On 10/5/21 19:48, Roger Sayle wrote:
This patch to the nvptx backend changes the backend's STORE_FLAG_VALUE
from -1 to 1, by using BImode predicates and selp instructions, instead
of set instructions (almost always followed by integer negation).
Historically, it was reasonable (through rare) for backends to use -1
for representing true during the RTL passes. However with tree-ssa,
GCC now emits lots of code that reads and writes _Bool values, requiring
STORE_FLAG_VALUE=-1 targets to frequently convert 0/-1 pseudos to 0/1
pseudos using integer negation. Unfortunately, this process prevents
or complicates many optimizations (negate isn't associative with logical
AND, OR and XOR, and interferes with range/vrp/nonzerobits bounds etc.).
The impact of this is that for a relatively simple logical expression
like "return (x==21) && (y==69);", the nvptx backend currently generates:
mov.u32 %r26, %ar0;
mov.u32 %r27, %ar1;
set.u32.eq.u32 %r30, %r26, 21;
neg.s32 %r31, %r30;
mov.u32 %r29, %r31;
set.u32.eq.u32 %r33, %r27, 69;
neg.s32 %r34, %r33;
mov.u32 %r32, %r34;
cvt.u16.u8 %r39, %r29;
mov.u16 %r36, %r39;
cvt.u16.u8 %r39, %r32;
mov.u16 %r37, %r39;
and.b16 %r35, %r36, %r37;
cvt.u32.u16 %r38, %r35;
cvt.u32.u8 %value, %r38;
This patch tweaks nvptx to generate 0/1 values instead, requiring the
same number of instructions, using (BImode) predicate registers and selp
instructions so as to now generate the almost identical:
mov.u32 %r26, %ar0;
mov.u32 %r27, %ar1;
setp.eq.u32 %r31, %r26, 21;
selp.u32 %r30, 1, 0, %r31;
mov.u32 %r29, %r30;
setp.eq.u32 %r34, %r27, 69;
selp.u32 %r33, 1, 0, %r34;
mov.u32 %r32, %r33;
cvt.u16.u8 %r39, %r29;
mov.u16 %r36, %r39;
cvt.u16.u8 %r39, %r32;
mov.u16 %r37, %r39;
and.b16 %r35, %r36, %r37;
cvt.u32.u16 %r38, %r35;
cvt.u32.u8 %value, %r38;
The hidden benefit is that this sequence can (in theory) be optimized
by the RTL passes to eventually generate a much shorter sequence using
an and.pred instruction (just like Nvidia's nvcc compiler).
This patch has been tested nvptx-none with a "make" and "make -k check"
(including newlib) hosted on x86_64-pc-linux-gnu with no new failures.
Ok for mainline?
Thanks for the patch, sounds reasonable.
Committed.
Thanks,
- Tom
2021-10-05 Roger Sayle <ro...@nextmovesoftware.com>
gcc/ChangeLog
* config/nvptx/nvptx.h (STORE_FLAG_VALUE): Change to 1.
* config/nvptx/nvptx.md (movbi): Use P1 constraint for true.
(setcc_from_bi): Remove SImode specific pattern.
(setcc<mode>_from_bi): Provide more general HSDIM pattern.
(extendbi<mode>2, zeroextendbi<mode>2): Provide instructions
for sign- and zero-extending BImode predicates to integers.
(setcc_int<mode>): Remove previous (-1-based) instructions.
(cstorebi4): Remove BImode to SImode specific expander.
(cstore<mode>4): Fix indentation. Expand using setccsi_from_bi.
(cstore<mode>4): For both integer and floating point modes.
Thanks in advance,
Roger
--
