A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
x64: Fix sinking loads band/bor/bxor (#13071)
Alex Crichtond3f444c876d7
diff --git a/cranelift/codegen/src/isa/x64/lower.isle b/cranelift/codegen/src/isa/x64/lower.isleindex 31a15ebd32..456efc8e9f 100644--- a/cranelift/codegen/src/isa/x64/lower.isle+++ b/cranelift/codegen/src/isa/x64/lower.isle@@ -335,8 +335,11 @@;; f32 and f64+;; Note that `x` and `y` are both constrained to being in registers here because+;; x86 instructions with a memory operand will load 128 bits, not the desired+;; bit width of 32 or 64 bits here.(rule 5 (lower (has_type (ty_scalar_float ty) (band _ x y)))- (sse_and ty x y))+ (sse_and ty x (put_in_xmm y)));; SSE.@@ -459,8 +462,11 @@;; f32 and f64+;; Note that `x` and `y` are both constrained to being in registers here because+;; x86 instructions with a memory operand will load 128 bits, not the desired+;; bit width of 32 or 64 bits here.(rule 5 (lower (has_type (ty_scalar_float ty) (bor _ x y)))- (sse_or ty x y))+ (sse_or ty x (put_in_xmm y)));; SSE.@@ -530,8 +536,11 @@;; f32 and f64+;; Note that `x` and `y` are both constrained to being in registers here because+;; x86 instructions with a memory operand will load 128 bits, not the desired+;; bit width of 32 or 64 bits here.(rule 5 (lower (has_type (ty_scalar_float ty) (bxor _ x y)))- (x64_xor_vector ty x y))+ (x64_xor_vector ty x (put_in_xmm y)));; SSE.@@ -624,7 +633,7 @@(unmasked Xmm (x64_psllw src (mov_rmi_to_xmm masked_amt)))(mask_addr SyntheticAmode (ishl_i8x16_mask masked_amt))(mask Reg (x64_movdqu_load mask_addr)))- (sse_and $I8X16 unmasked (RegMem.Reg mask))))+ (x64_pand unmasked (RegMem.Reg mask))));; Get the address of the mask to use when fixing up the lanes that weren't;; correctly generated by the 16x8 shift.@@ -728,9 +737,7 @@;; correct for half of the lanes; the others must be fixed up with;; the mask below.(unmasked Xmm (x64_psrlw src (mov_rmi_to_xmm masked_amt))))- (sse_and $I8X16- unmasked- (ushr_i8x16_mask masked_amt))))+ (x64_pand unmasked (ushr_i8x16_mask masked_amt))));; Get the address of the mask to use when fixing up the lanes that weren't;; correctly generated by the 16x8 shift.@@ -2491,12 +2498,12 @@(rule 1 (lower (has_type $I8X16 (popcnt _ src)))(if-let true (has_ssse3))(let ((low_mask XmmMem (emit_u128_le_const 0x0f0f0f0f0f0f0f0f0f0f0f0f0f0f0f0f))- (low_nibbles Xmm (sse_and $I8X16 src low_mask))+ (low_nibbles Xmm (x64_pand src low_mask));; Note that this is a 16x8 shift, but that's OK; we mask;; off anything that traverses from one byte to the next;; with the low_mask below.(shifted_src Xmm (x64_psrlw src (xmi_imm 4)))- (high_nibbles Xmm (sse_and $I8X16 shifted_src low_mask))+ (high_nibbles Xmm (x64_pand shifted_src low_mask))(lookup Xmm (x64_xmm_load_const $I8X16(emit_u128_le_const 0x04030302_03020201_03020201_02010100)))(bit_counts_low Xmm (x64_pshufb lookup low_nibbles))diff --git a/cranelift/filetests/filetests/isa/x64/float-bitops-mem-operand.clif b/cranelift/filetests/filetests/isa/x64/float-bitops-mem-operand.clifnew file mode 100644index 0000000000..db8809fc42--- /dev/null+++ b/cranelift/filetests/filetests/isa/x64/float-bitops-mem-operand.clif@@ -0,0 +1,173 @@+test compile precise-output+target x86_64++function %f32_band(i64, f32) -> f32 {+block0(v0: i64, v1: f32):+ v2 = load.f32 aligned v0+ v3 = band v1, v2+ return v3+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; movss (%rdi), %xmm4+; andps %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; movss (%rdi), %xmm4 ; trap: heap_oob+; andps %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq+++function %f64_band(i64, f64) -> f64 {+block0(v0: i64, v1: f64):+ v2 = load.f64 aligned v0+ v3 = band v1, v2+ return v3+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; movsd (%rdi), %xmm4+; andpd %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; movsd (%rdi), %xmm4 ; trap: heap_oob+; andpd %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq+++function %f32_bor(i64, f32) -> f32 {+block0(v0: i64, v1: f32):+ v2 = load.f32 aligned v0+ v3 = bor v1, v2+ return v3+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; movss (%rdi), %xmm4+; orps %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; movss (%rdi), %xmm4 ; trap: heap_oob+; orps %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq++function %f64_bor(i64, f64) -> f64 {+block0(v0: i64, v1: f64):+ v2 = load.f64 aligned v0+ v3 = bor v1, v2+ return v3+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; movsd (%rdi), %xmm4+; orpd %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; movsd (%rdi), %xmm4 ; trap: heap_oob+; orpd %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq++function %f32_bxor(i64, f32) -> f32 {+block0(v0: i64, v1: f32):+ v2 = load.f32 aligned v0+ v3 = bxor v1, v2+ return v3+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; movss (%rdi), %xmm4+; xorps %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; movss (%rdi), %xmm4 ; trap: heap_oob+; xorps %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq++function %f64_bxor(i64, f64) -> f64 {+block0(v0: i64, v1: f64):+ v2 = load.f64 aligned v0+ v3 = bxor v1, v2+ return v3+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; movsd (%rdi), %xmm4+; xorpd %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; movsd (%rdi), %xmm4 ; trap: heap_oob+; xorpd %xmm4, %xmm0+; movq %rbp, %rsp+; popq %rbp+; retq+