From 06af2fb38a531834cf5fa3ff43ccab2c1f2f2f20 Mon Sep 17 00:00:00 2001 From: Alex Beregszaszi Date: Tue, 14 Jul 2020 11:04:55 +0100 Subject: [PATCH] [ewasm] Change eq() in polyfill to branch-less version --- libyul/backends/wasm/polyfill/Arithmetic.yul | 6 +- libyul/backends/wasm/polyfill/Comparison.yul | 21 +++-- test/cmdlineTests/evm_to_wasm_break/output | 99 +++++++++++--------- 3 files changed, 71 insertions(+), 55 deletions(-) diff --git a/libyul/backends/wasm/polyfill/Arithmetic.yul b/libyul/backends/wasm/polyfill/Arithmetic.yul index 4a0e32bf3..7a4a80f5f 100644 --- a/libyul/backends/wasm/polyfill/Arithmetic.yul +++ b/libyul/backends/wasm/polyfill/Arithmetic.yul @@ -90,9 +90,9 @@ function mul_64x64_128(x, y) -> hi, lo { // value split into four 64 bit values. function mul_128x128_256(x1, x2, y1, y2) -> r1, r2, r3, r4 { let ah, al := mul_64x64_128(x1, y1) - let bh, bl := mul_64x64_128(x1, y2) - let ch, cl := mul_64x64_128(x2, y1) - let dh, dl := mul_64x64_128(x2, y2) + let bh, bl := mul_64x64_128(x1, y2) + let ch, cl := mul_64x64_128(x2, y1) + let dh, dl := mul_64x64_128(x2, y2) r4 := dl let carry1, carry2 let t1, t2 diff --git a/libyul/backends/wasm/polyfill/Comparison.yul b/libyul/backends/wasm/polyfill/Comparison.yul index 6f08c5b18..c82669f80 100644 --- a/libyul/backends/wasm/polyfill/Comparison.yul +++ b/libyul/backends/wasm/polyfill/Comparison.yul @@ -35,15 +35,18 @@ function iszero512(x1, x2, x3, x4, x5, x6, x7, x8) -> r:i32 { } function eq(x1, x2, x3, x4, y1, y2, y3, y4) -> r1, r2, r3, r4 { - if i64.eq(x1, y1) { - if i64.eq(x2, y2) { - if i64.eq(x3, y3) { - if i64.eq(x4, y4) { - r4 := 1 - } - } - } - } + r4 := i64.extend_i32_u( + i32.and( + i64.eq(x1, y1), + i32.and( + i64.eq(x2, y2), + i32.and( + i64.eq(x3, y3), + i64.eq(x4, y4) + ) + ) + ) + ) } // returns 0 if a == b, -1 if a < b and 1 if a > b diff --git a/test/cmdlineTests/evm_to_wasm_break/output b/test/cmdlineTests/evm_to_wasm_break/output index 8346f3c42..6b931671b 100644 --- a/test/cmdlineTests/evm_to_wasm_break/output +++ b/test/cmdlineTests/evm_to_wasm_break/output @@ -31,10 +31,9 @@ object "object" { let x_6 := x_2 let x_7 := x_3 let _2 := 1 - let _3 := i64.or(_1, _1) - let _4:i32 := i32.eqz(i32.eqz(i64.eqz(i64.or(_3, i64.or(_1, _2))))) + let _3:i32 := i32.eqz(i32.eqz(i64.eqz(i64.or(i64.or(_1, _1), i64.or(_1, _2))))) for { } - i32.eqz(_4) + i32.eqz(_3) { let x_8, x_9, x_10, x_11 := add(x_4, x_5, x_6, x_7, _1, _1, _1, _2) x_4 := x_8 @@ -43,10 +42,12 @@ object "object" { x_7 := x_11 } { - let _5, _6, _7, _8 := iszero_197_919_1528(_1, _1, _1, lt_199(x_4, x_5, x_6, x_7, _1, _1, _1, 10)) - if i32.eqz(i64.eqz(i64.or(i64.or(_5, _6), i64.or(_7, _8)))) { break } - if i32.eqz(i64.eqz(i64.or(_3, i64.or(_1, eq(x_4, x_5, x_6, x_7, _1, _1, _1, 2))))) { break } - if i32.eqz(i64.eqz(i64.or(_3, i64.or(_1, eq(x_4, x_5, x_6, x_7, _1, _1, _1, 4))))) { continue } + let _4, _5, _6, _7 := iszero_200_868_1493(_1, _1, _1, lt_202(x_4, x_5, x_6, x_7, _1, _1, _1, 10)) + if i32.eqz(i64.eqz(i64.or(i64.or(_4, _5), i64.or(_6, _7)))) { break } + let _8, _9, _10, _11 := eq_201_869_1494(x_4, x_5, x_6, x_7, _1, _1, _1, 2) + if i32.eqz(i64.eqz(i64.or(i64.or(_8, _9), i64.or(_10, _11)))) { break } + let _12, _13, _14, _15 := eq_201_869_1494(x_4, x_5, x_6, x_7, _1, _1, _1, 4) + if i32.eqz(i64.eqz(i64.or(i64.or(_12, _13), i64.or(_14, _15)))) { continue } } sstore(_1, _1, _1, _1, x_4, x_5, x_6, x_7) } @@ -67,19 +68,13 @@ object "object" { let r1_1, carry_2 := add_carry(x1, y1, carry_1) r1 := r1_1 } - function iszero_197_919_1528(x1, x2, x3, x4) -> r1, r2, r3, r4 + function iszero_200_868_1493(x1, x2, x3, x4) -> r1, r2, r3, r4 { r4 := i64.extend_i32_u(i64.eqz(i64.or(i64.or(x1, x2), i64.or(x3, x4)))) } - function eq(x1, x2, x3, x4, y1, y2, y3, y4) -> r4 + function eq_201_869_1494(x1, x2, x3, x4, y1, y2, y3, y4) -> r1, r2, r3, r4 { - if i64.eq(x1, y1) - { - if i64.eq(x2, y2) - { - if i64.eq(x3, y3) { if i64.eq(x4, y4) { r4 := 1 } } - } - } + r4 := i64.extend_i32_u(i32.and(i64.eq(x1, y1), i32.and(i64.eq(x2, y2), i32.and(i64.eq(x3, y3), i64.eq(x4, y4))))) } function cmp(a, b) -> r:i32 { @@ -87,7 +82,7 @@ object "object" { case 1:i32 { r := 0xffffffff:i32 } default { r := i64.ne(a, b) } } - function lt_199(x1, x2, x3, x4, y1, y2, y3, y4) -> z4 + function lt_202(x1, x2, x3, x4, y1, y2, y3, y4) -> z4 { let z:i32 := false switch cmp(x1, y1) @@ -178,7 +173,7 @@ object "object" { Binary representation: -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 Text representation: (module @@ -203,12 +198,19 @@ Text representation: (local $x_6 i64) (local $x_7 i64) (local $_2 i64) - (local $_3 i64) - (local $_4 i32) + (local $_3 i32) + (local $_4 i64) (local $_5 i64) (local $_6 i64) (local $_7 i64) (local $_8 i64) + (local $_9 i64) + (local $_10 i64) + (local $_11 i64) + (local $_12 i64) + (local $_13 i64) + (local $_14 i64) + (local $_15 i64) (local $x_8 i64) (local $x_9 i64) (local $x_10 i64) @@ -227,26 +229,39 @@ Text representation: (local.set $x_6 (local.get $x_2)) (local.set $x_7 (local.get $x_3)) (local.set $_2 (i64.const 1)) - (local.set $_3 (i64.or (local.get $_1) (local.get $_1))) - (local.set $_4 (i32.eqz (i32.eqz (i64.eqz (i64.or (local.get $_3) (i64.or (local.get $_1) (local.get $_2))))))) + (local.set $_3 (i32.eqz (i32.eqz (i64.eqz (i64.or (i64.or (local.get $_1) (local.get $_1)) (i64.or (local.get $_1) (local.get $_2))))))) (block $label__3 (loop $label__5 - (br_if $label__3 (i32.eqz (i32.eqz (local.get $_4)))) + (br_if $label__3 (i32.eqz (i32.eqz (local.get $_3)))) (block $label__4 (block - (local.set $_5 (call $iszero_197_919_1528 (local.get $_1) (local.get $_1) (local.get $_1) (call $lt_199 (local.get $x_4) (local.get $x_5) (local.get $x_6) (local.get $x_7) (local.get $_1) (local.get $_1) (local.get $_1) (i64.const 10)))) - (local.set $_6 (global.get $global_)) - (local.set $_7 (global.get $global__1)) - (local.set $_8 (global.get $global__2)) + (local.set $_4 (call $iszero_200_868_1493 (local.get $_1) (local.get $_1) (local.get $_1) (call $lt_202 (local.get $x_4) (local.get $x_5) (local.get $x_6) (local.get $x_7) (local.get $_1) (local.get $_1) (local.get $_1) (i64.const 10)))) + (local.set $_5 (global.get $global_)) + (local.set $_6 (global.get $global__1)) + (local.set $_7 (global.get $global__2)) ) - (if (i32.eqz (i64.eqz (i64.or (i64.or (local.get $_5) (local.get $_6)) (i64.or (local.get $_7) (local.get $_8))))) (then + (if (i32.eqz (i64.eqz (i64.or (i64.or (local.get $_4) (local.get $_5)) (i64.or (local.get $_6) (local.get $_7))))) (then (br $label__3) )) - (if (i32.eqz (i64.eqz (i64.or (local.get $_3) (i64.or (local.get $_1) (call $eq (local.get $x_4) (local.get $x_5) (local.get $x_6) (local.get $x_7) (local.get $_1) (local.get $_1) (local.get $_1) (i64.const 2)))))) (then + (block + (local.set $_8 (call $eq_201_869_1494 (local.get $x_4) (local.get $x_5) (local.get $x_6) (local.get $x_7) (local.get $_1) (local.get $_1) (local.get $_1) (i64.const 2))) + (local.set $_9 (global.get $global_)) + (local.set $_10 (global.get $global__1)) + (local.set $_11 (global.get $global__2)) + + ) + (if (i32.eqz (i64.eqz (i64.or (i64.or (local.get $_8) (local.get $_9)) (i64.or (local.get $_10) (local.get $_11))))) (then (br $label__3) )) - (if (i32.eqz (i64.eqz (i64.or (local.get $_3) (i64.or (local.get $_1) (call $eq (local.get $x_4) (local.get $x_5) (local.get $x_6) (local.get $x_7) (local.get $_1) (local.get $_1) (local.get $_1) (i64.const 4)))))) (then + (block + (local.set $_12 (call $eq_201_869_1494 (local.get $x_4) (local.get $x_5) (local.get $x_6) (local.get $x_7) (local.get $_1) (local.get $_1) (local.get $_1) (i64.const 4))) + (local.set $_13 (global.get $global_)) + (local.set $_14 (global.get $global__1)) + (local.set $_15 (global.get $global__2)) + + ) + (if (i32.eqz (i64.eqz (i64.or (i64.or (local.get $_12) (local.get $_13)) (i64.or (local.get $_14) (local.get $_15))))) (then (br $label__4) )) @@ -338,7 +353,7 @@ Text representation: (local.get $r1) ) -(func $iszero_197_919_1528 +(func $iszero_200_868_1493 (param $x1 i64) (param $x2 i64) (param $x3 i64) @@ -358,7 +373,7 @@ Text representation: (local.get $r1) ) -(func $eq +(func $eq_201_869_1494 (param $x1 i64) (param $x2 i64) (param $x3 i64) @@ -368,20 +383,18 @@ Text representation: (param $y3 i64) (param $y4 i64) (result i64) + (local $r1 i64) + (local $r2 i64) + (local $r3 i64) (local $r4 i64) (block $label__9 - (if (i64.eq (local.get $x1) (local.get $y1)) (then - (if (i64.eq (local.get $x2) (local.get $y2)) (then - (if (i64.eq (local.get $x3) (local.get $y3)) (then - (if (i64.eq (local.get $x4) (local.get $y4)) (then - (local.set $r4 (i64.const 1)) - )) - )) - )) - )) + (local.set $r4 (i64.extend_i32_u (i32.and (i64.eq (local.get $x1) (local.get $y1)) (i32.and (i64.eq (local.get $x2) (local.get $y2)) (i32.and (i64.eq (local.get $x3) (local.get $y3)) (i64.eq (local.get $x4) (local.get $y4))))))) ) - (local.get $r4) + (global.set $global_ (local.get $r2)) + (global.set $global__1 (local.get $r3)) + (global.set $global__2 (local.get $r4)) + (local.get $r1) ) (func $cmp @@ -405,7 +418,7 @@ Text representation: (local.get $r) ) -(func $lt_199 +(func $lt_202 (param $x1 i64) (param $x2 i64) (param $x3 i64)