diff options
author | Joel Sing <joel@sing.id.au> | 2023-04-29 20:29:41 +1000 |
---|---|---|
committer | Joel Sing <joel@sing.id.au> | 2023-07-31 16:39:36 +0000 |
commit | 6eaad824e56aec91266854bf7890a94c3f08b614 (patch) | |
tree | cbf6a147f0aaee20e26fae62da66a1a5b00d4378 /src/cmd/internal/obj/arm64/asm7.go | |
parent | 03bec7dc6ff6bbcdf077753230cee11211aa78ba (diff) | |
download | go-6eaad824e56aec91266854bf7890a94c3f08b614.tar.gz go-6eaad824e56aec91266854bf7890a94c3f08b614.zip |
cmd/internal/obj/arm64: improve splitting of 24 bit unsigned scaled immediates
The previous implementation would limit itself to 0xfff000 | 0xfff << shift,
while the maximum possible value is 0xfff000 + 0xfff << shift. In practical
terms, this means that an additional ((1 << shift) - 1) * 0x1000 of offset
is reachable for operations that use this splitting format. In the case of
an 8 byte load/store, this is an additional 0x7000 that can be reached
without needing to use the literal pool.
Updates #59615
Change-Id: Ice7023104042d31c115eafb9398c2b999bdd6583
Reviewed-on: https://go-review.googlesource.com/c/go/+/512540
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Joel Sing <joel@sing.id.au>
Diffstat (limited to 'src/cmd/internal/obj/arm64/asm7.go')
-rw-r--r-- | src/cmd/internal/obj/arm64/asm7.go | 19 |
1 files changed, 13 insertions, 6 deletions
diff --git a/src/cmd/internal/obj/arm64/asm7.go b/src/cmd/internal/obj/arm64/asm7.go index 05cf62773e..ea53a838e3 100644 --- a/src/cmd/internal/obj/arm64/asm7.go +++ b/src/cmd/internal/obj/arm64/asm7.go @@ -1420,13 +1420,20 @@ func splitImm24uScaled(v int32, shift int) (int32, int32, error) { if v < 0 { return 0, 0, fmt.Errorf("%d is not a 24 bit unsigned immediate", v) } + if v > 0xfff000+0xfff<<shift { + return 0, 0, fmt.Errorf("%d is too large for a scaled 24 bit unsigned immediate", v) + } if v&((1<<shift)-1) != 0 { return 0, 0, fmt.Errorf("%d is not a multiple of %d", v, 1<<shift) } lo := (v >> shift) & 0xfff hi := v - (lo << shift) - if hi&^0xfff000 != 0 { - return 0, 0, fmt.Errorf("%d is too large for a scaled 24 bit unsigned immediate %x %x", v, lo, hi) + if hi > 0xfff000 { + hi = 0xfff000 + lo = (v - hi) >> shift + } + if hi & ^0xfff000 != 0 { + panic(fmt.Sprintf("bad split for %x with shift %v (%x, %x)", v, shift, hi, lo)) } return hi, lo, nil } @@ -1975,28 +1982,28 @@ func (c *ctxt7) loadStoreClass(p *obj.Prog, lsc int, v int64) int { if cmp(C_UAUTO8K, lsc) || cmp(C_UOREG8K, lsc) { return lsc } - if v >= 0 && v <= 0xfffffe && v&1 == 0 { + if v >= 0 && v <= 0xfff000+0xfff<<1 && v&1 == 0 { needsPool = false } case AMOVW, AMOVWU, AFMOVS: if cmp(C_UAUTO16K, lsc) || cmp(C_UOREG16K, lsc) { return lsc } - if v >= 0 && v <= 0xfffffc && v&3 == 0 { + if v >= 0 && v <= 0xfff000+0xfff<<2 && v&3 == 0 { needsPool = false } case AMOVD, AFMOVD: if cmp(C_UAUTO32K, lsc) || cmp(C_UOREG32K, lsc) { return lsc } - if v >= 0 && v <= 0xfffff8 && v&7 == 0 { + if v >= 0 && v <= 0xfff000+0xfff<<3 && v&7 == 0 { needsPool = false } case AFMOVQ: if cmp(C_UAUTO64K, lsc) || cmp(C_UOREG64K, lsc) { return lsc } - if v >= 0 && v <= 0xfffff0 && v&15 == 0 { + if v >= 0 && v <= 0xfff000+0xfff<<4 && v&15 == 0 { needsPool = false } } |