David Anderson [:dvander] - inactive, e-mail if emergency Reporter
	Description • 14 years ago

+++ This bug was initially created as a clone of Bug #582152 +++

We've seen that loading doubles is very hot. We can do better on x86 when SSE4.1 is present.

From Moh:

On Penryn (SSE 4.1), we have other good instructions for this purpose.
More specifically, instead of the original sequence:

    mov [ebx + 0x80], eax
    mov [ebx + 0x84], ecx
    movsd xmm0, [ebx + 0x80]

a much better sequence is:
    movd   xmm0, eax 
    pinsrd xmm0, ecx, 1

	Alex Miller
	Comment 1 • 14 years ago

Introduces SSE4.1 checking and changes the assembler to use the movd and pinsrd instructions to load doubles.

	Alex Miller
	Comment 2 • 14 years ago

moved s_sseCheckState out of an #ifdef

	David Anderson [:dvander] - inactive, e-mail if emergency Reporter
	Comment 3 • 14 years ago

Comment on attachment 465529 [details] [diff] [review]
patch v2

Nice work! Seems like math heavy pieces of SunSpider get 1-2% faster.

	David Anderson [:dvander] - inactive, e-mail if emergency Reporter
	Comment 4 • 14 years ago

One nit: usage of pinsrd_rr() should assert that the SSE check resulted in 4.1.

	Sean Stangl [:sstangl]
	Comment 5 • 14 years ago

http://hg.mozilla.org/projects/jaegermonkey/rev/a16f3ba4b87c