From 737cf1d605a90b88e4bedac5af7c69a9dc485fd4 Mon Sep 17 00:00:00 2001 From: Andrew Brown Date: Mon, 29 Jun 2020 15:51:25 -0700 Subject: [PATCH] Implement `iabs` for x86 SIMD This only covers the types necessary for implementing the Wasm SIMD spec--`i8x16`, `i16x8`, `i32x4`. --- cranelift/codegen/meta/src/isa/x86/encodings.rs | 7 +++++++ cranelift/codegen/meta/src/isa/x86/opcodes.rs | 11 +++++++++++ .../filetests/isa/x86/simd-arithmetic-binemit.clif | 3 +++ .../filetests/isa/x86/simd-arithmetic-run.clif | 7 +++++++ 4 files changed, 28 insertions(+) diff --git a/cranelift/codegen/meta/src/isa/x86/encodings.rs b/cranelift/codegen/meta/src/isa/x86/encodings.rs index ad06f519c1..d4ae442085 100644 --- a/cranelift/codegen/meta/src/isa/x86/encodings.rs +++ b/cranelift/codegen/meta/src/isa/x86/encodings.rs @@ -1638,6 +1638,7 @@ fn define_simd( let fill_nop = shared.by_name("fill_nop"); let fmul = shared.by_name("fmul"); let fsub = shared.by_name("fsub"); + let iabs = shared.by_name("iabs"); let iadd = shared.by_name("iadd"); let icmp = shared.by_name("icmp"); let imul = shared.by_name("imul"); @@ -2184,6 +2185,12 @@ fn define_simd( e.enc_both_inferred(avgr, rec_fa.opcodes(opcodes)); } + // SIMD integer absolute value. + for (ty, opcodes) in &[(I8, &PABSB[..]), (I16, &PABSW[..]), (I32, &PABSD)] { + let iabs = iabs.bind(vector(*ty, sse_vector_size)); + e.enc_both_inferred_maybe_isap(iabs, rec_furm.opcodes(opcodes), Some(use_ssse3_simd)); + } + // SIMD logical operations let band = shared.by_name("band"); let band_not = shared.by_name("band_not"); diff --git a/cranelift/codegen/meta/src/isa/x86/opcodes.rs b/cranelift/codegen/meta/src/isa/x86/opcodes.rs index c38be80702..f7f7480f9b 100644 --- a/cranelift/codegen/meta/src/isa/x86/opcodes.rs +++ b/cranelift/codegen/meta/src/isa/x86/opcodes.rs @@ -303,6 +303,17 @@ pub static OR_IMM8_SIGN_EXTEND: [u8; 1] = [0x83]; /// Return the bitwise logical OR of packed single-precision values in xmm and x/m (SSE). pub static ORPS: [u8; 2] = [0x0f, 0x56]; +/// Compute the absolute value of bytes in xmm2/m128 and store the unsigned result in xmm1 (SSSE3). +pub static PABSB: [u8; 4] = [0x66, 0x0f, 0x38, 0x1c]; + +/// Compute the absolute value of 32-bit integers in xmm2/m128 and store the unsigned result in +/// xmm1 (SSSE3). +pub static PABSD: [u8; 4] = [0x66, 0x0f, 0x38, 0x1e]; + +/// Compute the absolute value of 16-bit integers in xmm2/m128 and store the unsigned result in +/// xmm1 (SSSE3). +pub static PABSW: [u8; 4] = [0x66, 0x0f, 0x38, 0x1d]; + /// Converts 8 packed signed word integers from xmm1 and from xxm2/m128 into 16 packed signed byte /// integers in xmm1 using signed saturation (SSE2). pub static PACKSSWB: [u8; 3] = [0x66, 0x0f, 0x63]; diff --git a/cranelift/filetests/filetests/isa/x86/simd-arithmetic-binemit.clif b/cranelift/filetests/filetests/isa/x86/simd-arithmetic-binemit.clif index 6a215a8e6c..8df5c2afa8 100644 --- a/cranelift/filetests/filetests/isa/x86/simd-arithmetic-binemit.clif +++ b/cranelift/filetests/filetests/isa/x86/simd-arithmetic-binemit.clif @@ -10,6 +10,7 @@ block0(v0: i8x16 [%xmm6], v1: i8x16 [%xmm2]): [-, %xmm6] v5 = ssub_sat v0, v1 ; bin: 66 0f e8 f2 [-, %xmm6] v6 = usub_sat v0, v1 ; bin: 66 0f d8 f2 [-, %xmm6] v7 = avg_round v0, v1 ; bin: 66 0f e0 f2 +[-, %xmm6] v9 = iabs v1 ; bin: 66 0f 38 1c f2 return } @@ -23,6 +24,7 @@ block0(v0: i16x8 [%xmm3], v1: i16x8 [%xmm5]): [-, %xmm3] v6 = ssub_sat v0, v1 ; bin: 66 0f e9 dd [-, %xmm3] v7 = usub_sat v0, v1 ; bin: 66 0f d9 dd [-, %xmm3] v8 = avg_round v0, v1 ; bin: 66 0f e3 dd +[-, %xmm3] v9 = iabs v1 ; bin: 66 0f 38 1d dd return } @@ -32,6 +34,7 @@ block0(v0: i32x4 [%xmm0], v1: i32x4 [%xmm1]): [-, %xmm0] v2 = iadd v0, v1 ; bin: 66 0f fe c1 [-, %xmm0] v3 = isub v0, v1 ; bin: 66 0f fa c1 [-, %xmm0] v4 = imul v0, v1 ; bin: 66 0f 38 40 c1 +[-, %xmm0] v5 = iabs v1 ; bin: 66 0f 38 1e c1 return } diff --git a/cranelift/filetests/filetests/isa/x86/simd-arithmetic-run.clif b/cranelift/filetests/filetests/isa/x86/simd-arithmetic-run.clif index 601defecfd..7acbe8e00b 100644 --- a/cranelift/filetests/filetests/isa/x86/simd-arithmetic-run.clif +++ b/cranelift/filetests/filetests/isa/x86/simd-arithmetic-run.clif @@ -270,3 +270,10 @@ block0: return v5 } ; run + +function %iabs(i32x4) -> i32x4 { +block0(v0: i32x4): + v1 = iabs v0 + return v1 +} +; run: %iabs([-42 -1 0 1]) == [42 1 0 1]