-
476c4cdf
by Sean D. Gillespie at 2026-03-02T10:14:37-05:00
Add SIMD absolute value on x86 and LLVM
On x86, absolute value of 32 bits or less is implemented with
PABSB/PABSW/PABSD if SSSE3 is available. Otherwise, there is a fallback
for SSE2. For 64 bit integers it uses VPABSQ, required by AVX-512VL,
with fallbacks for SSE4.2 and SSE2.
There is no dedicated instruction for floating point absolute value on
x86, so it is simulated using bitwise AND.
Absolute value for signed integers and floats are implemented by the
"llvm.abs/llvm.fabs" standard library intrinsics. This implementation
uses MachOps constructors, unlike non-vector floating point absolute
value, which uses CallishMachOps.
-
709448c0
by Sean D. Gillespie at 2026-03-02T10:14:46-05:00
Add SIMD floating point square root
On x86, this is implemented with the SQRTPS and SQRTPD instructions. On
LLVM, it uses the sqrt library intrinstic.
-
0deadf66
by Sean D. Gillespie at 2026-03-02T10:14:47-05:00
Improve error message for SIMD on aarch64
When encountering vector literals on aarch64, previously it would
throw:
<no location info>: error:
panic! (the 'impossible' happened)
GHC version 9.15.20251219:
getRegister' (CmmLit:CmmVec):
Now it is more consistent with the other vector operations:
<no location info>: error:
sorry! (unimplemented feature or known bug)
GHC version 9.15.20251219:
SIMD operations on AArch64 currently require the LLVM backend
-
7d64031b
by Vladislav Zavialov at 2026-03-03T11:09:28-05:00
Replace maybeAddSpace with spaceIfSingleQuote
Simplify pretty-printing of HsTypes by using spaceIfSingleQuote.
This allows us to drop the unwieldy lhsTypeHasLeadingPromotionQuote
helper function.
Follow-up to 178c1fd830c78377ef5d338406a41e1d8eb5f0da
-
6f7ac47b
by Vladislav Zavialov at 2026-03-05T15:39:06+03:00
TyCons in terms with -fdefer-out-of-scope-variables (#19966)
This fixes a regression introduced in 990ea991a1c35fdb894fcb91f919fb8f8fed33dd
which prevented the following program from compiling:
{-# OPTIONS -fdefer-out-of-scope-variables #-}
f = Int
In addition to that, we no longer abort type checking on
TcRnIllegalTermLevelUse, allowing GHC to report more errors.