf-conf-codegen at Glasgow Haskell Compiler / GHC


Commits:
5c756080 by Ben Gamari at 2020-10-26T13:58:13-04:00
CmmToAsm/X86: Improve code generation for MO_UF_Conv of literal

Previously we would call hs_word2float which is incredibly expensive
compared to just a MOV.

- - - - -
1b92bde8 by Ben Gamari at 2020-10-26T13:58:13-04:00
primops: Document semantics of Float/Int conversions

Fixes #18840.

- - - - -


2 changed files:

- compiler/GHC/Builtin/primops.txt.pp
- compiler/GHC/CmmToAsm/X86/CodeGen.hs


Changes:

=====================================
compiler/GHC/Builtin/primops.txt.pp
=====================================
@@ -607,10 +607,22 @@ primop   IntToWordOp "int2Word#" GenPrimOp Int# -> Word#
    with code_size = 0
 
 primop   IntToFloatOp   "int2Float#"      GenPrimOp  Int# -> Float#
+   {Convert an {\tt Int#} to the corrsponding {\tt Float#} with the same
+    integral value (up to truncation due to floating-point precision). e.g.
+    {\tt int2Float# 1# == 1.0#}}
 primop   IntToDoubleOp   "int2Double#"          GenPrimOp  Int# -> Double#
+   {Convert an {\tt Int#} to the corrsponding {\tt Double#} with the same
+    integral value (up to truncation due to floating-point precision). e.g.
+    {\tt int2Double# 1# == 1.0#}}
 
 primop   WordToFloatOp   "word2Float#"      GenPrimOp  Word# -> Float#
+   {Convert an {\tt Word#} to the corrsponding {\tt Float#} with the same
+    integral value (up to truncation due to floating-point precision). e.g.
+    {\tt int2Float# 1# == 1.0#}}
 primop   WordToDoubleOp   "word2Double#"          GenPrimOp  Word# -> Double#
+   {Convert an {\tt Word#} to the corrsponding {\tt Double#} with the same
+    integral value (up to truncation due to floating-point precision). e.g.
+    {\tt int2Double# 1# == 1.0#}}
 
 primop   ISllOp   "uncheckedIShiftL#" GenPrimOp  Int# -> Int# -> Int#
          {Shift left.  Result undefined if shift amount is not


=====================================
compiler/GHC/CmmToAsm/X86/CodeGen.hs
=====================================
@@ -2511,6 +2511,14 @@ genCCall' config is32Bit (PrimTarget (MO_Clz width)) dest_regs@[dst] args@[src]
     bw = widthInBits width
     lbl = mkCmmCodeLabel primUnitId (fsLit (clzLabel width))
 
+genCCall' config _ (PrimTarget (MO_UF_Conv width)) [dst] [CmmLit arg_lit] _
+  | CmmFloat n _ <- arg_lit = do
+    let fmt = floatFormat width
+    Amode amode amode_code <- memConstant (mkAlignment $ widthInBytes width) (CmmFloat (realToFrac n) width)
+    let platform = ncgPlatform config
+    let dst_r = getRegisterReg platform (CmmLocal dst)
+    return $ amode_code `appOL` unitOL (MOV fmt (OpAddr amode) (OpReg dst_r))
+
 genCCall' config is32Bit (PrimTarget (MO_UF_Conv width)) dest_regs args bid = do
     targetExpr <- cmmMakeDynamicReference config
                   CallReference lbl



View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/fa34519a875ed13c66f6c5e990a873239d53a234...1b92bde80e492895ff2be26d673a5f179543a539

-- 
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/fa34519a875ed13c66f6c5e990a873239d53a234...1b92bde80e492895ff2be26d673a5f179543a539
You're receiving this email because of your account on gitlab.haskell.org.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-commits/attachments/20201026/2cfb8fd0/attachment-0001.html>