[ 
https://issues.apache.org/jira/browse/GROOVY-12065?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Daniel Sun updated GROOVY-12065:
--------------------------------
    Summary: Implement peephole optimization for bytecode generation  (was: 
Implement peephole optimization for constant loading in bytecode generation)

> Implement peephole optimization for bytecode generation
> -------------------------------------------------------
>
>                 Key: GROOVY-12065
>                 URL: https://issues.apache.org/jira/browse/GROOVY-12065
>             Project: Groovy
>          Issue Type: Improvement
>            Reporter: Daniel Sun
>            Priority: Major
>
> h3. Background
> Groovy's bytecode generator relied on ad-hoc, per-type conditional branches 
> scattered across {{OperandStack}} and {{BytecodeHelper}} to select compact 
> constant-loading instructions ({{{}ICONST_x{}}}, {{{}LCONST_x{}}}, 
> {{{}FCONST_x{}}}, {{{}DCONST_x{}}}, {{{}BIPUSH{}}}, {{{}SIPUSH{}}}) over the 
> generic {{LDC}} instruction. This logic was fragile, inconsistent, and 
> incomplete — for instance, {{BytecodeHelper.pushConstant()}} was missing the 
> {{ICONST_M1}} case for the integer constant {{{}-1{}}}.
> h3. Change
> A new {{PeepholeOptimizingMethodVisitor}} — a {{MethodVisitor}} decorator — 
> is introduced and inserted into the per-method visitor chain in 
> {{{}AsmClassGenerator{}}}. It intercepts every {{visitLdcInsn}} and 
> {{visitIntInsn(BIPUSH/SIPUSH, ...)}} call and rewrites each to the shortest 
> valid equivalent opcode:
> ||Value||Emitted instruction||
> |{{int}} −1|{{ICONST_M1}}|
> |{{int}} 0–5|{{ICONST_0}} – {{ICONST_5}}|
> |{{int}} in [−128, 127]|{{BIPUSH}}|
> |{{int}} in [−32768, 32767]|{{SIPUSH}}|
> |{{long}} 0 / 1|{{LCONST_0}} / {{LCONST_1}}|
> |{{float}} 0 / 1 / 2|{{FCONST_0}} / {{FCONST_1}} / {{FCONST_2}}|
> |{{double}} 0 / 1|{{DCONST_0}} / {{DCONST_1}}|
> Signed-zero {{float}} and {{double}} values are handled correctly by 
> comparing raw bit patterns (via {{Float.floatToRawIntBits}} / 
> {{{}Double.doubleToRawLongBits{}}}) rather than {{{}=={}}}, preserving the 
> {{-0.0}} / {{+0.0}} distinction that a straight equality check would collapse.
> All per-type branches previously inlined in {{OperandStack}} are replaced 
> with plain {{visitLdcInsn(value)}} calls; the peephole visitor centralizes 
> the selection uniformly for every generated method. 
> {{BytecodeHelper.pushConstant()}} also gains the previously missing 
> {{ICONST_M1}} branch.
>  



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to