According to the POWER8 Processor User’s Manual and some
experimentation on power8 hardware, unaligned fp load/store on power8
does not invoke the alignment interrupt. Thus SLOW_UNALIGNED_ACCESS
should not return true for these accesses. It so happens that we
already have an undocumented switch to control whether VSX vector
access is slow, and since FPRs are a subset of VSX regs it seemed good
to me to use the same switch.
Bootstrapped and regtested powerpc64le-linux and powerpc64-linux.
* config/rs6000/rs6000.h (SLOW_UNALIGNED_ACCESS): Make scalar
float access depend on TARGET_EFFICIENT_UNALIGNED_VSX.
* config/rs6000/rs6000.opt (-mefficient-unaligned-vsx): Renamed
from -mefficient-unaligned-vector. Note that this affects fp too.
diff --git a/gcc/config/rs6000/rs6000.h b/gcc/config/rs6000/rs6000.h
index 9b3e83b..af77258 100644
--- a/gcc/config/rs6000/rs6000.h
+++ b/gcc/config/rs6000/rs6000.h
@@ -990,10 +990,10 @@ enum data_align { align_abi, align_opt, align_both };
aligned to 4 or 8 bytes. */
#define SLOW_UNALIGNED_ACCESS(MODE, ALIGN) \
(STRICT_ALIGNMENT \
- || (SCALAR_FLOAT_MODE_NOT_VECTOR_P (MODE) && (ALIGN) < 32) \
- || (!TARGET_EFFICIENT_UNALIGNED_VSX \
- && ((VECTOR_MODE_P (MODE) || FLOAT128_VECTOR_P (MODE)) \
- && (((int)(ALIGN)) < VECTOR_ALIGN (MODE)))))
+ || (!TARGET_EFFICIENT_UNALIGNED_VSX \
+ && ((SCALAR_FLOAT_MODE_NOT_VECTOR_P (MODE) && (ALIGN) < 32) \
+ || ((VECTOR_MODE_P (MODE) || FLOAT128_VECTOR_P (MODE)) \
+ && (int) (ALIGN) < VECTOR_ALIGN (MODE)))))
/* Standard register usage. */
diff --git a/gcc/config/rs6000/rs6000.opt b/gcc/config/rs6000/rs6000.opt
index 9e9d33e..e47f47f 100644
--- a/gcc/config/rs6000/rs6000.opt
+++ b/gcc/config/rs6000/rs6000.opt
@@ -211,9 +211,9 @@ mallow-movmisalign
Target Undocumented Var(TARGET_ALLOW_MOVMISALIGN) Init(-1) Save
; Allow the movmisalign in DF/DI vectors
-mefficient-unaligned-vector
+mefficient-unaligned-vsx
Target Undocumented Report Mask(EFFICIENT_UNALIGNED_VSX) Var(rs6000_isa_flags)
-; Consider unaligned VSX accesses to be efficient
+; Consider unaligned VSX vector and fp accesses to be efficient
mallow-df-permute
Target Undocumented Var(TARGET_ALLOW_DF_PERMUTE) Save
--
Alan Modra
Australia Development Lab, IBM
