T0b1/wasmtime
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add x86-specific instruction for i64x2 multiplication

Browse Source 
Without this special instruction, legalizing to the AVX512 instruction AND the SSE instruction sequence is impossible. This extra instruction would be rendered unnecessary by the x64 backend.
This commit is contained in:
Andrew Brown
2020-05-20 16:17:16 -07:00
parent 8701645493
commit 9ba9fd0f64
6 changed files with 35 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
cranelift/filetests/filetests/isa/x86/simd-arithmetic-legalize.clif
View File

@@ -69,3 +69,10 @@ block0:
; nextln: v1 = band v0, v4
return
}
function %imul(i64x2, i64x2) {
block0(v0:i64x2, v1:i64x2):
v2 = imul v0, v1
; check: v2 = x86_pmullq v0, v1
return
}

2
cranelift/filetests/filetests/isa/x86/simd-avx512-arithmetic-binemit.clif
View File

@@ -6,7 +6,7 @@ function %imul_i64x2() {
block0:
[-, %xmm1] v0 = vconst.i64x2 [1 2]
[-, %xmm2] v1 = vconst.i64x2 [2 2]
[-, %xmm14] v2 = imul v0, v1 ; bin: 62 72 f5 08 40 f2
[-, %xmm14] v2 = x86_pmullq v0, v1 ; bin: 62 72 f5 08 40 f2
; 62, mandatory EVEX prefix
; 72 = 0111 0010, R is set (MSB in %xmm14) while X, B, and R' are unset (note these are all inverted); mm is set to 0F38
; f5 = 1111 0101, W is set (64-bit op), vvvv set to 1 (inverted), bit 2 always set, pp set to 01