AuroraMiddleware/v8
1
0
Fork 0
Code Issues 2 Pull Requests Releases Wiki Activity

ARM: Tweak ECMAToInt32VFP to address regression on Nexus 4

Browse Source 
After r13912, we saw a small regression in Kraken crypto-aes and
crypto-ccm on Nexus 4. This patch, proposed by Rodolph Perfetta,
addresses that without regressing other platforms.

Instead of looking at the exponent of double_input and trying to figure
out if the conversion will overflow, eagerly do the VCVT and return
early unless it saturated.

BUG=none

Review URL: https://chromiumcodereview.appspot.com/12544025
Patch from Hans Wennborg <hans@chromium.org>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@13948 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This commit is contained in:
ulan@chromium.org 2013-03-14 15:28:15 +00:00
parent 6e86d34ba9
commit 80f490085f
1 changed files with 9 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
src/arm/macro-assembler-arm.cc
View File

@ -2531,7 +2531,15 @@ void MacroAssembler::ECMAToInt32VFP(Register result,
!scratch.is(input_low));
ASSERT(!double_input.is(double_scratch));
Label overflow, out_of_range, negate, done;
Label out_of_range, negate, done;
vcvt_s32_f64(double_scratch.low(), double_input);
vmov(result, double_scratch.low());
// If result is not saturated (0x7fffffff or 0x80000000), we are done.
sub(scratch, result, Operand(1));
cmp(scratch, Operand(0x7ffffffe));
b(lt, &done);
vmov(input_low, input_high, double_input);
Ubfx(scratch, input_high,
@ -2539,20 +2547,9 @@ void MacroAssembler::ECMAToInt32VFP(Register result,
// Load scratch with exponent - 1. This is faster than loading
// with exponent because Bias + 1 = 1024 which is an *ARM* immediate value.
sub(scratch, scratch, Operand(HeapNumber::kExponentBias + 1));
// Compare exponent with 31 (compare exponent - 1 with 30).
cmp(scratch, Operand(30));
b(ge, &overflow);
// Exponent is less than 31 so vcvt will never saturate.
// So, just return the result.
vcvt_s32_f64(double_scratch.low(), double_input);
vmov(result, double_scratch.low());
b(&done);
bind(&overflow);
// If exponent is greater than or equal to 84, the 32 less significant
// bits are 0s (2^84 = 1, 52 significant bits, 32 uncoded bits),
// the result is 0.
// This test also catch Nan and infinities which also return 0.
// Compare exponent with 84 (compare exponent - 1 with 83).
cmp(scratch, Operand(83));
b(ge, &out_of_range);