A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
cranelift(aarch64): fold an i32x4 horizontal add into a single addv (#13717)
Damilare Darmie Akinlaja5937fbcd7b6b
diff --git a/cranelift/codegen/src/isa/aarch64/lower.isle b/cranelift/codegen/src/isa/aarch64/lower.isleindex ffffd57f95..d733c8e095 100644--- a/cranelift/codegen/src/isa/aarch64/lower.isle+++ b/cranelift/codegen/src/isa/aarch64/lower.isle@@ -2772,6 +2772,15 @@(u8_from_uimm8 lane)))(fpu_move_from_vec val lane (vector_size vty)))+;; A full `i32x4` horizontal add — two (GVN-shared) `iadd_pairwise`es then a+;; lane-0 extract — is a single `addv`, replacing `addp; addp; umov`.+(rule 3 (lower (extractlane $I32+ (iadd_pairwise $I32X4+ (iadd_pairwise $I32X4 x x)+ (iadd_pairwise $I32X4 x x))+ (u8_from_uimm8 0)))+ (mov_from_vec (addv x (vector_size $I32X4)) 0 (scalar_size $I32)))+;;; Rules for `insertlane` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;(rule 1 (lower (insertlane _ vec @ (value_type vty)diff --git a/cranelift/filetests/filetests/isa/aarch64/simd-addv-reduce.clif b/cranelift/filetests/filetests/isa/aarch64/simd-addv-reduce.clifnew file mode 100644index 0000000000..aacb3c3ac6--- /dev/null+++ b/cranelift/filetests/filetests/isa/aarch64/simd-addv-reduce.clif@@ -0,0 +1,25 @@+test compile precise-output+set unwind_info=false+target aarch64++;; A complete `i32x4` horizontal-add reduction contracts to a single `addv`.+function %hsum_i32x4(i32x4) -> i32 {+block0(v0: i32x4):+ v1 = iadd_pairwise v0, v0+ v2 = iadd_pairwise v1, v1+ v3 = extractlane v2, 0+ return v3+}++; VCode:+; block0:+; addv s2, v0.4s+; mov w0, v2.s[0]+; ret+;+; Disassembled:+; block0: ; offset 0x0+; addv s2, v0.4s+; mov w0, v2.s[0]+; ret+diff --git a/cranelift/filetests/filetests/runtests/simd-addv-reduce.clif b/cranelift/filetests/filetests/runtests/simd-addv-reduce.clifnew file mode 100644index 0000000000..8ada1e67a5--- /dev/null+++ b/cranelift/filetests/filetests/runtests/simd-addv-reduce.clif@@ -0,0 +1,19 @@+test interpret+test run+target aarch64+target x86_64 has_sse3 has_ssse3 has_sse41+target s390x++;; A complete `i32x4` horizontal-add reduction, expressed as two pairwise adds+;; plus a lane-0 extract. On aarch64 this lowers to a single `addv`.+function %hsum_i32x4(i32x4) -> i32 {+block0(v0: i32x4):+ v1 = iadd_pairwise v0, v0+ v2 = iadd_pairwise v1, v1+ v3 = extractlane v2, 0+ return v3+}+; run: %hsum_i32x4([1 2 3 4]) == 10+; run: %hsum_i32x4([10 -5 100 -1]) == 104+; run: %hsum_i32x4([0 0 0 0]) == 0+; run: %hsum_i32x4([-1 -1 -1 -1]) == -4