On 2011/3/24 06:51 PM, Richard Earnshaw wrote: > > On Thu, 2011-03-24 at 12:56 +0900, Chung-Lin Tang wrote: >> Hi, >> PR48250 happens under TARGET_NEON, where DImode is included within the >> valid NEON modes. This turns the range of legitimate constant indexes to >> step-4 (coproc load/store), thus arm_legitimize_reload_address() when >> trying to decompose the [reg+index] reload address into >> [(reg+index_high)+index_low], can cause an ICE later when 'index_low' >> part is not aligned to 4. >> >> I'm not sure why the current DImode index is computed as: >> low = ((val & 0xf) ^ 0x8) - 0x8; the sign-extending into negative >> values, then subtracting back, actually creates further off indexes. >> e.g. in the supplied testcase, [sp+13] was turned into [(sp+16)-3]. >> > > Hysterical Raisins... the code there was clearly written for the days > before we had LDRD in the architecture. At that time the most efficient > way to load a 64-bit object was to use the LDM{ia,ib,da,db} > instructions. The computation here was (I think), intended to try and > make the most efficient use of an add/sub instruction followed by > LDM/STM offsetting. At that time the architecture had no unaligned > access either, so dealing with 64-bit that were less than 32-bit aligned > (in those days 32-bit was the maximum alignment) probably wasn't > considered, or couldn't even get through to reload. >
I see it now. The code in output_move_double() returning assembly for ldm/stm(db/da/ib) for offsets -8/-4/+4 seems to confirm this. I have changed the patch to let the new code handle the TARGET_LDRD case only. The pre-LDRD case is still handled by the original code, with an additional & ~0x3 for aligning the offset to 4. I've also added a comment for the pre-TARGET_LDRD case. Please see if the description is accurate enough. >> My patch changes the index decomposing to a more straightforward way; it >> also sort of outlines the way the other reload address indexes are >> broken by using and-masks, is not the most effective. The address is >> computed by addition, subtracting away the parts to obtain low+high >> should be the optimal way of giving the largest computable index range. >> >> I have included a few Thumb-2 bits in the patch; I know currently >> arm_legitimize_reload_address() is only used under TARGET_ARM, but I >> guess it might eventually be turned into TARGET_32BIT. >> > > I think this needs to be looked at carefully on ARMv4/ARMv4T to check > that it doesn't cause regressions there when we don't have LDRD in the > instruction set. I'll be testing the modified patch under an ARMv4/ARMv4T configuration. Okay for trunk if no regressions? Thanks, Chung-Lin PR target/48250 * config/arm/arm.c (arm_legitimize_reload_address): Adjust DImode constant index decomposing under TARGET_LDRD. Clear lower two bits for NEON, Thumb-2, and !TARGET_LDRD. Add comment for !TARGET_LDRD case.
Index: config/arm/arm.c =================================================================== --- config/arm/arm.c (revision 171652) +++ config/arm/arm.c (working copy) @@ -6420,7 +6420,32 @@ HOST_WIDE_INT low, high; if (mode == DImode || (mode == DFmode && TARGET_SOFT_FLOAT)) - low = ((val & 0xf) ^ 0x8) - 0x8; + { + if (TARGET_LDRD) + { + /* ??? There may be more adjustments later for Thumb-2, + which has a ldrd insn with +-1020 index range. */ + int max_idx = 255; + + /* low == val, if val is within range [-max_idx, +max_idx]. + If not, val is set to the boundary +-max_idx. */ + low = (-max_idx <= val && val <= max_idx + ? val : (val > 0 ? max_idx : -max_idx)); + + /* Thumb-2 ldrd, and NEON coprocessor load/store indexes + are in steps of 4, so the least two bits need to be + cleared to zero. */ + if (TARGET_NEON || TARGET_THUMB2) + low &= ~0x3; + } + else + { + /* For pre-ARMv5TE (without ldrd), we use ldm/stm(db/da/ib) + to access doublewords. The supported load/store offsets are + -8, -4, and 4, which we try to produce here. */ + low = (((val & 0xf) ^ 0x8) - 0x8) & ~0x3; + } + } else if (TARGET_MAVERICK && TARGET_HARD_FLOAT) /* Need to be careful, -256 is not a valid offset. */ low = val >= 0 ? (val & 0xff) : -((-val) & 0xff);